camchanimation commited on
Commit
7dd6673
1 Parent(s): ea3c4e3

Upload folder using huggingface_hub

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .eslintignore +5 -0
  2. .eslintrc.js +96 -0
  3. .git-blame-ignore-revs +2 -0
  4. .github/ISSUE_TEMPLATE/bug_report.yml +105 -0
  5. .github/ISSUE_TEMPLATE/config.yml +5 -0
  6. .github/ISSUE_TEMPLATE/feature_request.yml +40 -0
  7. .github/pull_request_template.md +15 -0
  8. .github/workflows/on_pull_request.yaml +38 -0
  9. .github/workflows/run_tests.yaml +107 -0
  10. .github/workflows/warns_merge_master.yml +19 -0
  11. .gitignore +44 -0
  12. .pylintrc +3 -0
  13. CHANGELOG.md +674 -0
  14. CITATION.cff +7 -0
  15. CODEOWNERS +1 -0
  16. LICENSE.txt +663 -0
  17. README.md +705 -7
  18. configs/alt-diffusion-inference.yaml +72 -0
  19. configs/alt-diffusion-m18-inference.yaml +73 -0
  20. configs/instruct-pix2pix.yaml +98 -0
  21. configs/sd_xl_inpaint.yaml +98 -0
  22. configs/v1-inference.yaml +70 -0
  23. configs/v1-inpainting-inference.yaml +70 -0
  24. environment-wsl2.yaml +11 -0
  25. extensions-builtin/LDSR/ldsr_model_arch.py +250 -0
  26. extensions-builtin/LDSR/preload.py +6 -0
  27. extensions-builtin/LDSR/scripts/ldsr_model.py +70 -0
  28. extensions-builtin/LDSR/sd_hijack_autoencoder.py +293 -0
  29. extensions-builtin/LDSR/sd_hijack_ddpm_v1.py +1443 -0
  30. extensions-builtin/LDSR/vqvae_quantize.py +147 -0
  31. extensions-builtin/Lora/extra_networks_lora.py +67 -0
  32. extensions-builtin/Lora/lora.py +9 -0
  33. extensions-builtin/Lora/lora_logger.py +33 -0
  34. extensions-builtin/Lora/lora_patches.py +6 -0
  35. extensions-builtin/Lora/network.py +190 -0
  36. extensions-builtin/Lora/networks.py +219 -0
  37. extensions-builtin/Lora/preload.py +8 -0
  38. extensions-builtin/Lora/scripts/lora_script.py +100 -0
  39. extensions-builtin/Lora/ui_edit_user_metadata.py +222 -0
  40. extensions-builtin/Lora/ui_extra_networks_lora.py +90 -0
  41. extensions-builtin/ScuNET/preload.py +6 -0
  42. extensions-builtin/ScuNET/scripts/scunet_model.py +74 -0
  43. extensions-builtin/SwinIR/preload.py +6 -0
  44. extensions-builtin/SwinIR/scripts/swinir_model.py +98 -0
  45. extensions-builtin/canvas-zoom-and-pan/javascript/zoom.js +968 -0
  46. extensions-builtin/canvas-zoom-and-pan/scripts/hotkey_config.py +17 -0
  47. extensions-builtin/canvas-zoom-and-pan/style.css +66 -0
  48. extensions-builtin/extra-options-section/scripts/extra_options_section.py +78 -0
  49. extensions-builtin/forge_legacy_preprocessors/.gitignore +185 -0
  50. extensions-builtin/forge_legacy_preprocessors/LICENSE +674 -0
.eslintignore ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ extensions
2
+ extensions-disabled
3
+ extensions-builtin/sd_forge_controlnet
4
+ repositories
5
+ venv
.eslintrc.js ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /* global module */
2
+ module.exports = {
3
+ env: {
4
+ browser: true,
5
+ es2021: true,
6
+ },
7
+ extends: "eslint:recommended",
8
+ parserOptions: {
9
+ ecmaVersion: "latest",
10
+ },
11
+ rules: {
12
+ "arrow-spacing": "error",
13
+ "block-spacing": "error",
14
+ "brace-style": "error",
15
+ "comma-dangle": ["error", "only-multiline"],
16
+ "comma-spacing": "error",
17
+ "comma-style": ["error", "last"],
18
+ "curly": ["error", "multi-line", "consistent"],
19
+ "eol-last": "error",
20
+ "func-call-spacing": "error",
21
+ "function-call-argument-newline": ["error", "consistent"],
22
+ "function-paren-newline": ["error", "consistent"],
23
+ "indent": ["error", 4],
24
+ "key-spacing": "error",
25
+ "keyword-spacing": "error",
26
+ "linebreak-style": ["error", "unix"],
27
+ "no-extra-semi": "error",
28
+ "no-mixed-spaces-and-tabs": "error",
29
+ "no-multi-spaces": "error",
30
+ "no-redeclare": ["error", {builtinGlobals: false}],
31
+ "no-trailing-spaces": "error",
32
+ "no-unused-vars": "off",
33
+ "no-whitespace-before-property": "error",
34
+ "object-curly-newline": ["error", {consistent: true, multiline: true}],
35
+ "object-curly-spacing": ["error", "never"],
36
+ "operator-linebreak": ["error", "after"],
37
+ "quote-props": ["error", "consistent-as-needed"],
38
+ "semi": ["error", "always"],
39
+ "semi-spacing": "error",
40
+ "semi-style": ["error", "last"],
41
+ "space-before-blocks": "error",
42
+ "space-before-function-paren": ["error", "never"],
43
+ "space-in-parens": ["error", "never"],
44
+ "space-infix-ops": "error",
45
+ "space-unary-ops": "error",
46
+ "switch-colon-spacing": "error",
47
+ "template-curly-spacing": ["error", "never"],
48
+ "unicode-bom": "error",
49
+ },
50
+ globals: {
51
+ //script.js
52
+ gradioApp: "readonly",
53
+ executeCallbacks: "readonly",
54
+ onAfterUiUpdate: "readonly",
55
+ onOptionsChanged: "readonly",
56
+ onUiLoaded: "readonly",
57
+ onUiUpdate: "readonly",
58
+ uiCurrentTab: "writable",
59
+ uiElementInSight: "readonly",
60
+ uiElementIsVisible: "readonly",
61
+ //ui.js
62
+ opts: "writable",
63
+ all_gallery_buttons: "readonly",
64
+ selected_gallery_button: "readonly",
65
+ selected_gallery_index: "readonly",
66
+ switch_to_txt2img: "readonly",
67
+ switch_to_img2img_tab: "readonly",
68
+ switch_to_img2img: "readonly",
69
+ switch_to_sketch: "readonly",
70
+ switch_to_inpaint: "readonly",
71
+ switch_to_inpaint_sketch: "readonly",
72
+ switch_to_extras: "readonly",
73
+ get_tab_index: "readonly",
74
+ create_submit_args: "readonly",
75
+ restart_reload: "readonly",
76
+ updateInput: "readonly",
77
+ onEdit: "readonly",
78
+ //extraNetworks.js
79
+ requestGet: "readonly",
80
+ popup: "readonly",
81
+ // from python
82
+ localization: "readonly",
83
+ // progrssbar.js
84
+ randomId: "readonly",
85
+ requestProgress: "readonly",
86
+ // imageviewer.js
87
+ modalPrevImage: "readonly",
88
+ modalNextImage: "readonly",
89
+ // localStorage.js
90
+ localSet: "readonly",
91
+ localGet: "readonly",
92
+ localRemove: "readonly",
93
+ // resizeHandle.js
94
+ setupResizeHandle: "writable"
95
+ }
96
+ };
.git-blame-ignore-revs ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ # Apply ESlint
2
+ 9c54b78d9dde5601e916f308d9a9d6953ec39430
.github/ISSUE_TEMPLATE/bug_report.yml ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name: Bug Report
2
+ description: You think something is broken in the UI
3
+ title: "[Bug]: "
4
+ labels: ["bug-report"]
5
+
6
+ body:
7
+ - type: markdown
8
+ attributes:
9
+ value: |
10
+ > The title of the bug report should be short and descriptive.
11
+ > Use relevant keywords for searchability.
12
+ > Do not leave it blank, but also do not put an entire error log in it.
13
+ - type: checkboxes
14
+ attributes:
15
+ label: Checklist
16
+ description: |
17
+ Please perform basic debugging to see if extensions or configuration is the cause of the issue.
18
+ Basic debug procedure
19
+  1. Disable all third-party extensions - check if extension is the cause
20
+  2. Update extensions and webui - sometimes things just need to be updated
21
+  3. Backup and remove your config.json and ui-config.json - check if the issue is caused by bad configuration
22
+  4. Delete venv with third-party extensions disabled - sometimes extensions might cause wrong libraries to be installed
23
+  5. Try a fresh installation webui in a different directory - see if a clean installation solves the issue
24
+ Before making a issue report please, check that the issue hasn't been reported recently.
25
+ options:
26
+ - label: The issue exists after disabling all extensions
27
+ - label: The issue exists on a clean installation of webui
28
+ - label: The issue is caused by an extension, but I believe it is caused by a bug in the webui
29
+ - label: The issue exists in the current version of the webui
30
+ - label: The issue has not been reported before recently
31
+ - label: The issue has been reported before but has not been fixed yet
32
+ - type: markdown
33
+ attributes:
34
+ value: |
35
+ > Please fill this form with as much information as possible. Don't forget to "Upload Sysinfo" and "What browsers" and provide screenshots if possible
36
+ - type: textarea
37
+ id: what-did
38
+ attributes:
39
+ label: What happened?
40
+ description: Tell us what happened in a very clear and simple way
41
+ placeholder: |
42
+ txt2img is not working as intended.
43
+ validations:
44
+ required: true
45
+ - type: textarea
46
+ id: steps
47
+ attributes:
48
+ label: Steps to reproduce the problem
49
+ description: Please provide us with precise step by step instructions on how to reproduce the bug
50
+ placeholder: |
51
+ 1. Go to ...
52
+ 2. Press ...
53
+ 3. ...
54
+ validations:
55
+ required: true
56
+ - type: textarea
57
+ id: what-should
58
+ attributes:
59
+ label: What should have happened?
60
+ description: Tell us what you think the normal behavior should be
61
+ placeholder: |
62
+ WebUI should ...
63
+ validations:
64
+ required: true
65
+ - type: dropdown
66
+ id: browsers
67
+ attributes:
68
+ label: What browsers do you use to access the UI ?
69
+ multiple: true
70
+ options:
71
+ - Mozilla Firefox
72
+ - Google Chrome
73
+ - Brave
74
+ - Apple Safari
75
+ - Microsoft Edge
76
+ - Android
77
+ - iOS
78
+ - Other
79
+ - type: textarea
80
+ id: sysinfo
81
+ attributes:
82
+ label: Sysinfo
83
+ description: System info file, generated by WebUI. You can generate it in settings, on the Sysinfo page. Drag the file into the field to upload it. If you submit your report without including the sysinfo file, the report will be closed. If needed, review the report to make sure it includes no personal information you don't want to share. If you can't start WebUI, you can use --dump-sysinfo commandline argument to generate the file.
84
+ placeholder: |
85
+ 1. Go to WebUI Settings -> Sysinfo -> Download system info.
86
+ If WebUI fails to launch, use --dump-sysinfo commandline argument to generate the file
87
+ 2. Upload the Sysinfo as a attached file, Do NOT paste it in as plain text.
88
+ validations:
89
+ required: true
90
+ - type: textarea
91
+ id: logs
92
+ attributes:
93
+ label: Console logs
94
+ description: Please provide **full** cmd/terminal logs from the moment you started UI to the end of it, after the bug occured. If it's very long, provide a link to pastebin or similar service.
95
+ render: Shell
96
+ validations:
97
+ required: true
98
+ - type: textarea
99
+ id: misc
100
+ attributes:
101
+ label: Additional information
102
+ description: |
103
+ Please provide us with any relevant additional info or context.
104
+ Examples:
105
+  I have updated my GPU driver recently.
.github/ISSUE_TEMPLATE/config.yml ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ blank_issues_enabled: false
2
+ contact_links:
3
+ - name: WebUI Community Support
4
+ url: https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions
5
+ about: Please ask and answer questions here.
.github/ISSUE_TEMPLATE/feature_request.yml ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name: Feature request
2
+ description: Suggest an idea for this project
3
+ title: "[Feature Request]: "
4
+ labels: ["enhancement"]
5
+
6
+ body:
7
+ - type: checkboxes
8
+ attributes:
9
+ label: Is there an existing issue for this?
10
+ description: Please search to see if an issue already exists for the feature you want, and that it's not implemented in a recent build/commit.
11
+ options:
12
+ - label: I have searched the existing issues and checked the recent builds/commits
13
+ required: true
14
+ - type: markdown
15
+ attributes:
16
+ value: |
17
+ *Please fill this form with as much information as possible, provide screenshots and/or illustrations of the feature if possible*
18
+ - type: textarea
19
+ id: feature
20
+ attributes:
21
+ label: What would your feature do ?
22
+ description: Tell us about your feature in a very clear and simple way, and what problem it would solve
23
+ validations:
24
+ required: true
25
+ - type: textarea
26
+ id: workflow
27
+ attributes:
28
+ label: Proposed workflow
29
+ description: Please provide us with step by step information on how you'd like the feature to be accessed and used
30
+ value: |
31
+ 1. Go to ....
32
+ 2. Press ....
33
+ 3. ...
34
+ validations:
35
+ required: true
36
+ - type: textarea
37
+ id: misc
38
+ attributes:
39
+ label: Additional information
40
+ description: Add any other context or screenshots about the feature request here.
.github/pull_request_template.md ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ## Description
2
+
3
+ * a simple description of what you're trying to accomplish
4
+ * a summary of changes in code
5
+ * which issues it fixes, if any
6
+
7
+ ## Screenshots/videos:
8
+
9
+
10
+ ## Checklist:
11
+
12
+ - [ ] I have read [contributing wiki page](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Contributing)
13
+ - [ ] I have performed a self-review of my own code
14
+ - [ ] My code follows the [style guidelines](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Contributing#code-style)
15
+ - [ ] My code passes [tests](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Tests)
.github/workflows/on_pull_request.yaml ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name: Linter
2
+
3
+ on:
4
+ - push
5
+ - pull_request
6
+
7
+ jobs:
8
+ lint-python:
9
+ name: ruff
10
+ runs-on: ubuntu-latest
11
+ if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name != github.event.pull_request.base.repo.full_name
12
+ steps:
13
+ - name: Checkout Code
14
+ uses: actions/checkout@v3
15
+ - uses: actions/setup-python@v4
16
+ with:
17
+ python-version: 3.11
18
+ # NB: there's no cache: pip here since we're not installing anything
19
+ # from the requirements.txt file(s) in the repository; it's faster
20
+ # not to have GHA download an (at the time of writing) 4 GB cache
21
+ # of PyTorch and other dependencies.
22
+ - name: Install Ruff
23
+ run: pip install ruff==0.1.6
24
+ - name: Run Ruff
25
+ run: ruff .
26
+ lint-js:
27
+ name: eslint
28
+ runs-on: ubuntu-latest
29
+ if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name != github.event.pull_request.base.repo.full_name
30
+ steps:
31
+ - name: Checkout Code
32
+ uses: actions/checkout@v3
33
+ - name: Install Node.js
34
+ uses: actions/setup-node@v3
35
+ with:
36
+ node-version: 18
37
+ - run: npm i --ci
38
+ - run: npm run lint
.github/workflows/run_tests.yaml ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name: Tests
2
+
3
+ on:
4
+ - push
5
+ - pull_request
6
+
7
+ env:
8
+ FORGE_CQ_TEST: "True"
9
+
10
+ jobs:
11
+ test:
12
+ name: tests on CPU
13
+ runs-on: ubuntu-latest
14
+ if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name != github.event.pull_request.base.repo.full_name
15
+ steps:
16
+ - name: Checkout Code
17
+ uses: actions/checkout@v3
18
+ - name: Set up Python 3.10
19
+ uses: actions/setup-python@v4
20
+ with:
21
+ python-version: 3.10.6
22
+ cache: pip
23
+ cache-dependency-path: |
24
+ **/requirements*txt
25
+ launch.py
26
+ - name: Cache models
27
+ id: cache-models
28
+ uses: actions/cache@v3
29
+ with:
30
+ path: models
31
+ key: "2023-12-30"
32
+ - name: Install test dependencies
33
+ run: pip install wait-for-it -r requirements-test.txt
34
+ env:
35
+ PIP_DISABLE_PIP_VERSION_CHECK: "1"
36
+ PIP_PROGRESS_BAR: "off"
37
+ - name: Setup environment
38
+ run: python launch.py --skip-torch-cuda-test --exit
39
+ env:
40
+ PIP_DISABLE_PIP_VERSION_CHECK: "1"
41
+ PIP_PROGRESS_BAR: "off"
42
+ TORCH_INDEX_URL: https://download.pytorch.org/whl/cpu
43
+ WEBUI_LAUNCH_LIVE_OUTPUT: "1"
44
+ PYTHONUNBUFFERED: "1"
45
+ - name: Print installed packages
46
+ run: pip freeze
47
+ - name: Download models
48
+ run: |
49
+ declare -a urls=(
50
+ "https://huggingface.co/lllyasviel/fav_models/resolve/main/fav/realisticVisionV51_v51VAE.safetensors"
51
+ )
52
+ for url in "${urls[@]}"; do
53
+ filename="models/Stable-diffusion/${url##*/}" # Extracts the last part of the URL
54
+ if [ ! -f "$filename" ]; then
55
+ curl -Lo "$filename" "$url"
56
+ fi
57
+ done
58
+ # - name: Download ControlNet models
59
+ # run: |
60
+ # declare -a urls=(
61
+ # "https://huggingface.co/lllyasviel/ControlNet-v1-1/resolve/main/control_v11p_sd15_canny.pth"
62
+ # )
63
+
64
+ # for url in "${urls[@]}"; do
65
+ # filename="models/ControlNet/${url##*/}" # Extracts the last part of the URL
66
+ # if [ ! -f "$filename" ]; then
67
+ # curl -Lo "$filename" "$url"
68
+ # fi
69
+ # done
70
+ - name: Start test server
71
+ run: >
72
+ python -m coverage run
73
+ --data-file=.coverage.server
74
+ launch.py
75
+ --skip-prepare-environment
76
+ --skip-torch-cuda-test
77
+ --test-server
78
+ --do-not-download-clip
79
+ --no-half
80
+ --disable-opt-split-attention
81
+ --always-cpu
82
+ --api-server-stop
83
+ --ckpt models/Stable-diffusion/realisticVisionV51_v51VAE.safetensors
84
+ 2>&1 | tee output.txt &
85
+ - name: Run tests
86
+ run: |
87
+ wait-for-it --service 127.0.0.1:7860 -t 20
88
+ python -m pytest -vv --junitxml=test/results.xml --cov . --cov-report=xml --verify-base-url test
89
+ # TODO(huchenlei): Enable ControlNet tests. Currently it is too slow to run these tests on CPU with
90
+ # real SD model. We need to find a way to load empty SD model.
91
+ # - name: Run ControlNet tests
92
+ # run: >
93
+ # python -m pytest
94
+ # --junitxml=test/results.xml
95
+ # --cov ./extensions-builtin/sd_forge_controlnet
96
+ # --cov-report=xml
97
+ # --verify-base-url
98
+ # ./extensions-builtin/sd_forge_controlnet/tests
99
+ - name: Kill test server
100
+ if: always()
101
+ run: curl -vv -XPOST http://127.0.0.1:7860/sdapi/v1/server-stop && sleep 10
102
+ - name: Upload main app output
103
+ uses: actions/upload-artifact@v3
104
+ if: always()
105
+ with:
106
+ name: output
107
+ path: output.txt
.github/workflows/warns_merge_master.yml ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name: Pull requests can't target master branch
2
+
3
+ "on":
4
+ pull_request:
5
+ types:
6
+ - opened
7
+ - synchronize
8
+ - reopened
9
+ branches:
10
+ - master
11
+
12
+ jobs:
13
+ check:
14
+ runs-on: ubuntu-latest
15
+ steps:
16
+ - name: Warning marge into master
17
+ run: |
18
+ echo -e "::warning::This pull request directly merge into \"master\" branch, normally development happens on \"dev\" branch."
19
+ exit 1
.gitignore ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ __pycache__
2
+ *.ckpt
3
+ *.safetensors
4
+ *.pth
5
+ /ESRGAN/*
6
+ /SwinIR/*
7
+ /repositories
8
+ /venv
9
+ /tmp
10
+ /output
11
+ /model.ckpt
12
+ /models/**/*
13
+ /GFPGANv1.3.pth
14
+ /gfpgan/weights/*.pth
15
+ /ui-config.json
16
+ /outputs
17
+ /config.json
18
+ /log
19
+ /webui.settings.bat
20
+ /embeddings
21
+ /styles.csv
22
+ /params.txt
23
+ /styles.csv.bak
24
+ /webui-user.bat
25
+ /webui-user.sh
26
+ /interrogate
27
+ /user.css
28
+ /.idea
29
+ notification.mp3
30
+ /SwinIR
31
+ /textual_inversion
32
+ .vscode
33
+ /extensions
34
+ /test/stdout.txt
35
+ /test/stderr.txt
36
+ /cache.json*
37
+ /config_states/
38
+ /node_modules
39
+ /package-lock.json
40
+ /.coverage*
41
+ /test/test_outputs
42
+ /test/results.xml
43
+ coverage.xml
44
+ **/tests/**/expectations
.pylintrc ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ # See https://pylint.pycqa.org/en/latest/user_guide/messages/message_control.html
2
+ [MESSAGES CONTROL]
3
+ disable=C,R,W,E,I
CHANGELOG.md ADDED
@@ -0,0 +1,674 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ## 1.7.0
2
+
3
+ ### Features:
4
+ * settings tab rework: add search field, add categories, split UI settings page into many
5
+ * add altdiffusion-m18 support ([#13364](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13364))
6
+ * support inference with LyCORIS GLora networks ([#13610](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13610))
7
+ * add lora-embedding bundle system ([#13568](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13568))
8
+ * option to move prompt from top row into generation parameters
9
+ * add support for SSD-1B ([#13865](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13865))
10
+ * support inference with OFT networks ([#13692](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13692))
11
+ * script metadata and DAG sorting mechanism ([#13944](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13944))
12
+ * support HyperTile optimization ([#13948](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13948))
13
+ * add support for SD 2.1 Turbo ([#14170](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14170))
14
+ * remove Train->Preprocessing tab and put all its functionality into Extras tab
15
+ * initial IPEX support for Intel Arc GPU ([#14171](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14171))
16
+
17
+ ### Minor:
18
+ * allow reading model hash from images in img2img batch mode ([#12767](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12767))
19
+ * add option to align with sgm repo's sampling implementation ([#12818](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12818))
20
+ * extra field for lora metadata viewer: `ss_output_name` ([#12838](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12838))
21
+ * add action in settings page to calculate all SD checkpoint hashes ([#12909](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12909))
22
+ * add button to copy prompt to style editor ([#12975](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12975))
23
+ * add --skip-load-model-at-start option ([#13253](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13253))
24
+ * write infotext to gif images
25
+ * read infotext from gif images ([#13068](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13068))
26
+ * allow configuring the initial state of InputAccordion in ui-config.json ([#13189](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13189))
27
+ * allow editing whitespace delimiters for ctrl+up/ctrl+down prompt editing ([#13444](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13444))
28
+ * prevent accidentally closing popup dialogs ([#13480](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13480))
29
+ * added option to play notification sound or not ([#13631](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13631))
30
+ * show the preview image in the full screen image viewer if available ([#13459](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13459))
31
+ * support for webui.settings.bat ([#13638](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13638))
32
+ * add an option to not print stack traces on ctrl+c
33
+ * start/restart generation by Ctrl (Alt) + Enter ([#13644](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13644))
34
+ * update prompts_from_file script to allow concatenating entries with the general prompt ([#13733](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13733))
35
+ * added a visible checkbox to input accordion
36
+ * added an option to hide all txt2img/img2img parameters in an accordion ([#13826](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13826))
37
+ * added 'Path' sorting option for Extra network cards ([#13968](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13968))
38
+ * enable prompt hotkeys in style editor ([#13931](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13931))
39
+ * option to show batch img2img results in UI ([#14009](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14009))
40
+ * infotext updates: add option to disregard certain infotext fields, add option to not include VAE in infotext, add explanation to infotext settings page, move some options to infotext settings page
41
+ * add FP32 fallback support on sd_vae_approx ([#14046](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14046))
42
+ * support XYZ scripts / split hires path from unet ([#14126](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14126))
43
+ * allow use of mutiple styles csv files ([#14125](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14125))
44
+
45
+ ### Extensions and API:
46
+ * update gradio to 3.41.2
47
+ * support installed extensions list api ([#12774](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12774))
48
+ * update pnginfo API to return dict with parsed values
49
+ * add noisy latent to `ExtraNoiseParams` for callback ([#12856](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12856))
50
+ * show extension datetime in UTC ([#12864](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12864), [#12865](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12865), [#13281](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13281))
51
+ * add an option to choose how to combine hires fix and refiner
52
+ * include program version in info response. ([#13135](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13135))
53
+ * sd_unet support for SDXL
54
+ * patch DDPM.register_betas so that users can put given_betas in model yaml ([#13276](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13276))
55
+ * xyz_grid: add prepare ([#13266](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13266))
56
+ * allow multiple localization files with same language in extensions ([#13077](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13077))
57
+ * add onEdit function for js and rework token-counter.js to use it
58
+ * fix the key error exception when processing override_settings keys ([#13567](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13567))
59
+ * ability for extensions to return custom data via api in response.images ([#13463](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13463))
60
+ * call state.jobnext() before postproces*() ([#13762](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13762))
61
+ * add option to set notification sound volume ([#13884](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13884))
62
+ * update Ruff to 0.1.6 ([#14059](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14059))
63
+ * add Block component creation callback ([#14119](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14119))
64
+ * catch uncaught exception with ui creation scripts ([#14120](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14120))
65
+ * use extension name for determining an extension is installed in the index ([#14063](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14063))
66
+ * update is_installed() from launch_utils.py to fix reinstalling already installed packages ([#14192](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14192))
67
+
68
+ ### Bug Fixes:
69
+ * fix pix2pix producing bad results
70
+ * fix defaults settings page breaking when any of main UI tabs are hidden
71
+ * fix error that causes some extra networks to be disabled if both <lora:> and <lyco:> are present in the prompt
72
+ * fix for Reload UI function: if you reload UI on one tab, other opened tabs will no longer stop working
73
+ * prevent duplicate resize handler ([#12795](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12795))
74
+ * small typo: vae resolve bug ([#12797](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12797))
75
+ * hide broken image crop tool ([#12792](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12792))
76
+ * don't show hidden samplers in dropdown for XYZ script ([#12780](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12780))
77
+ * fix style editing dialog breaking if it's opened in both img2img and txt2img tabs
78
+ * hide --gradio-auth and --api-auth values from /internal/sysinfo report
79
+ * add missing infotext for RNG in options ([#12819](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12819))
80
+ * fix notification not playing when built-in webui tab is inactive ([#12834](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12834))
81
+ * honor `--skip-install` for extension installers ([#12832](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12832))
82
+ * don't print blank stdout in extension installers ([#12833](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12833), [#12855](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12855))
83
+ * get progressbar to display correctly in extensions tab
84
+ * keep order in list of checkpoints when loading model that doesn't have a checksum
85
+ * fix inpainting models in txt2img creating black pictures
86
+ * fix generation params regex ([#12876](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12876))
87
+ * fix batch img2img output dir with script ([#12926](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12926))
88
+ * fix #13080 - Hypernetwork/TI preview generation ([#13084](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13084))
89
+ * fix bug with sigma min/max overrides. ([#12995](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12995))
90
+ * more accurate check for enabling cuDNN benchmark on 16XX cards ([#12924](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12924))
91
+ * don't use multicond parser for negative prompt counter ([#13118](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13118))
92
+ * fix data-sort-name containing spaces ([#13412](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13412))
93
+ * update card on correct tab when editing metadata ([#13411](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13411))
94
+ * fix viewing/editing metadata when filename contains an apostrophe ([#13395](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13395))
95
+ * fix: --sd_model in "Prompts from file or textbox" script is not working ([#13302](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13302))
96
+ * better Support for Portable Git ([#13231](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13231))
97
+ * fix issues when webui_dir is not work_dir ([#13210](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13210))
98
+ * fix: lora-bias-backup don't reset cache ([#13178](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13178))
99
+ * account for customizable extra network separators whyen removing extra network text from the prompt ([#12877](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12877))
100
+ * re fix batch img2img output dir with script ([#13170](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13170))
101
+ * fix `--ckpt-dir` path separator and option use `short name` for checkpoint dropdown ([#13139](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13139))
102
+ * consolidated allowed preview formats, Fix extra network `.gif` not woking as preview ([#13121](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13121))
103
+ * fix venv_dir=- environment variable not working as expected on linux ([#13469](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13469))
104
+ * repair unload sd checkpoint button
105
+ * edit-attention fixes ([#13533](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13533))
106
+ * fix bug when using --gfpgan-models-path ([#13718](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13718))
107
+ * properly apply sort order for extra network cards when selected from dropdown
108
+ * fixes generation restart not working for some users when 'Ctrl+Enter' is pressed ([#13962](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13962))
109
+ * thread safe extra network list_items ([#13014](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13014))
110
+ * fix not able to exit metadata popup when pop up is too big ([#14156](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14156))
111
+ * fix auto focal point crop for opencv >= 4.8 ([#14121](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14121))
112
+ * make 'use-cpu all' actually apply to 'all' ([#14131](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14131))
113
+ * extras tab batch: actually use original filename
114
+ * make webui not crash when running with --disable-all-extensions option
115
+
116
+ ### Other:
117
+ * non-local condition ([#12814](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12814))
118
+ * fix minor typos ([#12827](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12827))
119
+ * remove xformers Python version check ([#12842](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12842))
120
+ * style: file-metadata word-break ([#12837](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12837))
121
+ * revert SGM noise multiplier change for img2img because it breaks hires fix
122
+ * do not change quicksettings dropdown option when value returned is `None` ([#12854](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12854))
123
+ * [RC 1.6.0 - zoom is partly hidden] Update style.css ([#12839](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12839))
124
+ * chore: change extension time format ([#12851](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12851))
125
+ * WEBUI.SH - Use torch 2.1.0 release candidate for Navi 3 ([#12929](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12929))
126
+ * add Fallback at images.read_info_from_image if exif data was invalid ([#13028](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13028))
127
+ * update cmd arg description ([#12986](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12986))
128
+ * fix: update shared.opts.data when add_option ([#12957](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12957), [#13213](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13213))
129
+ * restore missing tooltips ([#12976](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12976))
130
+ * use default dropdown padding on mobile ([#12880](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12880))
131
+ * put enable console prompts option into settings from commandline args ([#13119](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13119))
132
+ * fix some deprecated types ([#12846](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12846))
133
+ * bump to torchsde==0.2.6 ([#13418](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13418))
134
+ * update dragdrop.js ([#13372](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13372))
135
+ * use orderdict as lru cache:opt/bug ([#13313](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13313))
136
+ * XYZ if not include sub grids do not save sub grid ([#13282](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13282))
137
+ * initialize state.time_start befroe state.job_count ([#13229](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13229))
138
+ * fix fieldname regex ([#13458](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13458))
139
+ * change denoising_strength default to None. ([#13466](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13466))
140
+ * fix regression ([#13475](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13475))
141
+ * fix IndexError ([#13630](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13630))
142
+ * fix: checkpoints_loaded:{checkpoint:state_dict}, model.load_state_dict issue in dict value empty ([#13535](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13535))
143
+ * update bug_report.yml ([#12991](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12991))
144
+ * requirements_versions httpx==0.24.1 ([#13839](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13839))
145
+ * fix parenthesis auto selection ([#13829](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13829))
146
+ * fix #13796 ([#13797](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13797))
147
+ * corrected a typo in `modules/cmd_args.py` ([#13855](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13855))
148
+ * feat: fix randn found element of type float at pos 2 ([#14004](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14004))
149
+ * adds tqdm handler to logging_config.py for progress bar integration ([#13996](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13996))
150
+ * hotfix: call shared.state.end() after postprocessing done ([#13977](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13977))
151
+ * fix dependency address patch 1 ([#13929](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13929))
152
+ * save sysinfo as .json ([#14035](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14035))
153
+ * move exception_records related methods to errors.py ([#14084](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14084))
154
+ * compatibility ([#13936](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13936))
155
+ * json.dump(ensure_ascii=False) ([#14108](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14108))
156
+ * dir buttons start with / so only the correct dir will be shown and no… ([#13957](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13957))
157
+ * alternate implementation for unet forward replacement that does not depend on hijack being applied
158
+ * re-add `keyedit_delimiters_whitespace` setting lost as part of commit e294e46 ([#14178](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14178))
159
+ * fix `save_samples` being checked early when saving masked composite ([#14177](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14177))
160
+ * slight optimization for mask and mask_composite ([#14181](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14181))
161
+ * add import_hook hack to work around basicsr/torchvision incompatibility ([#14186](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14186))
162
+
163
+ ## 1.6.1
164
+
165
+ ### Bug Fixes:
166
+ * fix an error causing the webui to fail to start ([#13839](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13839))
167
+
168
+ ## 1.6.0
169
+
170
+ ### Features:
171
+ * refiner support [#12371](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12371)
172
+ * add NV option for Random number generator source setting, which allows to generate same pictures on CPU/AMD/Mac as on NVidia videocards
173
+ * add style editor dialog
174
+ * hires fix: add an option to use a different checkpoint for second pass ([#12181](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12181))
175
+ * option to keep multiple loaded models in memory ([#12227](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12227))
176
+ * new samplers: Restart, DPM++ 2M SDE Exponential, DPM++ 2M SDE Heun, DPM++ 2M SDE Heun Karras, DPM++ 2M SDE Heun Exponential, DPM++ 3M SDE, DPM++ 3M SDE Karras, DPM++ 3M SDE Exponential ([#12300](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12300), [#12519](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12519), [#12542](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12542))
177
+ * rework DDIM, PLMS, UniPC to use CFG denoiser same as in k-diffusion samplers:
178
+ * makes all of them work with img2img
179
+ * makes prompt composition posssible (AND)
180
+ * makes them available for SDXL
181
+ * always show extra networks tabs in the UI ([#11808](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/11808))
182
+ * use less RAM when creating models ([#11958](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/11958), [#12599](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12599))
183
+ * textual inversion inference support for SDXL
184
+ * extra networks UI: show metadata for SD checkpoints
185
+ * checkpoint merger: add metadata support
186
+ * prompt editing and attention: add support for whitespace after the number ([ red : green : 0.5 ]) (seed breaking change) ([#12177](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12177))
187
+ * VAE: allow selecting own VAE for each checkpoint (in user metadata editor)
188
+ * VAE: add selected VAE to infotext
189
+ * options in main UI: add own separate setting for txt2img and img2img, correctly read values from pasted infotext, add setting for column count ([#12551](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12551))
190
+ * add resize handle to txt2img and img2img tabs, allowing to change the amount of horizontable space given to generation parameters and resulting image gallery ([#12687](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12687), [#12723](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12723))
191
+ * change default behavior for batching cond/uncond -- now it's on by default, and is disabled by an UI setting (Optimizatios -> Batch cond/uncond) - if you are on lowvram/medvram and are getting OOM exceptions, you will need to enable it
192
+ * show current position in queue and make it so that requests are processed in the order of arrival ([#12707](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12707))
193
+ * add `--medvram-sdxl` flag that only enables `--medvram` for SDXL models
194
+ * prompt editing timeline has separate range for first pass and hires-fix pass (seed breaking change) ([#12457](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12457))
195
+
196
+ ### Minor:
197
+ * img2img batch: RAM savings, VRAM savings, .tif, .tiff in img2img batch ([#12120](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12120), [#12514](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12514), [#12515](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12515))
198
+ * postprocessing/extras: RAM savings ([#12479](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12479))
199
+ * XYZ: in the axis labels, remove pathnames from model filenames
200
+ * XYZ: support hires sampler ([#12298](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12298))
201
+ * XYZ: new option: use text inputs instead of dropdowns ([#12491](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12491))
202
+ * add gradio version warning
203
+ * sort list of VAE checkpoints ([#12297](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12297))
204
+ * use transparent white for mask in inpainting, along with an option to select the color ([#12326](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12326))
205
+ * move some settings to their own section: img2img, VAE
206
+ * add checkbox to show/hide dirs for extra networks
207
+ * Add TAESD(or more) options for all the VAE encode/decode operation ([#12311](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12311))
208
+ * gradio theme cache, new gradio themes, along with explanation that the user can input his own values ([#12346](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12346), [#12355](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12355))
209
+ * sampler fixes/tweaks: s_tmax, s_churn, s_noise, s_tmax ([#12354](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12354), [#12356](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12356), [#12357](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12357), [#12358](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12358), [#12375](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12375), [#12521](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12521))
210
+ * update README.md with correct instructions for Linux installation ([#12352](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12352))
211
+ * option to not save incomplete images, on by default ([#12338](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12338))
212
+ * enable cond cache by default
213
+ * git autofix for repos that are corrupted ([#12230](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12230))
214
+ * allow to open images in new browser tab by middle mouse button ([#12379](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12379))
215
+ * automatically open webui in browser when running "locally" ([#12254](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12254))
216
+ * put commonly used samplers on top, make DPM++ 2M Karras the default choice
217
+ * zoom and pan: option to auto-expand a wide image, improved integration ([#12413](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12413), [#12727](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12727))
218
+ * option to cache Lora networks in memory
219
+ * rework hires fix UI to use accordion
220
+ * face restoration and tiling moved to settings - use "Options in main UI" setting if you want them back
221
+ * change quicksettings items to have variable width
222
+ * Lora: add Norm module, add support for bias ([#12503](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12503))
223
+ * Lora: output warnings in UI rather than fail for unfitting loras; switch to logging for error output in console
224
+ * support search and display of hashes for all extra network items ([#12510](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12510))
225
+ * add extra noise param for img2img operations ([#12564](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12564))
226
+ * support for Lora with bias ([#12584](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12584))
227
+ * make interrupt quicker ([#12634](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12634))
228
+ * configurable gallery height ([#12648](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12648))
229
+ * make results column sticky ([#12645](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12645))
230
+ * more hash filename patterns ([#12639](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12639))
231
+ * make image viewer actually fit the whole page ([#12635](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12635))
232
+ * make progress bar work independently from live preview display which results in it being updated a lot more often
233
+ * forbid Full live preview method for medvram and add a setting to undo the forbidding
234
+ * make it possible to localize tooltips and placeholders
235
+ * add option to align with sgm repo's sampling implementation ([#12818](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12818))
236
+ * Restore faces and Tiling generation parameters have been moved to settings out of main UI
237
+ * if you want to put them back into main UI, use `Options in main UI` setting on the UI page.
238
+
239
+ ### Extensions and API:
240
+ * gradio 3.41.2
241
+ * also bump versions for packages: transformers, GitPython, accelerate, scikit-image, timm, tomesd
242
+ * support tooltip kwarg for gradio elements: gr.Textbox(label='hello', tooltip='world')
243
+ * properly clear the total console progressbar when using txt2img and img2img from API
244
+ * add cmd_arg --disable-extra-extensions and --disable-all-extensions ([#12294](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12294))
245
+ * shared.py and webui.py split into many files
246
+ * add --loglevel commandline argument for logging
247
+ * add a custom UI element that combines accordion and checkbox
248
+ * avoid importing gradio in tests because it spams warnings
249
+ * put infotext label for setting into OptionInfo definition rather than in a separate list
250
+ * make `StableDiffusionProcessingImg2Img.mask_blur` a property, make more inline with PIL `GaussianBlur` ([#12470](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12470))
251
+ * option to make scripts UI without gr.Group
252
+ * add a way for scripts to register a callback for before/after just a single component's creation
253
+ * use dataclass for StableDiffusionProcessing
254
+ * store patches for Lora in a specialized module instead of inside torch
255
+ * support http/https URLs in API ([#12663](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12663), [#12698](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12698))
256
+ * add extra noise callback ([#12616](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12616))
257
+ * dump current stack traces when exiting with SIGINT
258
+ * add type annotations for extra fields of shared.sd_model
259
+
260
+ ### Bug Fixes:
261
+ * Don't crash if out of local storage quota for javascriot localStorage
262
+ * XYZ plot do not fail if an exception occurs
263
+ * fix missing TI hash in infotext if generation uses both negative and positive TI ([#12269](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12269))
264
+ * localization fixes ([#12307](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12307))
265
+ * fix sdxl model invalid configuration after the hijack
266
+ * correctly toggle extras checkbox for infotext paste ([#12304](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12304))
267
+ * open raw sysinfo link in new page ([#12318](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12318))
268
+ * prompt parser: Account for empty field in alternating words syntax ([#12319](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12319))
269
+ * add tab and carriage return to invalid filename chars ([#12327](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12327))
270
+ * fix api only Lora not working ([#12387](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12387))
271
+ * fix options in main UI misbehaving when there's just one element
272
+ * make it possible to use a sampler from infotext even if it's hidden in the dropdown
273
+ * fix styles missing from the prompt in infotext when making a grid of batch of multiplie images
274
+ * prevent bogus progress output in console when calculating hires fix dimensions
275
+ * fix --use-textbox-seed
276
+ * fix broken `Lora/Networks: use old method` option ([#12466](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12466))
277
+ * properly return `None` for VAE hash when using `--no-hashing` ([#12463](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12463))
278
+ * MPS/macOS fixes and optimizations ([#12526](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12526))
279
+ * add second_order to samplers that mistakenly didn't have it
280
+ * when refreshing cards in extra networks UI, do not discard user's custom resolution
281
+ * fix processing error that happens if batch_size is not a multiple of how many prompts/negative prompts there are ([#12509](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12509))
282
+ * fix inpaint upload for alpha masks ([#12588](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12588))
283
+ * fix exception when image sizes are not integers ([#12586](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12586))
284
+ * fix incorrect TAESD Latent scale ([#12596](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12596))
285
+ * auto add data-dir to gradio-allowed-path ([#12603](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12603))
286
+ * fix exception if extensuions dir is missing ([#12607](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12607))
287
+ * fix issues with api model-refresh and vae-refresh ([#12638](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12638))
288
+ * fix img2img background color for transparent images option not being used ([#12633](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12633))
289
+ * attempt to resolve NaN issue with unstable VAEs in fp32 mk2 ([#12630](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12630))
290
+ * implement missing undo hijack for SDXL
291
+ * fix xyz swap axes ([#12684](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12684))
292
+ * fix errors in backup/restore tab if any of config files are broken ([#12689](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12689))
293
+ * fix SD VAE switch error after model reuse ([#12685](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12685))
294
+ * fix trying to create images too large for the chosen format ([#12667](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12667))
295
+ * create Gradio temp directory if necessary ([#12717](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12717))
296
+ * prevent possible cache loss if exiting as it's being written by using an atomic operation to replace the cache with the new version
297
+ * set devices.dtype_unet correctly
298
+ * run RealESRGAN on GPU for non-CUDA devices ([#12737](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12737))
299
+ * prevent extra network buttons being obscured by description for very small card sizes ([#12745](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12745))
300
+ * fix error that causes some extra networks to be disabled if both <lora:> and <lyco:> are present in the prompt
301
+ * fix defaults settings page breaking when any of main UI tabs are hidden
302
+ * fix incorrect save/display of new values in Defaults page in settings
303
+ * fix for Reload UI function: if you reload UI on one tab, other opened tabs will no longer stop working
304
+ * fix an error that prevents VAE being reloaded after an option change if a VAE near the checkpoint exists ([#12797](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12737))
305
+ * hide broken image crop tool ([#12792](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12737))
306
+ * don't show hidden samplers in dropdown for XYZ script ([#12780](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12737))
307
+ * fix style editing dialog breaking if it's opened in both img2img and txt2img tabs
308
+ * fix a bug allowing users to bypass gradio and API authentication (reported by vysecurity)
309
+ * fix notification not playing when built-in webui tab is inactive ([#12834](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12834))
310
+ * honor `--skip-install` for extension installers ([#12832](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12832))
311
+ * don't print blank stdout in extension installers ([#12833](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12832), [#12855](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12855))
312
+ * do not change quicksettings dropdown option when value returned is `None` ([#12854](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12854))
313
+ * get progressbar to display correctly in extensions tab
314
+
315
+
316
+ ## 1.5.2
317
+
318
+ ### Bug Fixes:
319
+ * fix memory leak when generation fails
320
+ * update doggettx cross attention optimization to not use an unreasonable amount of memory in some edge cases -- suggestion by MorkTheOrk
321
+
322
+
323
+ ## 1.5.1
324
+
325
+ ### Minor:
326
+ * support parsing text encoder blocks in some new LoRAs
327
+ * delete scale checker script due to user demand
328
+
329
+ ### Extensions and API:
330
+ * add postprocess_batch_list script callback
331
+
332
+ ### Bug Fixes:
333
+ * fix TI training for SD1
334
+ * fix reload altclip model error
335
+ * prepend the pythonpath instead of overriding it
336
+ * fix typo in SD_WEBUI_RESTARTING
337
+ * if txt2img/img2img raises an exception, finally call state.end()
338
+ * fix composable diffusion weight parsing
339
+ * restyle Startup profile for black users
340
+ * fix webui not launching with --nowebui
341
+ * catch exception for non git extensions
342
+ * fix some options missing from /sdapi/v1/options
343
+ * fix for extension update status always saying "unknown"
344
+ * fix display of extra network cards that have `<>` in the name
345
+ * update lora extension to work with python 3.8
346
+
347
+
348
+ ## 1.5.0
349
+
350
+ ### Features:
351
+ * SD XL support
352
+ * user metadata system for custom networks
353
+ * extended Lora metadata editor: set activation text, default weight, view tags, training info
354
+ * Lora extension rework to include other types of networks (all that were previously handled by LyCORIS extension)
355
+ * show github stars for extenstions
356
+ * img2img batch mode can read extra stuff from png info
357
+ * img2img batch works with subdirectories
358
+ * hotkeys to move prompt elements: alt+left/right
359
+ * restyle time taken/VRAM display
360
+ * add textual inversion hashes to infotext
361
+ * optimization: cache git extension repo information
362
+ * move generate button next to the generated picture for mobile clients
363
+ * hide cards for networks of incompatible Stable Diffusion version in Lora extra networks interface
364
+ * skip installing packages with pip if they all are already installed - startup speedup of about 2 seconds
365
+
366
+ ### Minor:
367
+ * checkbox to check/uncheck all extensions in the Installed tab
368
+ * add gradio user to infotext and to filename patterns
369
+ * allow gif for extra network previews
370
+ * add options to change colors in grid
371
+ * use natural sort for items in extra networks
372
+ * Mac: use empty_cache() from torch 2 to clear VRAM
373
+ * added automatic support for installing the right libraries for Navi3 (AMD)
374
+ * add option SWIN_torch_compile to accelerate SwinIR upscale
375
+ * suppress printing TI embedding info at start to console by default
376
+ * speedup extra networks listing
377
+ * added `[none]` filename token.
378
+ * removed thumbs extra networks view mode (use settings tab to change width/height/scale to get thumbs)
379
+ * add always_discard_next_to_last_sigma option to XYZ plot
380
+ * automatically switch to 32-bit float VAE if the generated picture has NaNs without the need for `--no-half-vae` commandline flag.
381
+
382
+ ### Extensions and API:
383
+ * api endpoints: /sdapi/v1/server-kill, /sdapi/v1/server-restart, /sdapi/v1/server-stop
384
+ * allow Script to have custom metaclass
385
+ * add model exists status check /sdapi/v1/options
386
+ * rename --add-stop-route to --api-server-stop
387
+ * add `before_hr` script callback
388
+ * add callback `after_extra_networks_activate`
389
+ * disable rich exception output in console for API by default, use WEBUI_RICH_EXCEPTIONS env var to enable
390
+ * return http 404 when thumb file not found
391
+ * allow replacing extensions index with environment variable
392
+
393
+ ### Bug Fixes:
394
+ * fix for catch errors when retrieving extension index #11290
395
+ * fix very slow loading speed of .safetensors files when reading from network drives
396
+ * API cache cleanup
397
+ * fix UnicodeEncodeError when writing to file CLIP Interrogator batch mode
398
+ * fix warning of 'has_mps' deprecated from PyTorch
399
+ * fix problem with extra network saving images as previews losing generation info
400
+ * fix throwing exception when trying to resize image with I;16 mode
401
+ * fix for #11534: canvas zoom and pan extension hijacking shortcut keys
402
+ * fixed launch script to be runnable from any directory
403
+ * don't add "Seed Resize: -1x-1" to API image metadata
404
+ * correctly remove end parenthesis with ctrl+up/down
405
+ * fixing --subpath on newer gradio version
406
+ * fix: check fill size none zero when resize (fixes #11425)
407
+ * use submit and blur for quick settings textbox
408
+ * save img2img batch with images.save_image()
409
+ * prevent running preload.py for disabled extensions
410
+ * fix: previously, model name was added together with directory name to infotext and to [model_name] filename pattern; directory name is now not included
411
+
412
+
413
+ ## 1.4.1
414
+
415
+ ### Bug Fixes:
416
+ * add queue lock for refresh-checkpoints
417
+
418
+ ## 1.4.0
419
+
420
+ ### Features:
421
+ * zoom controls for inpainting
422
+ * run basic torch calculation at startup in parallel to reduce the performance impact of first generation
423
+ * option to pad prompt/neg prompt to be same length
424
+ * remove taming_transformers dependency
425
+ * custom k-diffusion scheduler settings
426
+ * add an option to show selected settings in main txt2img/img2img UI
427
+ * sysinfo tab in settings
428
+ * infer styles from prompts when pasting params into the UI
429
+ * an option to control the behavior of the above
430
+
431
+ ### Minor:
432
+ * bump Gradio to 3.32.0
433
+ * bump xformers to 0.0.20
434
+ * Add option to disable token counters
435
+ * tooltip fixes & optimizations
436
+ * make it possible to configure filename for the zip download
437
+ * `[vae_filename]` pattern for filenames
438
+ * Revert discarding penultimate sigma for DPM-Solver++(2M) SDE
439
+ * change UI reorder setting to multiselect
440
+ * read version info form CHANGELOG.md if git version info is not available
441
+ * link footer API to Wiki when API is not active
442
+ * persistent conds cache (opt-in optimization)
443
+
444
+ ### Extensions:
445
+ * After installing extensions, webui properly restarts the process rather than reloads the UI
446
+ * Added VAE listing to web API. Via: /sdapi/v1/sd-vae
447
+ * custom unet support
448
+ * Add onAfterUiUpdate callback
449
+ * refactor EmbeddingDatabase.register_embedding() to allow unregistering
450
+ * add before_process callback for scripts
451
+ * add ability for alwayson scripts to specify section and let user reorder those sections
452
+
453
+ ### Bug Fixes:
454
+ * Fix dragging text to prompt
455
+ * fix incorrect quoting for infotext values with colon in them
456
+ * fix "hires. fix" prompt sharing same labels with txt2img_prompt
457
+ * Fix s_min_uncond default type int
458
+ * Fix for #10643 (Inpainting mask sometimes not working)
459
+ * fix bad styling for thumbs view in extra networks #10639
460
+ * fix for empty list of optimizations #10605
461
+ * small fixes to prepare_tcmalloc for Debian/Ubuntu compatibility
462
+ * fix --ui-debug-mode exit
463
+ * patch GitPython to not use leaky persistent processes
464
+ * fix duplicate Cross attention optimization after UI reload
465
+ * torch.cuda.is_available() check for SdOptimizationXformers
466
+ * fix hires fix using wrong conds in second pass if using Loras.
467
+ * handle exception when parsing generation parameters from png info
468
+ * fix upcast attention dtype error
469
+ * forcing Torch Version to 1.13.1 for RX 5000 series GPUs
470
+ * split mask blur into X and Y components, patch Outpainting MK2 accordingly
471
+ * don't die when a LoRA is a broken symlink
472
+ * allow activation of Generate Forever during generation
473
+
474
+
475
+ ## 1.3.2
476
+
477
+ ### Bug Fixes:
478
+ * fix files served out of tmp directory even if they are saved to disk
479
+ * fix postprocessing overwriting parameters
480
+
481
+ ## 1.3.1
482
+
483
+ ### Features:
484
+ * revert default cross attention optimization to Doggettx
485
+
486
+ ### Bug Fixes:
487
+ * fix bug: LoRA don't apply on dropdown list sd_lora
488
+ * fix png info always added even if setting is not enabled
489
+ * fix some fields not applying in xyz plot
490
+ * fix "hires. fix" prompt sharing same labels with txt2img_prompt
491
+ * fix lora hashes not being added properly to infotex if there is only one lora
492
+ * fix --use-cpu failing to work properly at startup
493
+ * make --disable-opt-split-attention command line option work again
494
+
495
+ ## 1.3.0
496
+
497
+ ### Features:
498
+ * add UI to edit defaults
499
+ * token merging (via dbolya/tomesd)
500
+ * settings tab rework: add a lot of additional explanations and links
501
+ * load extensions' Git metadata in parallel to loading the main program to save a ton of time during startup
502
+ * update extensions table: show branch, show date in separate column, and show version from tags if available
503
+ * TAESD - another option for cheap live previews
504
+ * allow choosing sampler and prompts for second pass of hires fix - hidden by default, enabled in settings
505
+ * calculate hashes for Lora
506
+ * add lora hashes to infotext
507
+ * when pasting infotext, use infotext's lora hashes to find local loras for `<lora:xxx:1>` entries whose hashes match loras the user has
508
+ * select cross attention optimization from UI
509
+
510
+ ### Minor:
511
+ * bump Gradio to 3.31.0
512
+ * bump PyTorch to 2.0.1 for macOS and Linux AMD
513
+ * allow setting defaults for elements in extensions' tabs
514
+ * allow selecting file type for live previews
515
+ * show "Loading..." for extra networks when displaying for the first time
516
+ * suppress ENSD infotext for samplers that don't use it
517
+ * clientside optimizations
518
+ * add options to show/hide hidden files and dirs in extra networks, and to not list models/files in hidden directories
519
+ * allow whitespace in styles.csv
520
+ * add option to reorder tabs
521
+ * move some functionality (swap resolution and set seed to -1) to client
522
+ * option to specify editor height for img2img
523
+ * button to copy image resolution into img2img width/height sliders
524
+ * switch from pyngrok to ngrok-py
525
+ * lazy-load images in extra networks UI
526
+ * set "Navigate image viewer with gamepad" option to false by default, by request
527
+ * change upscalers to download models into user-specified directory (from commandline args) rather than the default models/<...>
528
+ * allow hiding buttons in ui-config.json
529
+
530
+ ### Extensions:
531
+ * add /sdapi/v1/script-info api
532
+ * use Ruff to lint Python code
533
+ * use ESlint to lint Javascript code
534
+ * add/modify CFG callbacks for Self-Attention Guidance extension
535
+ * add command and endpoint for graceful server stopping
536
+ * add some locals (prompts/seeds/etc) from processing function into the Processing class as fields
537
+ * rework quoting for infotext items that have commas in them to use JSON (should be backwards compatible except for cases where it didn't work previously)
538
+ * add /sdapi/v1/refresh-loras api checkpoint post request
539
+ * tests overhaul
540
+
541
+ ### Bug Fixes:
542
+ * fix an issue preventing the program from starting if the user specifies a bad Gradio theme
543
+ * fix broken prompts from file script
544
+ * fix symlink scanning for extra networks
545
+ * fix --data-dir ignored when launching via webui-user.bat COMMANDLINE_ARGS
546
+ * allow web UI to be ran fully offline
547
+ * fix inability to run with --freeze-settings
548
+ * fix inability to merge checkpoint without adding metadata
549
+ * fix extra networks' save preview image not adding infotext for jpeg/webm
550
+ * remove blinking effect from text in hires fix and scale resolution preview
551
+ * make links to `http://<...>.git` extensions work in the extension tab
552
+ * fix bug with webui hanging at startup due to hanging git process
553
+
554
+
555
+ ## 1.2.1
556
+
557
+ ### Features:
558
+ * add an option to always refer to LoRA by filenames
559
+
560
+ ### Bug Fixes:
561
+ * never refer to LoRA by an alias if multiple LoRAs have same alias or the alias is called none
562
+ * fix upscalers disappearing after the user reloads UI
563
+ * allow bf16 in safe unpickler (resolves problems with loading some LoRAs)
564
+ * allow web UI to be ran fully offline
565
+ * fix localizations not working
566
+ * fix error for LoRAs: `'LatentDiffusion' object has no attribute 'lora_layer_mapping'`
567
+
568
+ ## 1.2.0
569
+
570
+ ### Features:
571
+ * do not wait for Stable Diffusion model to load at startup
572
+ * add filename patterns: `[denoising]`
573
+ * directory hiding for extra networks: dirs starting with `.` will hide their cards on extra network tabs unless specifically searched for
574
+ * LoRA: for the `<...>` text in prompt, use name of LoRA that is in the metdata of the file, if present, instead of filename (both can be used to activate LoRA)
575
+ * LoRA: read infotext params from kohya-ss's extension parameters if they are present and if his extension is not active
576
+ * LoRA: fix some LoRAs not working (ones that have 3x3 convolution layer)
577
+ * LoRA: add an option to use old method of applying LoRAs (producing same results as with kohya-ss)
578
+ * add version to infotext, footer and console output when starting
579
+ * add links to wiki for filename pattern settings
580
+ * add extended info for quicksettings setting and use multiselect input instead of a text field
581
+
582
+ ### Minor:
583
+ * bump Gradio to 3.29.0
584
+ * bump PyTorch to 2.0.1
585
+ * `--subpath` option for gradio for use with reverse proxy
586
+ * Linux/macOS: use existing virtualenv if already active (the VIRTUAL_ENV environment variable)
587
+ * do not apply localizations if there are none (possible frontend optimization)
588
+ * add extra `None` option for VAE in XYZ plot
589
+ * print error to console when batch processing in img2img fails
590
+ * create HTML for extra network pages only on demand
591
+ * allow directories starting with `.` to still list their models for LoRA, checkpoints, etc
592
+ * put infotext options into their own category in settings tab
593
+ * do not show licenses page when user selects Show all pages in settings
594
+
595
+ ### Extensions:
596
+ * tooltip localization support
597
+ * add API method to get LoRA models with prompt
598
+
599
+ ### Bug Fixes:
600
+ * re-add `/docs` endpoint
601
+ * fix gamepad navigation
602
+ * make the lightbox fullscreen image function properly
603
+ * fix squished thumbnails in extras tab
604
+ * keep "search" filter for extra networks when user refreshes the tab (previously it showed everthing after you refreshed)
605
+ * fix webui showing the same image if you configure the generation to always save results into same file
606
+ * fix bug with upscalers not working properly
607
+ * fix MPS on PyTorch 2.0.1, Intel Macs
608
+ * make it so that custom context menu from contextMenu.js only disappears after user's click, ignoring non-user click events
609
+ * prevent Reload UI button/link from reloading the page when it's not yet ready
610
+ * fix prompts from file script failing to read contents from a drag/drop file
611
+
612
+
613
+ ## 1.1.1
614
+ ### Bug Fixes:
615
+ * fix an error that prevents running webui on PyTorch<2.0 without --disable-safe-unpickle
616
+
617
+ ## 1.1.0
618
+ ### Features:
619
+ * switch to PyTorch 2.0.0 (except for AMD GPUs)
620
+ * visual improvements to custom code scripts
621
+ * add filename patterns: `[clip_skip]`, `[hasprompt<>]`, `[batch_number]`, `[generation_number]`
622
+ * add support for saving init images in img2img, and record their hashes in infotext for reproducability
623
+ * automatically select current word when adjusting weight with ctrl+up/down
624
+ * add dropdowns for X/Y/Z plot
625
+ * add setting: Stable Diffusion/Random number generator source: makes it possible to make images generated from a given manual seed consistent across different GPUs
626
+ * support Gradio's theme API
627
+ * use TCMalloc on Linux by default; possible fix for memory leaks
628
+ * add optimization option to remove negative conditioning at low sigma values #9177
629
+ * embed model merge metadata in .safetensors file
630
+ * extension settings backup/restore feature #9169
631
+ * add "resize by" and "resize to" tabs to img2img
632
+ * add option "keep original size" to textual inversion images preprocess
633
+ * image viewer scrolling via analog stick
634
+ * button to restore the progress from session lost / tab reload
635
+
636
+ ### Minor:
637
+ * bump Gradio to 3.28.1
638
+ * change "scale to" to sliders in Extras tab
639
+ * add labels to tool buttons to make it possible to hide them
640
+ * add tiled inference support for ScuNET
641
+ * add branch support for extension installation
642
+ * change Linux installation script to install into current directory rather than `/home/username`
643
+ * sort textual inversion embeddings by name (case-insensitive)
644
+ * allow styles.csv to be symlinked or mounted in docker
645
+ * remove the "do not add watermark to images" option
646
+ * make selected tab configurable with UI config
647
+ * make the extra networks UI fixed height and scrollable
648
+ * add `disable_tls_verify` arg for use with self-signed certs
649
+
650
+ ### Extensions:
651
+ * add reload callback
652
+ * add `is_hr_pass` field for processing
653
+
654
+ ### Bug Fixes:
655
+ * fix broken batch image processing on 'Extras/Batch Process' tab
656
+ * add "None" option to extra networks dropdowns
657
+ * fix FileExistsError for CLIP Interrogator
658
+ * fix /sdapi/v1/txt2img endpoint not working on Linux #9319
659
+ * fix disappearing live previews and progressbar during slow tasks
660
+ * fix fullscreen image view not working properly in some cases
661
+ * prevent alwayson_scripts args param resizing script_arg list when they are inserted in it
662
+ * fix prompt schedule for second order samplers
663
+ * fix image mask/composite for weird resolutions #9628
664
+ * use correct images for previews when using AND (see #9491)
665
+ * one broken image in img2img batch won't stop all processing
666
+ * fix image orientation bug in train/preprocess
667
+ * fix Ngrok recreating tunnels every reload
668
+ * fix `--realesrgan-models-path` and `--ldsr-models-path` not working
669
+ * fix `--skip-install` not working
670
+ * use SAMPLE file format in Outpainting Mk2 & Poorman
671
+ * do not fail all LoRAs if some have failed to load when making a picture
672
+
673
+ ## 1.0.0
674
+ * everything
CITATION.cff ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ cff-version: 1.2.0
2
+ message: "If you use this software, please cite it as below."
3
+ authors:
4
+ - given-names: AUTOMATIC1111
5
+ title: "Stable Diffusion Web UI"
6
+ date-released: 2022-08-22
7
+ url: "https://github.com/AUTOMATIC1111/stable-diffusion-webui"
CODEOWNERS ADDED
@@ -0,0 +1 @@
 
 
1
+ * @lllyasviel
LICENSE.txt ADDED
@@ -0,0 +1,663 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ GNU AFFERO GENERAL PUBLIC LICENSE
2
+ Version 3, 19 November 2007
3
+
4
+ Copyright (c) 2023 AUTOMATIC1111
5
+
6
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
7
+ Everyone is permitted to copy and distribute verbatim copies
8
+ of this license document, but changing it is not allowed.
9
+
10
+ Preamble
11
+
12
+ The GNU Affero General Public License is a free, copyleft license for
13
+ software and other kinds of works, specifically designed to ensure
14
+ cooperation with the community in the case of network server software.
15
+
16
+ The licenses for most software and other practical works are designed
17
+ to take away your freedom to share and change the works. By contrast,
18
+ our General Public Licenses are intended to guarantee your freedom to
19
+ share and change all versions of a program--to make sure it remains free
20
+ software for all its users.
21
+
22
+ When we speak of free software, we are referring to freedom, not
23
+ price. Our General Public Licenses are designed to make sure that you
24
+ have the freedom to distribute copies of free software (and charge for
25
+ them if you wish), that you receive source code or can get it if you
26
+ want it, that you can change the software or use pieces of it in new
27
+ free programs, and that you know you can do these things.
28
+
29
+ Developers that use our General Public Licenses protect your rights
30
+ with two steps: (1) assert copyright on the software, and (2) offer
31
+ you this License which gives you legal permission to copy, distribute
32
+ and/or modify the software.
33
+
34
+ A secondary benefit of defending all users' freedom is that
35
+ improvements made in alternate versions of the program, if they
36
+ receive widespread use, become available for other developers to
37
+ incorporate. Many developers of free software are heartened and
38
+ encouraged by the resulting cooperation. However, in the case of
39
+ software used on network servers, this result may fail to come about.
40
+ The GNU General Public License permits making a modified version and
41
+ letting the public access it on a server without ever releasing its
42
+ source code to the public.
43
+
44
+ The GNU Affero General Public License is designed specifically to
45
+ ensure that, in such cases, the modified source code becomes available
46
+ to the community. It requires the operator of a network server to
47
+ provide the source code of the modified version running there to the
48
+ users of that server. Therefore, public use of a modified version, on
49
+ a publicly accessible server, gives the public access to the source
50
+ code of the modified version.
51
+
52
+ An older license, called the Affero General Public License and
53
+ published by Affero, was designed to accomplish similar goals. This is
54
+ a different license, not a version of the Affero GPL, but Affero has
55
+ released a new version of the Affero GPL which permits relicensing under
56
+ this license.
57
+
58
+ The precise terms and conditions for copying, distribution and
59
+ modification follow.
60
+
61
+ TERMS AND CONDITIONS
62
+
63
+ 0. Definitions.
64
+
65
+ "This License" refers to version 3 of the GNU Affero General Public License.
66
+
67
+ "Copyright" also means copyright-like laws that apply to other kinds of
68
+ works, such as semiconductor masks.
69
+
70
+ "The Program" refers to any copyrightable work licensed under this
71
+ License. Each licensee is addressed as "you". "Licensees" and
72
+ "recipients" may be individuals or organizations.
73
+
74
+ To "modify" a work means to copy from or adapt all or part of the work
75
+ in a fashion requiring copyright permission, other than the making of an
76
+ exact copy. The resulting work is called a "modified version" of the
77
+ earlier work or a work "based on" the earlier work.
78
+
79
+ A "covered work" means either the unmodified Program or a work based
80
+ on the Program.
81
+
82
+ To "propagate" a work means to do anything with it that, without
83
+ permission, would make you directly or secondarily liable for
84
+ infringement under applicable copyright law, except executing it on a
85
+ computer or modifying a private copy. Propagation includes copying,
86
+ distribution (with or without modification), making available to the
87
+ public, and in some countries other activities as well.
88
+
89
+ To "convey" a work means any kind of propagation that enables other
90
+ parties to make or receive copies. Mere interaction with a user through
91
+ a computer network, with no transfer of a copy, is not conveying.
92
+
93
+ An interactive user interface displays "Appropriate Legal Notices"
94
+ to the extent that it includes a convenient and prominently visible
95
+ feature that (1) displays an appropriate copyright notice, and (2)
96
+ tells the user that there is no warranty for the work (except to the
97
+ extent that warranties are provided), that licensees may convey the
98
+ work under this License, and how to view a copy of this License. If
99
+ the interface presents a list of user commands or options, such as a
100
+ menu, a prominent item in the list meets this criterion.
101
+
102
+ 1. Source Code.
103
+
104
+ The "source code" for a work means the preferred form of the work
105
+ for making modifications to it. "Object code" means any non-source
106
+ form of a work.
107
+
108
+ A "Standard Interface" means an interface that either is an official
109
+ standard defined by a recognized standards body, or, in the case of
110
+ interfaces specified for a particular programming language, one that
111
+ is widely used among developers working in that language.
112
+
113
+ The "System Libraries" of an executable work include anything, other
114
+ than the work as a whole, that (a) is included in the normal form of
115
+ packaging a Major Component, but which is not part of that Major
116
+ Component, and (b) serves only to enable use of the work with that
117
+ Major Component, or to implement a Standard Interface for which an
118
+ implementation is available to the public in source code form. A
119
+ "Major Component", in this context, means a major essential component
120
+ (kernel, window system, and so on) of the specific operating system
121
+ (if any) on which the executable work runs, or a compiler used to
122
+ produce the work, or an object code interpreter used to run it.
123
+
124
+ The "Corresponding Source" for a work in object code form means all
125
+ the source code needed to generate, install, and (for an executable
126
+ work) run the object code and to modify the work, including scripts to
127
+ control those activities. However, it does not include the work's
128
+ System Libraries, or general-purpose tools or generally available free
129
+ programs which are used unmodified in performing those activities but
130
+ which are not part of the work. For example, Corresponding Source
131
+ includes interface definition files associated with source files for
132
+ the work, and the source code for shared libraries and dynamically
133
+ linked subprograms that the work is specifically designed to require,
134
+ such as by intimate data communication or control flow between those
135
+ subprograms and other parts of the work.
136
+
137
+ The Corresponding Source need not include anything that users
138
+ can regenerate automatically from other parts of the Corresponding
139
+ Source.
140
+
141
+ The Corresponding Source for a work in source code form is that
142
+ same work.
143
+
144
+ 2. Basic Permissions.
145
+
146
+ All rights granted under this License are granted for the term of
147
+ copyright on the Program, and are irrevocable provided the stated
148
+ conditions are met. This License explicitly affirms your unlimited
149
+ permission to run the unmodified Program. The output from running a
150
+ covered work is covered by this License only if the output, given its
151
+ content, constitutes a covered work. This License acknowledges your
152
+ rights of fair use or other equivalent, as provided by copyright law.
153
+
154
+ You may make, run and propagate covered works that you do not
155
+ convey, without conditions so long as your license otherwise remains
156
+ in force. You may convey covered works to others for the sole purpose
157
+ of having them make modifications exclusively for you, or provide you
158
+ with facilities for running those works, provided that you comply with
159
+ the terms of this License in conveying all material for which you do
160
+ not control copyright. Those thus making or running the covered works
161
+ for you must do so exclusively on your behalf, under your direction
162
+ and control, on terms that prohibit them from making any copies of
163
+ your copyrighted material outside their relationship with you.
164
+
165
+ Conveying under any other circumstances is permitted solely under
166
+ the conditions stated below. Sublicensing is not allowed; section 10
167
+ makes it unnecessary.
168
+
169
+ 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
170
+
171
+ No covered work shall be deemed part of an effective technological
172
+ measure under any applicable law fulfilling obligations under article
173
+ 11 of the WIPO copyright treaty adopted on 20 December 1996, or
174
+ similar laws prohibiting or restricting circumvention of such
175
+ measures.
176
+
177
+ When you convey a covered work, you waive any legal power to forbid
178
+ circumvention of technological measures to the extent such circumvention
179
+ is effected by exercising rights under this License with respect to
180
+ the covered work, and you disclaim any intention to limit operation or
181
+ modification of the work as a means of enforcing, against the work's
182
+ users, your or third parties' legal rights to forbid circumvention of
183
+ technological measures.
184
+
185
+ 4. Conveying Verbatim Copies.
186
+
187
+ You may convey verbatim copies of the Program's source code as you
188
+ receive it, in any medium, provided that you conspicuously and
189
+ appropriately publish on each copy an appropriate copyright notice;
190
+ keep intact all notices stating that this License and any
191
+ non-permissive terms added in accord with section 7 apply to the code;
192
+ keep intact all notices of the absence of any warranty; and give all
193
+ recipients a copy of this License along with the Program.
194
+
195
+ You may charge any price or no price for each copy that you convey,
196
+ and you may offer support or warranty protection for a fee.
197
+
198
+ 5. Conveying Modified Source Versions.
199
+
200
+ You may convey a work based on the Program, or the modifications to
201
+ produce it from the Program, in the form of source code under the
202
+ terms of section 4, provided that you also meet all of these conditions:
203
+
204
+ a) The work must carry prominent notices stating that you modified
205
+ it, and giving a relevant date.
206
+
207
+ b) The work must carry prominent notices stating that it is
208
+ released under this License and any conditions added under section
209
+ 7. This requirement modifies the requirement in section 4 to
210
+ "keep intact all notices".
211
+
212
+ c) You must license the entire work, as a whole, under this
213
+ License to anyone who comes into possession of a copy. This
214
+ License will therefore apply, along with any applicable section 7
215
+ additional terms, to the whole of the work, and all its parts,
216
+ regardless of how they are packaged. This License gives no
217
+ permission to license the work in any other way, but it does not
218
+ invalidate such permission if you have separately received it.
219
+
220
+ d) If the work has interactive user interfaces, each must display
221
+ Appropriate Legal Notices; however, if the Program has interactive
222
+ interfaces that do not display Appropriate Legal Notices, your
223
+ work need not make them do so.
224
+
225
+ A compilation of a covered work with other separate and independent
226
+ works, which are not by their nature extensions of the covered work,
227
+ and which are not combined with it such as to form a larger program,
228
+ in or on a volume of a storage or distribution medium, is called an
229
+ "aggregate" if the compilation and its resulting copyright are not
230
+ used to limit the access or legal rights of the compilation's users
231
+ beyond what the individual works permit. Inclusion of a covered work
232
+ in an aggregate does not cause this License to apply to the other
233
+ parts of the aggregate.
234
+
235
+ 6. Conveying Non-Source Forms.
236
+
237
+ You may convey a covered work in object code form under the terms
238
+ of sections 4 and 5, provided that you also convey the
239
+ machine-readable Corresponding Source under the terms of this License,
240
+ in one of these ways:
241
+
242
+ a) Convey the object code in, or embodied in, a physical product
243
+ (including a physical distribution medium), accompanied by the
244
+ Corresponding Source fixed on a durable physical medium
245
+ customarily used for software interchange.
246
+
247
+ b) Convey the object code in, or embodied in, a physical product
248
+ (including a physical distribution medium), accompanied by a
249
+ written offer, valid for at least three years and valid for as
250
+ long as you offer spare parts or customer support for that product
251
+ model, to give anyone who possesses the object code either (1) a
252
+ copy of the Corresponding Source for all the software in the
253
+ product that is covered by this License, on a durable physical
254
+ medium customarily used for software interchange, for a price no
255
+ more than your reasonable cost of physically performing this
256
+ conveying of source, or (2) access to copy the
257
+ Corresponding Source from a network server at no charge.
258
+
259
+ c) Convey individual copies of the object code with a copy of the
260
+ written offer to provide the Corresponding Source. This
261
+ alternative is allowed only occasionally and noncommercially, and
262
+ only if you received the object code with such an offer, in accord
263
+ with subsection 6b.
264
+
265
+ d) Convey the object code by offering access from a designated
266
+ place (gratis or for a charge), and offer equivalent access to the
267
+ Corresponding Source in the same way through the same place at no
268
+ further charge. You need not require recipients to copy the
269
+ Corresponding Source along with the object code. If the place to
270
+ copy the object code is a network server, the Corresponding Source
271
+ may be on a different server (operated by you or a third party)
272
+ that supports equivalent copying facilities, provided you maintain
273
+ clear directions next to the object code saying where to find the
274
+ Corresponding Source. Regardless of what server hosts the
275
+ Corresponding Source, you remain obligated to ensure that it is
276
+ available for as long as needed to satisfy these requirements.
277
+
278
+ e) Convey the object code using peer-to-peer transmission, provided
279
+ you inform other peers where the object code and Corresponding
280
+ Source of the work are being offered to the general public at no
281
+ charge under subsection 6d.
282
+
283
+ A separable portion of the object code, whose source code is excluded
284
+ from the Corresponding Source as a System Library, need not be
285
+ included in conveying the object code work.
286
+
287
+ A "User Product" is either (1) a "consumer product", which means any
288
+ tangible personal property which is normally used for personal, family,
289
+ or household purposes, or (2) anything designed or sold for incorporation
290
+ into a dwelling. In determining whether a product is a consumer product,
291
+ doubtful cases shall be resolved in favor of coverage. For a particular
292
+ product received by a particular user, "normally used" refers to a
293
+ typical or common use of that class of product, regardless of the status
294
+ of the particular user or of the way in which the particular user
295
+ actually uses, or expects or is expected to use, the product. A product
296
+ is a consumer product regardless of whether the product has substantial
297
+ commercial, industrial or non-consumer uses, unless such uses represent
298
+ the only significant mode of use of the product.
299
+
300
+ "Installation Information" for a User Product means any methods,
301
+ procedures, authorization keys, or other information required to install
302
+ and execute modified versions of a covered work in that User Product from
303
+ a modified version of its Corresponding Source. The information must
304
+ suffice to ensure that the continued functioning of the modified object
305
+ code is in no case prevented or interfered with solely because
306
+ modification has been made.
307
+
308
+ If you convey an object code work under this section in, or with, or
309
+ specifically for use in, a User Product, and the conveying occurs as
310
+ part of a transaction in which the right of possession and use of the
311
+ User Product is transferred to the recipient in perpetuity or for a
312
+ fixed term (regardless of how the transaction is characterized), the
313
+ Corresponding Source conveyed under this section must be accompanied
314
+ by the Installation Information. But this requirement does not apply
315
+ if neither you nor any third party retains the ability to install
316
+ modified object code on the User Product (for example, the work has
317
+ been installed in ROM).
318
+
319
+ The requirement to provide Installation Information does not include a
320
+ requirement to continue to provide support service, warranty, or updates
321
+ for a work that has been modified or installed by the recipient, or for
322
+ the User Product in which it has been modified or installed. Access to a
323
+ network may be denied when the modification itself materially and
324
+ adversely affects the operation of the network or violates the rules and
325
+ protocols for communication across the network.
326
+
327
+ Corresponding Source conveyed, and Installation Information provided,
328
+ in accord with this section must be in a format that is publicly
329
+ documented (and with an implementation available to the public in
330
+ source code form), and must require no special password or key for
331
+ unpacking, reading or copying.
332
+
333
+ 7. Additional Terms.
334
+
335
+ "Additional permissions" are terms that supplement the terms of this
336
+ License by making exceptions from one or more of its conditions.
337
+ Additional permissions that are applicable to the entire Program shall
338
+ be treated as though they were included in this License, to the extent
339
+ that they are valid under applicable law. If additional permissions
340
+ apply only to part of the Program, that part may be used separately
341
+ under those permissions, but the entire Program remains governed by
342
+ this License without regard to the additional permissions.
343
+
344
+ When you convey a copy of a covered work, you may at your option
345
+ remove any additional permissions from that copy, or from any part of
346
+ it. (Additional permissions may be written to require their own
347
+ removal in certain cases when you modify the work.) You may place
348
+ additional permissions on material, added by you to a covered work,
349
+ for which you have or can give appropriate copyright permission.
350
+
351
+ Notwithstanding any other provision of this License, for material you
352
+ add to a covered work, you may (if authorized by the copyright holders of
353
+ that material) supplement the terms of this License with terms:
354
+
355
+ a) Disclaiming warranty or limiting liability differently from the
356
+ terms of sections 15 and 16 of this License; or
357
+
358
+ b) Requiring preservation of specified reasonable legal notices or
359
+ author attributions in that material or in the Appropriate Legal
360
+ Notices displayed by works containing it; or
361
+
362
+ c) Prohibiting misrepresentation of the origin of that material, or
363
+ requiring that modified versions of such material be marked in
364
+ reasonable ways as different from the original version; or
365
+
366
+ d) Limiting the use for publicity purposes of names of licensors or
367
+ authors of the material; or
368
+
369
+ e) Declining to grant rights under trademark law for use of some
370
+ trade names, trademarks, or service marks; or
371
+
372
+ f) Requiring indemnification of licensors and authors of that
373
+ material by anyone who conveys the material (or modified versions of
374
+ it) with contractual assumptions of liability to the recipient, for
375
+ any liability that these contractual assumptions directly impose on
376
+ those licensors and authors.
377
+
378
+ All other non-permissive additional terms are considered "further
379
+ restrictions" within the meaning of section 10. If the Program as you
380
+ received it, or any part of it, contains a notice stating that it is
381
+ governed by this License along with a term that is a further
382
+ restriction, you may remove that term. If a license document contains
383
+ a further restriction but permits relicensing or conveying under this
384
+ License, you may add to a covered work material governed by the terms
385
+ of that license document, provided that the further restriction does
386
+ not survive such relicensing or conveying.
387
+
388
+ If you add terms to a covered work in accord with this section, you
389
+ must place, in the relevant source files, a statement of the
390
+ additional terms that apply to those files, or a notice indicating
391
+ where to find the applicable terms.
392
+
393
+ Additional terms, permissive or non-permissive, may be stated in the
394
+ form of a separately written license, or stated as exceptions;
395
+ the above requirements apply either way.
396
+
397
+ 8. Termination.
398
+
399
+ You may not propagate or modify a covered work except as expressly
400
+ provided under this License. Any attempt otherwise to propagate or
401
+ modify it is void, and will automatically terminate your rights under
402
+ this License (including any patent licenses granted under the third
403
+ paragraph of section 11).
404
+
405
+ However, if you cease all violation of this License, then your
406
+ license from a particular copyright holder is reinstated (a)
407
+ provisionally, unless and until the copyright holder explicitly and
408
+ finally terminates your license, and (b) permanently, if the copyright
409
+ holder fails to notify you of the violation by some reasonable means
410
+ prior to 60 days after the cessation.
411
+
412
+ Moreover, your license from a particular copyright holder is
413
+ reinstated permanently if the copyright holder notifies you of the
414
+ violation by some reasonable means, this is the first time you have
415
+ received notice of violation of this License (for any work) from that
416
+ copyright holder, and you cure the violation prior to 30 days after
417
+ your receipt of the notice.
418
+
419
+ Termination of your rights under this section does not terminate the
420
+ licenses of parties who have received copies or rights from you under
421
+ this License. If your rights have been terminated and not permanently
422
+ reinstated, you do not qualify to receive new licenses for the same
423
+ material under section 10.
424
+
425
+ 9. Acceptance Not Required for Having Copies.
426
+
427
+ You are not required to accept this License in order to receive or
428
+ run a copy of the Program. Ancillary propagation of a covered work
429
+ occurring solely as a consequence of using peer-to-peer transmission
430
+ to receive a copy likewise does not require acceptance. However,
431
+ nothing other than this License grants you permission to propagate or
432
+ modify any covered work. These actions infringe copyright if you do
433
+ not accept this License. Therefore, by modifying or propagating a
434
+ covered work, you indicate your acceptance of this License to do so.
435
+
436
+ 10. Automatic Licensing of Downstream Recipients.
437
+
438
+ Each time you convey a covered work, the recipient automatically
439
+ receives a license from the original licensors, to run, modify and
440
+ propagate that work, subject to this License. You are not responsible
441
+ for enforcing compliance by third parties with this License.
442
+
443
+ An "entity transaction" is a transaction transferring control of an
444
+ organization, or substantially all assets of one, or subdividing an
445
+ organization, or merging organizations. If propagation of a covered
446
+ work results from an entity transaction, each party to that
447
+ transaction who receives a copy of the work also receives whatever
448
+ licenses to the work the party's predecessor in interest had or could
449
+ give under the previous paragraph, plus a right to possession of the
450
+ Corresponding Source of the work from the predecessor in interest, if
451
+ the predecessor has it or can get it with reasonable efforts.
452
+
453
+ You may not impose any further restrictions on the exercise of the
454
+ rights granted or affirmed under this License. For example, you may
455
+ not impose a license fee, royalty, or other charge for exercise of
456
+ rights granted under this License, and you may not initiate litigation
457
+ (including a cross-claim or counterclaim in a lawsuit) alleging that
458
+ any patent claim is infringed by making, using, selling, offering for
459
+ sale, or importing the Program or any portion of it.
460
+
461
+ 11. Patents.
462
+
463
+ A "contributor" is a copyright holder who authorizes use under this
464
+ License of the Program or a work on which the Program is based. The
465
+ work thus licensed is called the contributor's "contributor version".
466
+
467
+ A contributor's "essential patent claims" are all patent claims
468
+ owned or controlled by the contributor, whether already acquired or
469
+ hereafter acquired, that would be infringed by some manner, permitted
470
+ by this License, of making, using, or selling its contributor version,
471
+ but do not include claims that would be infringed only as a
472
+ consequence of further modification of the contributor version. For
473
+ purposes of this definition, "control" includes the right to grant
474
+ patent sublicenses in a manner consistent with the requirements of
475
+ this License.
476
+
477
+ Each contributor grants you a non-exclusive, worldwide, royalty-free
478
+ patent license under the contributor's essential patent claims, to
479
+ make, use, sell, offer for sale, import and otherwise run, modify and
480
+ propagate the contents of its contributor version.
481
+
482
+ In the following three paragraphs, a "patent license" is any express
483
+ agreement or commitment, however denominated, not to enforce a patent
484
+ (such as an express permission to practice a patent or covenant not to
485
+ sue for patent infringement). To "grant" such a patent license to a
486
+ party means to make such an agreement or commitment not to enforce a
487
+ patent against the party.
488
+
489
+ If you convey a covered work, knowingly relying on a patent license,
490
+ and the Corresponding Source of the work is not available for anyone
491
+ to copy, free of charge and under the terms of this License, through a
492
+ publicly available network server or other readily accessible means,
493
+ then you must either (1) cause the Corresponding Source to be so
494
+ available, or (2) arrange to deprive yourself of the benefit of the
495
+ patent license for this particular work, or (3) arrange, in a manner
496
+ consistent with the requirements of this License, to extend the patent
497
+ license to downstream recipients. "Knowingly relying" means you have
498
+ actual knowledge that, but for the patent license, your conveying the
499
+ covered work in a country, or your recipient's use of the covered work
500
+ in a country, would infringe one or more identifiable patents in that
501
+ country that you have reason to believe are valid.
502
+
503
+ If, pursuant to or in connection with a single transaction or
504
+ arrangement, you convey, or propagate by procuring conveyance of, a
505
+ covered work, and grant a patent license to some of the parties
506
+ receiving the covered work authorizing them to use, propagate, modify
507
+ or convey a specific copy of the covered work, then the patent license
508
+ you grant is automatically extended to all recipients of the covered
509
+ work and works based on it.
510
+
511
+ A patent license is "discriminatory" if it does not include within
512
+ the scope of its coverage, prohibits the exercise of, or is
513
+ conditioned on the non-exercise of one or more of the rights that are
514
+ specifically granted under this License. You may not convey a covered
515
+ work if you are a party to an arrangement with a third party that is
516
+ in the business of distributing software, under which you make payment
517
+ to the third party based on the extent of your activity of conveying
518
+ the work, and under which the third party grants, to any of the
519
+ parties who would receive the covered work from you, a discriminatory
520
+ patent license (a) in connection with copies of the covered work
521
+ conveyed by you (or copies made from those copies), or (b) primarily
522
+ for and in connection with specific products or compilations that
523
+ contain the covered work, unless you entered into that arrangement,
524
+ or that patent license was granted, prior to 28 March 2007.
525
+
526
+ Nothing in this License shall be construed as excluding or limiting
527
+ any implied license or other defenses to infringement that may
528
+ otherwise be available to you under applicable patent law.
529
+
530
+ 12. No Surrender of Others' Freedom.
531
+
532
+ If conditions are imposed on you (whether by court order, agreement or
533
+ otherwise) that contradict the conditions of this License, they do not
534
+ excuse you from the conditions of this License. If you cannot convey a
535
+ covered work so as to satisfy simultaneously your obligations under this
536
+ License and any other pertinent obligations, then as a consequence you may
537
+ not convey it at all. For example, if you agree to terms that obligate you
538
+ to collect a royalty for further conveying from those to whom you convey
539
+ the Program, the only way you could satisfy both those terms and this
540
+ License would be to refrain entirely from conveying the Program.
541
+
542
+ 13. Remote Network Interaction; Use with the GNU General Public License.
543
+
544
+ Notwithstanding any other provision of this License, if you modify the
545
+ Program, your modified version must prominently offer all users
546
+ interacting with it remotely through a computer network (if your version
547
+ supports such interaction) an opportunity to receive the Corresponding
548
+ Source of your version by providing access to the Corresponding Source
549
+ from a network server at no charge, through some standard or customary
550
+ means of facilitating copying of software. This Corresponding Source
551
+ shall include the Corresponding Source for any work covered by version 3
552
+ of the GNU General Public License that is incorporated pursuant to the
553
+ following paragraph.
554
+
555
+ Notwithstanding any other provision of this License, you have
556
+ permission to link or combine any covered work with a work licensed
557
+ under version 3 of the GNU General Public License into a single
558
+ combined work, and to convey the resulting work. The terms of this
559
+ License will continue to apply to the part which is the covered work,
560
+ but the work with which it is combined will remain governed by version
561
+ 3 of the GNU General Public License.
562
+
563
+ 14. Revised Versions of this License.
564
+
565
+ The Free Software Foundation may publish revised and/or new versions of
566
+ the GNU Affero General Public License from time to time. Such new versions
567
+ will be similar in spirit to the present version, but may differ in detail to
568
+ address new problems or concerns.
569
+
570
+ Each version is given a distinguishing version number. If the
571
+ Program specifies that a certain numbered version of the GNU Affero General
572
+ Public License "or any later version" applies to it, you have the
573
+ option of following the terms and conditions either of that numbered
574
+ version or of any later version published by the Free Software
575
+ Foundation. If the Program does not specify a version number of the
576
+ GNU Affero General Public License, you may choose any version ever published
577
+ by the Free Software Foundation.
578
+
579
+ If the Program specifies that a proxy can decide which future
580
+ versions of the GNU Affero General Public License can be used, that proxy's
581
+ public statement of acceptance of a version permanently authorizes you
582
+ to choose that version for the Program.
583
+
584
+ Later license versions may give you additional or different
585
+ permissions. However, no additional obligations are imposed on any
586
+ author or copyright holder as a result of your choosing to follow a
587
+ later version.
588
+
589
+ 15. Disclaimer of Warranty.
590
+
591
+ THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
592
+ APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
593
+ HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
594
+ OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
595
+ THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
596
+ PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
597
+ IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
598
+ ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
599
+
600
+ 16. Limitation of Liability.
601
+
602
+ IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
603
+ WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
604
+ THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
605
+ GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
606
+ USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
607
+ DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
608
+ PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
609
+ EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
610
+ SUCH DAMAGES.
611
+
612
+ 17. Interpretation of Sections 15 and 16.
613
+
614
+ If the disclaimer of warranty and limitation of liability provided
615
+ above cannot be given local legal effect according to their terms,
616
+ reviewing courts shall apply local law that most closely approximates
617
+ an absolute waiver of all civil liability in connection with the
618
+ Program, unless a warranty or assumption of liability accompanies a
619
+ copy of the Program in return for a fee.
620
+
621
+ END OF TERMS AND CONDITIONS
622
+
623
+ How to Apply These Terms to Your New Programs
624
+
625
+ If you develop a new program, and you want it to be of the greatest
626
+ possible use to the public, the best way to achieve this is to make it
627
+ free software which everyone can redistribute and change under these terms.
628
+
629
+ To do so, attach the following notices to the program. It is safest
630
+ to attach them to the start of each source file to most effectively
631
+ state the exclusion of warranty; and each file should have at least
632
+ the "copyright" line and a pointer to where the full notice is found.
633
+
634
+ <one line to give the program's name and a brief idea of what it does.>
635
+ Copyright (C) <year> <name of author>
636
+
637
+ This program is free software: you can redistribute it and/or modify
638
+ it under the terms of the GNU Affero General Public License as published by
639
+ the Free Software Foundation, either version 3 of the License, or
640
+ (at your option) any later version.
641
+
642
+ This program is distributed in the hope that it will be useful,
643
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
644
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
645
+ GNU Affero General Public License for more details.
646
+
647
+ You should have received a copy of the GNU Affero General Public License
648
+ along with this program. If not, see <https://www.gnu.org/licenses/>.
649
+
650
+ Also add information on how to contact you by electronic and paper mail.
651
+
652
+ If your software can interact with users remotely through a computer
653
+ network, you should also make sure that it provides a way for users to
654
+ get its source. For example, if your program is a web application, its
655
+ interface could display a "Source" link that leads users to an archive
656
+ of the code. There are many ways you could offer source, and different
657
+ solutions will be better for different programs; see section 13 for the
658
+ specific requirements.
659
+
660
+ You should also get your employer (if you work as a programmer) or school,
661
+ if any, to sign a "copyright disclaimer" for the program, if necessary.
662
+ For more information on this, and how to apply and follow the GNU AGPL, see
663
+ <https://www.gnu.org/licenses/>.
README.md CHANGED
@@ -1,12 +1,710 @@
1
  ---
2
- title: Webui
3
- emoji: 🐢
4
- colorFrom: purple
5
- colorTo: gray
6
  sdk: gradio
7
  sdk_version: 4.20.1
8
- app_file: app.py
9
- pinned: false
10
  ---
 
11
 
12
- Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ title: webui
3
+ app_file: C:\Users\Camch\AppData\Local\Programs\Python\Python310\app.py
 
 
4
  sdk: gradio
5
  sdk_version: 4.20.1
 
 
6
  ---
7
+ # Stable Diffusion WebUI Forge
8
 
9
+ Stable Diffusion WebUI Forge is a platform on top of [Stable Diffusion WebUI](https://github.com/AUTOMATIC1111/stable-diffusion-webui) (based on [Gradio](https://www.gradio.app/)) to make development easier, optimize resource management, and speed up inference.
10
+
11
+ The name "Forge" is inspired from "Minecraft Forge". This project is aimed at becoming SD WebUI's Forge.
12
+
13
+ Compared to original WebUI (for SDXL inference at 1024px), you can expect the below speed-ups:
14
+
15
+ 1. If you use common GPU like 8GB vram, you can expect to get about **30~45% speed up** in inference speed (it/s), the GPU memory peak (in task manager) will drop about 700MB to 1.3GB, the maximum diffusion resolution (that will not OOM) will increase about 2x to 3x, and the maximum diffusion batch size (that will not OOM) will increase about 4x to 6x.
16
+
17
+ 2. If you use less powerful GPU like 6GB vram, you can expect to get about **60~75% speed up** in inference speed (it/s), the GPU memory peak (in task manager) will drop about 800MB to 1.5GB, the maximum diffusion resolution (that will not OOM) will increase about 3x, and the maximum diffusion batch size (that will not OOM) will increase about 4x.
18
+
19
+ 3. If you use powerful GPU like 4090 with 24GB vram, you can expect to get about **3~6% speed up** in inference speed (it/s), the GPU memory peak (in task manager) will drop about 1GB to 1.4GB, the maximum diffusion resolution (that will not OOM) will increase about 1.6x, and the maximum diffusion batch size (that will not OOM) will increase about 2x.
20
+
21
+ 4. If you use ControlNet for SDXL, the maximum ControlNet count (that will not OOM) will increase about 2x, the speed with SDXL+ControlNet will **speed up about 30~45%**.
22
+
23
+ Another very important change that Forge brings is **Unet Patcher**. Using Unet Patcher, methods like Self-Attention Guidance, Kohya High Res Fix, FreeU, StyleAlign, Hypertile can all be implemented in about 100 lines of codes.
24
+
25
+ Thanks to Unet Patcher, many new things are possible now and supported in Forge, including SVD, Z123, masked Ip-adapter, masked controlnet, photomaker, etc.
26
+
27
+ **No need to monkeypatch UNet and conflict other extensions anymore!**
28
+
29
+ Forge also adds a few samplers, including but not limited to DDPM, DDPM Karras, DPM++ 2M Turbo, DPM++ 2M SDE Turbo, LCM Karras, Euler A Turbo, etc. (LCM is already in original webui since 1.7.0).
30
+
31
+ Finally, Forge promise that we will only do our jobs. Forge will never add unnecessary opinioned changes to the user interface. You are still using 100% Automatic1111 WebUI.
32
+
33
+ # Installing Forge
34
+
35
+ If you are proficient in Git and you want to install Forge as another branch of SD-WebUI, please see [here](https://github.com/continue-revolution/sd-webui-animatediff/blob/forge/master/docs/how-to-use.md#you-have-a1111-and-you-know-git). In this way, you can reuse all SD checkpoints and all extensions you installed previously in your OG SD-WebUI, but you should know what you are doing.
36
+
37
+ If you know what you are doing, you can install Forge using same method as SD-WebUI. (Install Git, Python, Git Clone the forge repo `https://github.com/lllyasviel/stable-diffusion-webui-forge.git` and then run webui-user.bat).
38
+
39
+ **Or you can just use this one-click installation package (with git and python included).**
40
+
41
+ [>>> Click Here to Download One-Click Package<<<](https://github.com/lllyasviel/stable-diffusion-webui-forge/releases/download/latest/webui_forge_cu121_torch21.7z)
42
+
43
+ After you download, you uncompress, use `update.bat` to update, and use `run.bat` to run.
44
+
45
+ Note that running `update.bat` is important, otherwise you may be using a previous version with potential bugs unfixed.
46
+
47
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/c49bd60d-82bd-4086-9859-88d472582b94)
48
+
49
+ # Screenshots of Comparison
50
+
51
+ I tested with several devices, and this is a typical result from 8GB VRAM (3070ti laptop) with SDXL.
52
+
53
+ **This is original WebUI:**
54
+
55
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/16893937-9ed9-4f8e-b960-70cd5d1e288f)
56
+
57
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/7bbc16fe-64ef-49e2-a595-d91bb658bd94)
58
+
59
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/de1747fd-47bc-482d-a5c6-0728dd475943)
60
+
61
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/96e5e171-2d74-41ba-9dcc-11bf68be7e16)
62
+
63
+ (average about 7.4GB/8GB, peak at about 7.9GB/8GB)
64
+
65
+ **This is WebUI Forge:**
66
+
67
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/ca5e05ed-bd86-4ced-8662-f41034648e8c)
68
+
69
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/3629ee36-4a99-4d9b-b371-12efb260a283)
70
+
71
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/6d13ebb7-c30d-4aa8-9242-c0b5a1af8c95)
72
+
73
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/c4f723c3-6ea7-4539-980b-0708ed2a69aa)
74
+
75
+ (average and peak are all 6.3GB/8GB)
76
+
77
+ You can see that Forge does not change WebUI results. Installing Forge is not a seed breaking change.
78
+
79
+ Forge can perfectly keep WebUI unchanged even for most complicated prompts like `fantasy landscape with a [mountain:lake:0.25] and [an oak:a christmas tree:0.75][ in foreground::0.6][ in background:0.25] [shoddy:masterful:0.5]`.
80
+
81
+ All your previous works still work in Forge!
82
+
83
+ # Forge Backend
84
+
85
+ Forge backend removes all WebUI's codes related to resource management and reworked everything. All previous CMD flags like `medvram, lowvram, medvram-sdxl, precision full, no half, no half vae, attention_xxx, upcast unet`, ... are all **REMOVED**. Adding these flags will not cause error but they will not do anything now. **We highly encourage Forge users to remove all cmd flags and let Forge to decide how to load models.**
86
+
87
+ Without any cmd flag, Forge can run SDXL with 4GB vram and SD1.5 with 2GB vram.
88
+
89
+ **Some flags that you may still pay attention to:**
90
+
91
+ 1. `--always-offload-from-vram` (This flag will make things **slower** but less risky). This option will let Forge always unload models from VRAM. This can be useful if you use multiple software together and want Forge to use less VRAM and give some VRAM to other software, or when you are using some old extensions that will compete vram with Forge, or (very rarely) when you get OOM.
92
+
93
+ 2. `--cuda-malloc` (This flag will make things **faster** but more risky). This will ask pytorch to use *cudaMallocAsync* for tensor malloc. On some profilers I can observe performance gain at millisecond level, but the real speed up on most my devices are often unnoticed (about or less than 0.1 second per image). This cannot be set as default because many users reported issues that the async malloc will crash the program. Users need to enable this cmd flag at their own risk.
94
+
95
+ 3. `--cuda-stream` (This flag will make things **faster** but more risky). This will use pytorch CUDA streams (a special type of thread on GPU) to move models and compute tensors simultaneously. This can almost eliminate all model moving time, and speed up SDXL on 30XX/40XX devices with small VRAM (eg, RTX 4050 6GB, RTX 3060 Laptop 6GB, etc) by about 15\% to 25\%. However, this unfortunately cannot be set as default because I observe higher possibility of pure black images (Nan outputs) on 2060, and higher chance of OOM on 1080 and 2060. When the resolution is large, there is a chance that the computation time of one single attention layer is longer than the time for moving entire model to GPU. When that happens, the next attention layer will OOM since the GPU is filled with the entire model, and no remaining space is available for computing another attention layer. Most overhead detecting methods are not robust enough to be reliable on old devices (in my tests). Users need to enable this cmd flag at their own risk.
96
+
97
+ 4. `--pin-shared-memory` (This flag will make things **faster** but more risky). Effective only when used together with `--cuda-stream`. This will offload modules to Shared GPU Memory instead of system RAM when offloading models. On some 30XX/40XX devices with small VRAM (eg, RTX 4050 6GB, RTX 3060 Laptop 6GB, etc), I can observe significant (at least 20\%) speed-up for SDXL. However, this unfortunately cannot be set as default because the OOM of Shared GPU Memory is a much more severe problem than common GPU memory OOM. Pytorch does not provide any robust method to unload or detect Shared GPU Memory. Once the Shared GPU Memory OOM, the entire program will crash (observed with SDXL on GTX 1060/1050/1066), and there is no dynamic method to prevent or recover from the crash. Users need to enable this cmd flag at their own risk.
98
+
99
+ If you really want to play with cmd flags, you can additionally control the GPU with:
100
+
101
+ (extreme VRAM cases)
102
+
103
+ --always-gpu
104
+ --always-cpu
105
+
106
+ (rare attention cases)
107
+
108
+ --attention-split
109
+ --attention-quad
110
+ --attention-pytorch
111
+ --disable-xformers
112
+ --disable-attention-upcast
113
+
114
+ (float point type)
115
+
116
+ --all-in-fp32
117
+ --all-in-fp16
118
+ --unet-in-bf16
119
+ --unet-in-fp16
120
+ --unet-in-fp8-e4m3fn
121
+ --unet-in-fp8-e5m2
122
+ --vae-in-fp16
123
+ --vae-in-fp32
124
+ --vae-in-bf16
125
+ --clip-in-fp8-e4m3fn
126
+ --clip-in-fp8-e5m2
127
+ --clip-in-fp16
128
+ --clip-in-fp32
129
+
130
+ (rare platforms)
131
+
132
+ --directml
133
+ --disable-ipex-hijack
134
+ --pytorch-deterministic
135
+
136
+ Again, Forge do not recommend users to use any cmd flags unless you are very sure that you really need these.
137
+
138
+ # UNet Patcher
139
+
140
+ Note that [Forge does not use any other software as backend](https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/169). The full name of the backend is `Stable Diffusion WebUI with Forge backend`, or for simplicity, the `Forge backend`. The API and python symbols are made similar to previous software only for reducing the learning cost of developers.
141
+
142
+ Now developing an extension is super simple. We finally have a patchable UNet.
143
+
144
+ Below is using one single file with 80 lines of codes to support FreeU:
145
+
146
+ `extensions-builtin/sd_forge_freeu/scripts/forge_freeu.py`
147
+
148
+ ```python
149
+ import torch
150
+ import gradio as gr
151
+ from modules import scripts
152
+
153
+
154
+ def Fourier_filter(x, threshold, scale):
155
+ x_freq = torch.fft.fftn(x.float(), dim=(-2, -1))
156
+ x_freq = torch.fft.fftshift(x_freq, dim=(-2, -1))
157
+ B, C, H, W = x_freq.shape
158
+ mask = torch.ones((B, C, H, W), device=x.device)
159
+ crow, ccol = H // 2, W //2
160
+ mask[..., crow - threshold:crow + threshold, ccol - threshold:ccol + threshold] = scale
161
+ x_freq = x_freq * mask
162
+ x_freq = torch.fft.ifftshift(x_freq, dim=(-2, -1))
163
+ x_filtered = torch.fft.ifftn(x_freq, dim=(-2, -1)).real
164
+ return x_filtered.to(x.dtype)
165
+
166
+
167
+ def set_freeu_v2_patch(model, b1, b2, s1, s2):
168
+ model_channels = model.model.model_config.unet_config["model_channels"]
169
+ scale_dict = {model_channels * 4: (b1, s1), model_channels * 2: (b2, s2)}
170
+
171
+ def output_block_patch(h, hsp, *args, **kwargs):
172
+ scale = scale_dict.get(h.shape[1], None)
173
+ if scale is not None:
174
+ hidden_mean = h.mean(1).unsqueeze(1)
175
+ B = hidden_mean.shape[0]
176
+ hidden_max, _ = torch.max(hidden_mean.view(B, -1), dim=-1, keepdim=True)
177
+ hidden_min, _ = torch.min(hidden_mean.view(B, -1), dim=-1, keepdim=True)
178
+ hidden_mean = (hidden_mean - hidden_min.unsqueeze(2).unsqueeze(3)) / \
179
+ (hidden_max - hidden_min).unsqueeze(2).unsqueeze(3)
180
+ h[:, :h.shape[1] // 2] = h[:, :h.shape[1] // 2] * ((scale[0] - 1) * hidden_mean + 1)
181
+ hsp = Fourier_filter(hsp, threshold=1, scale=scale[1])
182
+ return h, hsp
183
+
184
+ m = model.clone()
185
+ m.set_model_output_block_patch(output_block_patch)
186
+ return m
187
+
188
+
189
+ class FreeUForForge(scripts.Script):
190
+ def title(self):
191
+ return "FreeU Integrated"
192
+
193
+ def show(self, is_img2img):
194
+ # make this extension visible in both txt2img and img2img tab.
195
+ return scripts.AlwaysVisible
196
+
197
+ def ui(self, *args, **kwargs):
198
+ with gr.Accordion(open=False, label=self.title()):
199
+ freeu_enabled = gr.Checkbox(label='Enabled', value=False)
200
+ freeu_b1 = gr.Slider(label='B1', minimum=0, maximum=2, step=0.01, value=1.01)
201
+ freeu_b2 = gr.Slider(label='B2', minimum=0, maximum=2, step=0.01, value=1.02)
202
+ freeu_s1 = gr.Slider(label='S1', minimum=0, maximum=4, step=0.01, value=0.99)
203
+ freeu_s2 = gr.Slider(label='S2', minimum=0, maximum=4, step=0.01, value=0.95)
204
+
205
+ return freeu_enabled, freeu_b1, freeu_b2, freeu_s1, freeu_s2
206
+
207
+ def process_before_every_sampling(self, p, *script_args, **kwargs):
208
+ # This will be called before every sampling.
209
+ # If you use highres fix, this will be called twice.
210
+
211
+ freeu_enabled, freeu_b1, freeu_b2, freeu_s1, freeu_s2 = script_args
212
+
213
+ if not freeu_enabled:
214
+ return
215
+
216
+ unet = p.sd_model.forge_objects.unet
217
+
218
+ unet = set_freeu_v2_patch(unet, freeu_b1, freeu_b2, freeu_s1, freeu_s2)
219
+
220
+ p.sd_model.forge_objects.unet = unet
221
+
222
+ # Below codes will add some logs to the texts below the image outputs on UI.
223
+ # The extra_generation_params does not influence results.
224
+ p.extra_generation_params.update(dict(
225
+ freeu_enabled=freeu_enabled,
226
+ freeu_b1=freeu_b1,
227
+ freeu_b2=freeu_b2,
228
+ freeu_s1=freeu_s1,
229
+ freeu_s2=freeu_s2,
230
+ ))
231
+
232
+ return
233
+ ```
234
+
235
+ It looks like this:
236
+
237
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/277bac6e-5ea7-4bff-b71a-e55a60cfc03c)
238
+
239
+ Similar components like HyperTile, KohyaHighResFix, SAG, can all be implemented within 100 lines of codes (see also the codes).
240
+
241
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/06472b03-b833-4816-ab47-70712ac024d3)
242
+
243
+ ControlNets can finally be called by different extensions.
244
+
245
+ Implementing Stable Video Diffusion and Zero123 are also super simple now (see also the codes).
246
+
247
+ *Stable Video Diffusion:*
248
+
249
+ `extensions-builtin/sd_forge_svd/scripts/forge_svd.py`
250
+
251
+ ```python
252
+ import torch
253
+ import gradio as gr
254
+ import os
255
+ import pathlib
256
+
257
+ from modules import script_callbacks
258
+ from modules.paths import models_path
259
+ from modules.ui_common import ToolButton, refresh_symbol
260
+ from modules import shared
261
+
262
+ from modules_forge.forge_util import numpy_to_pytorch, pytorch_to_numpy
263
+ from ldm_patched.modules.sd import load_checkpoint_guess_config
264
+ from ldm_patched.contrib.external_video_model import VideoLinearCFGGuidance, SVD_img2vid_Conditioning
265
+ from ldm_patched.contrib.external import KSampler, VAEDecode
266
+
267
+
268
+ opVideoLinearCFGGuidance = VideoLinearCFGGuidance()
269
+ opSVD_img2vid_Conditioning = SVD_img2vid_Conditioning()
270
+ opKSampler = KSampler()
271
+ opVAEDecode = VAEDecode()
272
+
273
+ svd_root = os.path.join(models_path, 'svd')
274
+ os.makedirs(svd_root, exist_ok=True)
275
+ svd_filenames = []
276
+
277
+
278
+ def update_svd_filenames():
279
+ global svd_filenames
280
+ svd_filenames = [
281
+ pathlib.Path(x).name for x in
282
+ shared.walk_files(svd_root, allowed_extensions=[".pt", ".ckpt", ".safetensors"])
283
+ ]
284
+ return svd_filenames
285
+
286
+
287
+ @torch.inference_mode()
288
+ @torch.no_grad()
289
+ def predict(filename, width, height, video_frames, motion_bucket_id, fps, augmentation_level,
290
+ sampling_seed, sampling_steps, sampling_cfg, sampling_sampler_name, sampling_scheduler,
291
+ sampling_denoise, guidance_min_cfg, input_image):
292
+ filename = os.path.join(svd_root, filename)
293
+ model_raw, _, vae, clip_vision = \
294
+ load_checkpoint_guess_config(filename, output_vae=True, output_clip=False, output_clipvision=True)
295
+ model = opVideoLinearCFGGuidance.patch(model_raw, guidance_min_cfg)[0]
296
+ init_image = numpy_to_pytorch(input_image)
297
+ positive, negative, latent_image = opSVD_img2vid_Conditioning.encode(
298
+ clip_vision, init_image, vae, width, height, video_frames, motion_bucket_id, fps, augmentation_level)
299
+ output_latent = opKSampler.sample(model, sampling_seed, sampling_steps, sampling_cfg,
300
+ sampling_sampler_name, sampling_scheduler, positive,
301
+ negative, latent_image, sampling_denoise)[0]
302
+ output_pixels = opVAEDecode.decode(vae, output_latent)[0]
303
+ outputs = pytorch_to_numpy(output_pixels)
304
+ return outputs
305
+
306
+
307
+ def on_ui_tabs():
308
+ with gr.Blocks() as svd_block:
309
+ with gr.Row():
310
+ with gr.Column():
311
+ input_image = gr.Image(label='Input Image', source='upload', type='numpy', height=400)
312
+
313
+ with gr.Row():
314
+ filename = gr.Dropdown(label="SVD Checkpoint Filename",
315
+ choices=svd_filenames,
316
+ value=svd_filenames[0] if len(svd_filenames) > 0 else None)
317
+ refresh_button = ToolButton(value=refresh_symbol, tooltip="Refresh")
318
+ refresh_button.click(
319
+ fn=lambda: gr.update(choices=update_svd_filenames),
320
+ inputs=[], outputs=filename)
321
+
322
+ width = gr.Slider(label='Width', minimum=16, maximum=8192, step=8, value=1024)
323
+ height = gr.Slider(label='Height', minimum=16, maximum=8192, step=8, value=576)
324
+ video_frames = gr.Slider(label='Video Frames', minimum=1, maximum=4096, step=1, value=14)
325
+ motion_bucket_id = gr.Slider(label='Motion Bucket Id', minimum=1, maximum=1023, step=1, value=127)
326
+ fps = gr.Slider(label='Fps', minimum=1, maximum=1024, step=1, value=6)
327
+ augmentation_level = gr.Slider(label='Augmentation Level', minimum=0.0, maximum=10.0, step=0.01,
328
+ value=0.0)
329
+ sampling_steps = gr.Slider(label='Sampling Steps', minimum=1, maximum=200, step=1, value=20)
330
+ sampling_cfg = gr.Slider(label='CFG Scale', minimum=0.0, maximum=50.0, step=0.1, value=2.5)
331
+ sampling_denoise = gr.Slider(label='Sampling Denoise', minimum=0.0, maximum=1.0, step=0.01, value=1.0)
332
+ guidance_min_cfg = gr.Slider(label='Guidance Min Cfg', minimum=0.0, maximum=100.0, step=0.5, value=1.0)
333
+ sampling_sampler_name = gr.Radio(label='Sampler Name',
334
+ choices=['euler', 'euler_ancestral', 'heun', 'heunpp2', 'dpm_2',
335
+ 'dpm_2_ancestral', 'lms', 'dpm_fast', 'dpm_adaptive',
336
+ 'dpmpp_2s_ancestral', 'dpmpp_sde', 'dpmpp_sde_gpu',
337
+ 'dpmpp_2m', 'dpmpp_2m_sde', 'dpmpp_2m_sde_gpu',
338
+ 'dpmpp_3m_sde', 'dpmpp_3m_sde_gpu', 'ddpm', 'lcm', 'ddim',
339
+ 'uni_pc', 'uni_pc_bh2'], value='euler')
340
+ sampling_scheduler = gr.Radio(label='Scheduler',
341
+ choices=['normal', 'karras', 'exponential', 'sgm_uniform', 'simple',
342
+ 'ddim_uniform'], value='karras')
343
+ sampling_seed = gr.Number(label='Seed', value=12345, precision=0)
344
+
345
+ generate_button = gr.Button(value="Generate")
346
+
347
+ ctrls = [filename, width, height, video_frames, motion_bucket_id, fps, augmentation_level,
348
+ sampling_seed, sampling_steps, sampling_cfg, sampling_sampler_name, sampling_scheduler,
349
+ sampling_denoise, guidance_min_cfg, input_image]
350
+
351
+ with gr.Column():
352
+ output_gallery = gr.Gallery(label='Gallery', show_label=False, object_fit='contain',
353
+ visible=True, height=1024, columns=4)
354
+
355
+ generate_button.click(predict, inputs=ctrls, outputs=[output_gallery])
356
+ return [(svd_block, "SVD", "svd")]
357
+
358
+
359
+ update_svd_filenames()
360
+ script_callbacks.on_ui_tabs(on_ui_tabs)
361
+ ```
362
+
363
+ Note that although the above codes look like independent codes, they actually will automatically offload/unload any other models. For example, below is me opening webui, load SDXL, generated an image, then go to SVD, then generated image frames. You can see that the GPU memory is perfectly managed and the SDXL is moved to RAM then SVD is moved to GPU.
364
+
365
+ Note that this management is fully automatic. This makes writing extensions super simple.
366
+
367
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/de1a2d05-344a-44d7-bab8-9ecc0a58a8d3)
368
+
369
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/14bcefcf-599f-42c3-bce9-3fd5e428dd91)
370
+
371
+ Similarly, Zero123:
372
+
373
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/7685019c-7239-47fb-9cb5-2b7b33943285)
374
+
375
+ ### Write a simple ControlNet:
376
+
377
+ Below is a simple extension to have a completely independent pass of ControlNet that never conflicts any other extensions:
378
+
379
+ `extensions-builtin/sd_forge_controlnet_example/scripts/sd_forge_controlnet_example.py`
380
+
381
+ Note that this extension is hidden because it is only for developers. To see it in UI, use `--show-controlnet-example`.
382
+
383
+ The memory optimization in this example is fully automatic. You do not need to care about memory and inference speed, but you may want to cache objects if you wish.
384
+
385
+ ```python
386
+ # Use --show-controlnet-example to see this extension.
387
+
388
+ import cv2
389
+ import gradio as gr
390
+ import torch
391
+
392
+ from modules import scripts
393
+ from modules.shared_cmd_options import cmd_opts
394
+ from modules_forge.shared import supported_preprocessors
395
+ from modules.modelloader import load_file_from_url
396
+ from ldm_patched.modules.controlnet import load_controlnet
397
+ from modules_forge.controlnet import apply_controlnet_advanced
398
+ from modules_forge.forge_util import numpy_to_pytorch
399
+ from modules_forge.shared import controlnet_dir
400
+
401
+
402
+ class ControlNetExampleForge(scripts.Script):
403
+ model = None
404
+
405
+ def title(self):
406
+ return "ControlNet Example for Developers"
407
+
408
+ def show(self, is_img2img):
409
+ # make this extension visible in both txt2img and img2img tab.
410
+ return scripts.AlwaysVisible
411
+
412
+ def ui(self, *args, **kwargs):
413
+ with gr.Accordion(open=False, label=self.title()):
414
+ gr.HTML('This is an example controlnet extension for developers.')
415
+ gr.HTML('You see this extension because you used --show-controlnet-example')
416
+ input_image = gr.Image(source='upload', type='numpy')
417
+ funny_slider = gr.Slider(label='This slider does nothing. It just shows you how to transfer parameters.',
418
+ minimum=0.0, maximum=1.0, value=0.5)
419
+
420
+ return input_image, funny_slider
421
+
422
+ def process(self, p, *script_args, **kwargs):
423
+ input_image, funny_slider = script_args
424
+
425
+ # This slider does nothing. It just shows you how to transfer parameters.
426
+ del funny_slider
427
+
428
+ if input_image is None:
429
+ return
430
+
431
+ # controlnet_canny_path = load_file_from_url(
432
+ # url='https://huggingface.co/lllyasviel/sd_control_collection/resolve/main/sai_xl_canny_256lora.safetensors',
433
+ # model_dir=model_dir,
434
+ # file_name='sai_xl_canny_256lora.safetensors'
435
+ # )
436
+ controlnet_canny_path = load_file_from_url(
437
+ url='https://huggingface.co/lllyasviel/fav_models/resolve/main/fav/control_v11p_sd15_canny_fp16.safetensors',
438
+ model_dir=controlnet_dir,
439
+ file_name='control_v11p_sd15_canny_fp16.safetensors'
440
+ )
441
+ print('The model [control_v11p_sd15_canny_fp16.safetensors] download finished.')
442
+
443
+ self.model = load_controlnet(controlnet_canny_path)
444
+ print('Controlnet loaded.')
445
+
446
+ return
447
+
448
+ def process_before_every_sampling(self, p, *script_args, **kwargs):
449
+ # This will be called before every sampling.
450
+ # If you use highres fix, this will be called twice.
451
+
452
+ input_image, funny_slider = script_args
453
+
454
+ if input_image is None or self.model is None:
455
+ return
456
+
457
+ B, C, H, W = kwargs['noise'].shape # latent_shape
458
+ height = H * 8
459
+ width = W * 8
460
+ batch_size = p.batch_size
461
+
462
+ preprocessor = supported_preprocessors['canny']
463
+
464
+ # detect control at certain resolution
465
+ control_image = preprocessor(
466
+ input_image, resolution=512, slider_1=100, slider_2=200, slider_3=None)
467
+
468
+ # here we just use nearest neighbour to align input shape.
469
+ # You may want crop and resize, or crop and fill, or others.
470
+ control_image = cv2.resize(
471
+ control_image, (width, height), interpolation=cv2.INTER_NEAREST)
472
+
473
+ # Output preprocessor result. Now called every sampling. Cache in your own way.
474
+ p.extra_result_images.append(control_image)
475
+
476
+ print('Preprocessor Canny finished.')
477
+
478
+ control_image_bchw = numpy_to_pytorch(control_image).movedim(-1, 1)
479
+
480
+ unet = p.sd_model.forge_objects.unet
481
+
482
+ # Unet has input, middle, output blocks, and we can give different weights
483
+ # to each layers in all blocks.
484
+ # Below is an example for stronger control in middle block.
485
+ # This is helpful for some high-res fix passes. (p.is_hr_pass)
486
+ positive_advanced_weighting = {
487
+ 'input': [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2],
488
+ 'middle': [1.0],
489
+ 'output': [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2]
490
+ }
491
+ negative_advanced_weighting = {
492
+ 'input': [0.15, 0.25, 0.35, 0.45, 0.55, 0.65, 0.75, 0.85, 0.95, 1.05, 1.15, 1.25],
493
+ 'middle': [1.05],
494
+ 'output': [0.15, 0.25, 0.35, 0.45, 0.55, 0.65, 0.75, 0.85, 0.95, 1.05, 1.15, 1.25]
495
+ }
496
+
497
+ # The advanced_frame_weighting is a weight applied to each image in a batch.
498
+ # The length of this list must be same with batch size
499
+ # For example, if batch size is 5, the below list is [0.2, 0.4, 0.6, 0.8, 1.0]
500
+ # If you view the 5 images as 5 frames in a video, this will lead to
501
+ # progressively stronger control over time.
502
+ advanced_frame_weighting = [float(i + 1) / float(batch_size) for i in range(batch_size)]
503
+
504
+ # The advanced_sigma_weighting allows you to dynamically compute control
505
+ # weights given diffusion timestep (sigma).
506
+ # For example below code can softly make beginning steps stronger than ending steps.
507
+ sigma_max = unet.model.model_sampling.sigma_max
508
+ sigma_min = unet.model.model_sampling.sigma_min
509
+ advanced_sigma_weighting = lambda s: (s - sigma_min) / (sigma_max - sigma_min)
510
+
511
+ # You can even input a tensor to mask all control injections
512
+ # The mask will be automatically resized during inference in UNet.
513
+ # The size should be B 1 H W and the H and W are not important
514
+ # because they will be resized automatically
515
+ advanced_mask_weighting = torch.ones(size=(1, 1, 512, 512))
516
+
517
+ # But in this simple example we do not use them
518
+ positive_advanced_weighting = None
519
+ negative_advanced_weighting = None
520
+ advanced_frame_weighting = None
521
+ advanced_sigma_weighting = None
522
+ advanced_mask_weighting = None
523
+
524
+ unet = apply_controlnet_advanced(unet=unet, controlnet=self.model, image_bchw=control_image_bchw,
525
+ strength=0.6, start_percent=0.0, end_percent=0.8,
526
+ positive_advanced_weighting=positive_advanced_weighting,
527
+ negative_advanced_weighting=negative_advanced_weighting,
528
+ advanced_frame_weighting=advanced_frame_weighting,
529
+ advanced_sigma_weighting=advanced_sigma_weighting,
530
+ advanced_mask_weighting=advanced_mask_weighting)
531
+
532
+ p.sd_model.forge_objects.unet = unet
533
+
534
+ # Below codes will add some logs to the texts below the image outputs on UI.
535
+ # The extra_generation_params does not influence results.
536
+ p.extra_generation_params.update(dict(
537
+ controlnet_info='You should see these texts below output images!',
538
+ ))
539
+
540
+ return
541
+
542
+
543
+ # Use --show-controlnet-example to see this extension.
544
+ if not cmd_opts.show_controlnet_example:
545
+ del ControlNetExampleForge
546
+
547
+ ```
548
+
549
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/822fa2fc-c9f4-4f58-8669-4b6680b91063)
550
+
551
+
552
+ ### Add a preprocessor
553
+
554
+ Below is the full codes to add a normalbae preprocessor with perfect memory managements.
555
+
556
+ You can use arbitrary independent extensions to add a preprocessor.
557
+
558
+ Your preprocessor will be read by all other extensions using `modules_forge.shared.preprocessors`
559
+
560
+ Below codes are in `extensions-builtin\forge_preprocessor_normalbae\scripts\preprocessor_normalbae.py`
561
+
562
+ ```python
563
+ from modules_forge.supported_preprocessor import Preprocessor, PreprocessorParameter
564
+ from modules_forge.shared import preprocessor_dir, add_supported_preprocessor
565
+ from modules_forge.forge_util import resize_image_with_pad
566
+ from modules.modelloader import load_file_from_url
567
+
568
+ import types
569
+ import torch
570
+ import numpy as np
571
+
572
+ from einops import rearrange
573
+ from annotator.normalbae.models.NNET import NNET
574
+ from annotator.normalbae import load_checkpoint
575
+ from torchvision import transforms
576
+
577
+
578
+ class PreprocessorNormalBae(Preprocessor):
579
+ def __init__(self):
580
+ super().__init__()
581
+ self.name = 'normalbae'
582
+ self.tags = ['NormalMap']
583
+ self.model_filename_filters = ['normal']
584
+ self.slider_resolution = PreprocessorParameter(
585
+ label='Resolution', minimum=128, maximum=2048, value=512, step=8, visible=True)
586
+ self.slider_1 = PreprocessorParameter(visible=False)
587
+ self.slider_2 = PreprocessorParameter(visible=False)
588
+ self.slider_3 = PreprocessorParameter(visible=False)
589
+ self.show_control_mode = True
590
+ self.do_not_need_model = False
591
+ self.sorting_priority = 100 # higher goes to top in the list
592
+
593
+ def load_model(self):
594
+ if self.model_patcher is not None:
595
+ return
596
+
597
+ model_path = load_file_from_url(
598
+ "https://huggingface.co/lllyasviel/Annotators/resolve/main/scannet.pt",
599
+ model_dir=preprocessor_dir)
600
+
601
+ args = types.SimpleNamespace()
602
+ args.mode = 'client'
603
+ args.architecture = 'BN'
604
+ args.pretrained = 'scannet'
605
+ args.sampling_ratio = 0.4
606
+ args.importance_ratio = 0.7
607
+ model = NNET(args)
608
+ model = load_checkpoint(model_path, model)
609
+ self.norm = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
610
+
611
+ self.model_patcher = self.setup_model_patcher(model)
612
+
613
+ def __call__(self, input_image, resolution, slider_1=None, slider_2=None, slider_3=None, **kwargs):
614
+ input_image, remove_pad = resize_image_with_pad(input_image, resolution)
615
+
616
+ self.load_model()
617
+
618
+ self.move_all_model_patchers_to_gpu()
619
+
620
+ assert input_image.ndim == 3
621
+ image_normal = input_image
622
+
623
+ with torch.no_grad():
624
+ image_normal = self.send_tensor_to_model_device(torch.from_numpy(image_normal))
625
+ image_normal = image_normal / 255.0
626
+ image_normal = rearrange(image_normal, 'h w c -> 1 c h w')
627
+ image_normal = self.norm(image_normal)
628
+
629
+ normal = self.model_patcher.model(image_normal)
630
+ normal = normal[0][-1][:, :3]
631
+ normal = ((normal + 1) * 0.5).clip(0, 1)
632
+
633
+ normal = rearrange(normal[0], 'c h w -> h w c').cpu().numpy()
634
+ normal_image = (normal * 255.0).clip(0, 255).astype(np.uint8)
635
+
636
+ return remove_pad(normal_image)
637
+
638
+
639
+ add_supported_preprocessor(PreprocessorNormalBae())
640
+
641
+ ```
642
+
643
+ # New features (that are not available in original WebUI)
644
+
645
+ Thanks to Unet Patcher, many new things are possible now and supported in Forge, including SVD, Z123, masked Ip-adapter, masked controlnet, photomaker, etc.
646
+
647
+ Masked Ip-Adapter
648
+
649
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/d26630f9-922d-4483-8bf9-f364dca5fd50)
650
+
651
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/03580ef7-235c-4b03-9ca6-a27677a5a175)
652
+
653
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/d9ed4a01-70d4-45b4-a6a7-2f765f158fae)
654
+
655
+ Masked ControlNet
656
+
657
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/872d4785-60e4-4431-85c7-665c781dddaa)
658
+
659
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/335a3b33-1ef8-46ff-a462-9f1b4f2c49fc)
660
+
661
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/b3684a15-8895-414e-8188-487269dfcada)
662
+
663
+ PhotoMaker
664
+
665
+ (Note that photomaker is a special control that need you to add the trigger word "photomaker". Your prompt should be like "a photo of photomaker")
666
+
667
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/07b0b626-05b5-473b-9d69-3657624d59be)
668
+
669
+ Marigold Depth
670
+
671
+ ![image](https://github.com/lllyasviel/stable-diffusion-webui-forge/assets/19834515/bdf54148-892d-410d-8ed9-70b4b121b6e7)
672
+
673
+ # New Samplers (that are not in origin)
674
+
675
+ DDPM
676
+ DDPM Karras
677
+ DPM++ 2M Turbo
678
+ DPM++ 2M SDE Turbo
679
+ LCM Karras
680
+ Euler A Turbo
681
+
682
+ # About Extensions
683
+
684
+ ControlNet and TiledVAE are integrated, and you should uninstall these two extensions:
685
+
686
+ sd-webui-controlnet
687
+ multidiffusion-upscaler-for-automatic1111
688
+
689
+ Note that **AnimateDiff** is under construction by [continue-revolution](https://github.com/continue-revolution) at [sd-webui-animatediff forge/master branch](https://github.com/continue-revolution/sd-webui-animatediff/tree/forge/master) and [sd-forge-animatediff](https://github.com/continue-revolution/sd-forge-animatediff) (they are in sync). (continue-revolution original words: prompt travel, inf t2v, controlnet v2v have been proven to work well; motion lora, i2i batch still under construction and may be finished in a week")
690
+
691
+ Other extensions should work without problems, like:
692
+
693
+ canvas-zoom
694
+ translations/localizations
695
+ Dynamic Prompts
696
+ Adetailer
697
+ Ultimate SD Upscale
698
+ Reactor
699
+
700
+ However, if newer extensions use Forge, their codes can be much shorter.
701
+
702
+ Usually if an old extension rework using Forge's unet patcher, 80% codes can be removed, especially when they need to call controlnet.
703
+
704
+ # Contribution
705
+
706
+ Forge uses a bot to get commits and codes from https://github.com/AUTOMATIC1111/stable-diffusion-webui/tree/dev every afternoon (if merge is automatically successful by a git bot, or by my compiler, or by my ChatGPT bot) or mid-night (if my compiler and my ChatGPT bot both failed to merge and I review it manually).
707
+
708
+ All PRs that can be implemented in https://github.com/AUTOMATIC1111/stable-diffusion-webui/tree/dev should submit PRs there.
709
+
710
+ Feel free to submit PRs related to the functionality of Forge here.
configs/alt-diffusion-inference.yaml ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ base_learning_rate: 1.0e-04
3
+ target: ldm.models.diffusion.ddpm.LatentDiffusion
4
+ params:
5
+ linear_start: 0.00085
6
+ linear_end: 0.0120
7
+ num_timesteps_cond: 1
8
+ log_every_t: 200
9
+ timesteps: 1000
10
+ first_stage_key: "jpg"
11
+ cond_stage_key: "txt"
12
+ image_size: 64
13
+ channels: 4
14
+ cond_stage_trainable: false # Note: different from the one we trained before
15
+ conditioning_key: crossattn
16
+ monitor: val/loss_simple_ema
17
+ scale_factor: 0.18215
18
+ use_ema: False
19
+
20
+ scheduler_config: # 10000 warmup steps
21
+ target: ldm.lr_scheduler.LambdaLinearScheduler
22
+ params:
23
+ warm_up_steps: [ 10000 ]
24
+ cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
25
+ f_start: [ 1.e-6 ]
26
+ f_max: [ 1. ]
27
+ f_min: [ 1. ]
28
+
29
+ unet_config:
30
+ target: ldm.modules.diffusionmodules.openaimodel.UNetModel
31
+ params:
32
+ image_size: 32 # unused
33
+ in_channels: 4
34
+ out_channels: 4
35
+ model_channels: 320
36
+ attention_resolutions: [ 4, 2, 1 ]
37
+ num_res_blocks: 2
38
+ channel_mult: [ 1, 2, 4, 4 ]
39
+ num_heads: 8
40
+ use_spatial_transformer: True
41
+ transformer_depth: 1
42
+ context_dim: 768
43
+ use_checkpoint: True
44
+ legacy: False
45
+
46
+ first_stage_config:
47
+ target: ldm.models.autoencoder.AutoencoderKL
48
+ params:
49
+ embed_dim: 4
50
+ monitor: val/rec_loss
51
+ ddconfig:
52
+ double_z: true
53
+ z_channels: 4
54
+ resolution: 256
55
+ in_channels: 3
56
+ out_ch: 3
57
+ ch: 128
58
+ ch_mult:
59
+ - 1
60
+ - 2
61
+ - 4
62
+ - 4
63
+ num_res_blocks: 2
64
+ attn_resolutions: []
65
+ dropout: 0.0
66
+ lossconfig:
67
+ target: torch.nn.Identity
68
+
69
+ cond_stage_config:
70
+ target: modules.xlmr.BertSeriesModelWithTransformation
71
+ params:
72
+ name: "XLMR-Large"
configs/alt-diffusion-m18-inference.yaml ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ base_learning_rate: 1.0e-04
3
+ target: ldm.models.diffusion.ddpm.LatentDiffusion
4
+ params:
5
+ linear_start: 0.00085
6
+ linear_end: 0.0120
7
+ num_timesteps_cond: 1
8
+ log_every_t: 200
9
+ timesteps: 1000
10
+ first_stage_key: "jpg"
11
+ cond_stage_key: "txt"
12
+ image_size: 64
13
+ channels: 4
14
+ cond_stage_trainable: false # Note: different from the one we trained before
15
+ conditioning_key: crossattn
16
+ monitor: val/loss_simple_ema
17
+ scale_factor: 0.18215
18
+ use_ema: False
19
+
20
+ scheduler_config: # 10000 warmup steps
21
+ target: ldm.lr_scheduler.LambdaLinearScheduler
22
+ params:
23
+ warm_up_steps: [ 10000 ]
24
+ cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
25
+ f_start: [ 1.e-6 ]
26
+ f_max: [ 1. ]
27
+ f_min: [ 1. ]
28
+
29
+ unet_config:
30
+ target: ldm.modules.diffusionmodules.openaimodel.UNetModel
31
+ params:
32
+ image_size: 32 # unused
33
+ in_channels: 4
34
+ out_channels: 4
35
+ model_channels: 320
36
+ attention_resolutions: [ 4, 2, 1 ]
37
+ num_res_blocks: 2
38
+ channel_mult: [ 1, 2, 4, 4 ]
39
+ num_head_channels: 64
40
+ use_spatial_transformer: True
41
+ use_linear_in_transformer: True
42
+ transformer_depth: 1
43
+ context_dim: 1024
44
+ use_checkpoint: True
45
+ legacy: False
46
+
47
+ first_stage_config:
48
+ target: ldm.models.autoencoder.AutoencoderKL
49
+ params:
50
+ embed_dim: 4
51
+ monitor: val/rec_loss
52
+ ddconfig:
53
+ double_z: true
54
+ z_channels: 4
55
+ resolution: 256
56
+ in_channels: 3
57
+ out_ch: 3
58
+ ch: 128
59
+ ch_mult:
60
+ - 1
61
+ - 2
62
+ - 4
63
+ - 4
64
+ num_res_blocks: 2
65
+ attn_resolutions: []
66
+ dropout: 0.0
67
+ lossconfig:
68
+ target: torch.nn.Identity
69
+
70
+ cond_stage_config:
71
+ target: modules.xlmr_m18.BertSeriesModelWithTransformation
72
+ params:
73
+ name: "XLMR-Large"
configs/instruct-pix2pix.yaml ADDED
@@ -0,0 +1,98 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # File modified by authors of InstructPix2Pix from original (https://github.com/CompVis/stable-diffusion).
2
+ # See more details in LICENSE.
3
+
4
+ model:
5
+ base_learning_rate: 1.0e-04
6
+ target: modules.models.diffusion.ddpm_edit.LatentDiffusion
7
+ params:
8
+ linear_start: 0.00085
9
+ linear_end: 0.0120
10
+ num_timesteps_cond: 1
11
+ log_every_t: 200
12
+ timesteps: 1000
13
+ first_stage_key: edited
14
+ cond_stage_key: edit
15
+ # image_size: 64
16
+ # image_size: 32
17
+ image_size: 16
18
+ channels: 4
19
+ cond_stage_trainable: false # Note: different from the one we trained before
20
+ conditioning_key: hybrid
21
+ monitor: val/loss_simple_ema
22
+ scale_factor: 0.18215
23
+ use_ema: false
24
+
25
+ scheduler_config: # 10000 warmup steps
26
+ target: ldm.lr_scheduler.LambdaLinearScheduler
27
+ params:
28
+ warm_up_steps: [ 0 ]
29
+ cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
30
+ f_start: [ 1.e-6 ]
31
+ f_max: [ 1. ]
32
+ f_min: [ 1. ]
33
+
34
+ unet_config:
35
+ target: ldm.modules.diffusionmodules.openaimodel.UNetModel
36
+ params:
37
+ image_size: 32 # unused
38
+ in_channels: 8
39
+ out_channels: 4
40
+ model_channels: 320
41
+ attention_resolutions: [ 4, 2, 1 ]
42
+ num_res_blocks: 2
43
+ channel_mult: [ 1, 2, 4, 4 ]
44
+ num_heads: 8
45
+ use_spatial_transformer: True
46
+ transformer_depth: 1
47
+ context_dim: 768
48
+ use_checkpoint: True
49
+ legacy: False
50
+
51
+ first_stage_config:
52
+ target: ldm.models.autoencoder.AutoencoderKL
53
+ params:
54
+ embed_dim: 4
55
+ monitor: val/rec_loss
56
+ ddconfig:
57
+ double_z: true
58
+ z_channels: 4
59
+ resolution: 256
60
+ in_channels: 3
61
+ out_ch: 3
62
+ ch: 128
63
+ ch_mult:
64
+ - 1
65
+ - 2
66
+ - 4
67
+ - 4
68
+ num_res_blocks: 2
69
+ attn_resolutions: []
70
+ dropout: 0.0
71
+ lossconfig:
72
+ target: torch.nn.Identity
73
+
74
+ cond_stage_config:
75
+ target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
76
+
77
+ data:
78
+ target: main.DataModuleFromConfig
79
+ params:
80
+ batch_size: 128
81
+ num_workers: 1
82
+ wrap: false
83
+ validation:
84
+ target: edit_dataset.EditDataset
85
+ params:
86
+ path: data/clip-filtered-dataset
87
+ cache_dir: data/
88
+ cache_name: data_10k
89
+ split: val
90
+ min_text_sim: 0.2
91
+ min_image_sim: 0.75
92
+ min_direction_sim: 0.2
93
+ max_samples_per_prompt: 1
94
+ min_resize_res: 512
95
+ max_resize_res: 512
96
+ crop_res: 512
97
+ output_as_edit: False
98
+ real_input: True
configs/sd_xl_inpaint.yaml ADDED
@@ -0,0 +1,98 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ target: sgm.models.diffusion.DiffusionEngine
3
+ params:
4
+ scale_factor: 0.13025
5
+ disable_first_stage_autocast: True
6
+
7
+ denoiser_config:
8
+ target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
9
+ params:
10
+ num_idx: 1000
11
+
12
+ weighting_config:
13
+ target: sgm.modules.diffusionmodules.denoiser_weighting.EpsWeighting
14
+ scaling_config:
15
+ target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling
16
+ discretization_config:
17
+ target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
18
+
19
+ network_config:
20
+ target: sgm.modules.diffusionmodules.openaimodel.UNetModel
21
+ params:
22
+ adm_in_channels: 2816
23
+ num_classes: sequential
24
+ use_checkpoint: True
25
+ in_channels: 9
26
+ out_channels: 4
27
+ model_channels: 320
28
+ attention_resolutions: [4, 2]
29
+ num_res_blocks: 2
30
+ channel_mult: [1, 2, 4]
31
+ num_head_channels: 64
32
+ use_spatial_transformer: True
33
+ use_linear_in_transformer: True
34
+ transformer_depth: [1, 2, 10] # note: the first is unused (due to attn_res starting at 2) 32, 16, 8 --> 64, 32, 16
35
+ context_dim: 2048
36
+ spatial_transformer_attn_type: softmax-xformers
37
+ legacy: False
38
+
39
+ conditioner_config:
40
+ target: sgm.modules.GeneralConditioner
41
+ params:
42
+ emb_models:
43
+ # crossattn cond
44
+ - is_trainable: False
45
+ input_key: txt
46
+ target: sgm.modules.encoders.modules.FrozenCLIPEmbedder
47
+ params:
48
+ layer: hidden
49
+ layer_idx: 11
50
+ # crossattn and vector cond
51
+ - is_trainable: False
52
+ input_key: txt
53
+ target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder2
54
+ params:
55
+ arch: ViT-bigG-14
56
+ version: laion2b_s39b_b160k
57
+ freeze: True
58
+ layer: penultimate
59
+ always_return_pooled: True
60
+ legacy: False
61
+ # vector cond
62
+ - is_trainable: False
63
+ input_key: original_size_as_tuple
64
+ target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
65
+ params:
66
+ outdim: 256 # multiplied by two
67
+ # vector cond
68
+ - is_trainable: False
69
+ input_key: crop_coords_top_left
70
+ target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
71
+ params:
72
+ outdim: 256 # multiplied by two
73
+ # vector cond
74
+ - is_trainable: False
75
+ input_key: target_size_as_tuple
76
+ target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
77
+ params:
78
+ outdim: 256 # multiplied by two
79
+
80
+ first_stage_config:
81
+ target: sgm.models.autoencoder.AutoencoderKLInferenceWrapper
82
+ params:
83
+ embed_dim: 4
84
+ monitor: val/rec_loss
85
+ ddconfig:
86
+ attn_type: vanilla-xformers
87
+ double_z: true
88
+ z_channels: 4
89
+ resolution: 256
90
+ in_channels: 3
91
+ out_ch: 3
92
+ ch: 128
93
+ ch_mult: [1, 2, 4, 4]
94
+ num_res_blocks: 2
95
+ attn_resolutions: []
96
+ dropout: 0.0
97
+ lossconfig:
98
+ target: torch.nn.Identity
configs/v1-inference.yaml ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ base_learning_rate: 1.0e-04
3
+ target: ldm.models.diffusion.ddpm.LatentDiffusion
4
+ params:
5
+ linear_start: 0.00085
6
+ linear_end: 0.0120
7
+ num_timesteps_cond: 1
8
+ log_every_t: 200
9
+ timesteps: 1000
10
+ first_stage_key: "jpg"
11
+ cond_stage_key: "txt"
12
+ image_size: 64
13
+ channels: 4
14
+ cond_stage_trainable: false # Note: different from the one we trained before
15
+ conditioning_key: crossattn
16
+ monitor: val/loss_simple_ema
17
+ scale_factor: 0.18215
18
+ use_ema: False
19
+
20
+ scheduler_config: # 10000 warmup steps
21
+ target: ldm.lr_scheduler.LambdaLinearScheduler
22
+ params:
23
+ warm_up_steps: [ 10000 ]
24
+ cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
25
+ f_start: [ 1.e-6 ]
26
+ f_max: [ 1. ]
27
+ f_min: [ 1. ]
28
+
29
+ unet_config:
30
+ target: ldm.modules.diffusionmodules.openaimodel.UNetModel
31
+ params:
32
+ image_size: 32 # unused
33
+ in_channels: 4
34
+ out_channels: 4
35
+ model_channels: 320
36
+ attention_resolutions: [ 4, 2, 1 ]
37
+ num_res_blocks: 2
38
+ channel_mult: [ 1, 2, 4, 4 ]
39
+ num_heads: 8
40
+ use_spatial_transformer: True
41
+ transformer_depth: 1
42
+ context_dim: 768
43
+ use_checkpoint: True
44
+ legacy: False
45
+
46
+ first_stage_config:
47
+ target: ldm.models.autoencoder.AutoencoderKL
48
+ params:
49
+ embed_dim: 4
50
+ monitor: val/rec_loss
51
+ ddconfig:
52
+ double_z: true
53
+ z_channels: 4
54
+ resolution: 256
55
+ in_channels: 3
56
+ out_ch: 3
57
+ ch: 128
58
+ ch_mult:
59
+ - 1
60
+ - 2
61
+ - 4
62
+ - 4
63
+ num_res_blocks: 2
64
+ attn_resolutions: []
65
+ dropout: 0.0
66
+ lossconfig:
67
+ target: torch.nn.Identity
68
+
69
+ cond_stage_config:
70
+ target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
configs/v1-inpainting-inference.yaml ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ base_learning_rate: 7.5e-05
3
+ target: ldm.models.diffusion.ddpm.LatentInpaintDiffusion
4
+ params:
5
+ linear_start: 0.00085
6
+ linear_end: 0.0120
7
+ num_timesteps_cond: 1
8
+ log_every_t: 200
9
+ timesteps: 1000
10
+ first_stage_key: "jpg"
11
+ cond_stage_key: "txt"
12
+ image_size: 64
13
+ channels: 4
14
+ cond_stage_trainable: false # Note: different from the one we trained before
15
+ conditioning_key: hybrid # important
16
+ monitor: val/loss_simple_ema
17
+ scale_factor: 0.18215
18
+ finetune_keys: null
19
+
20
+ scheduler_config: # 10000 warmup steps
21
+ target: ldm.lr_scheduler.LambdaLinearScheduler
22
+ params:
23
+ warm_up_steps: [ 2500 ] # NOTE for resuming. use 10000 if starting from scratch
24
+ cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
25
+ f_start: [ 1.e-6 ]
26
+ f_max: [ 1. ]
27
+ f_min: [ 1. ]
28
+
29
+ unet_config:
30
+ target: ldm.modules.diffusionmodules.openaimodel.UNetModel
31
+ params:
32
+ image_size: 32 # unused
33
+ in_channels: 9 # 4 data + 4 downscaled image + 1 mask
34
+ out_channels: 4
35
+ model_channels: 320
36
+ attention_resolutions: [ 4, 2, 1 ]
37
+ num_res_blocks: 2
38
+ channel_mult: [ 1, 2, 4, 4 ]
39
+ num_heads: 8
40
+ use_spatial_transformer: True
41
+ transformer_depth: 1
42
+ context_dim: 768
43
+ use_checkpoint: True
44
+ legacy: False
45
+
46
+ first_stage_config:
47
+ target: ldm.models.autoencoder.AutoencoderKL
48
+ params:
49
+ embed_dim: 4
50
+ monitor: val/rec_loss
51
+ ddconfig:
52
+ double_z: true
53
+ z_channels: 4
54
+ resolution: 256
55
+ in_channels: 3
56
+ out_ch: 3
57
+ ch: 128
58
+ ch_mult:
59
+ - 1
60
+ - 2
61
+ - 4
62
+ - 4
63
+ num_res_blocks: 2
64
+ attn_resolutions: []
65
+ dropout: 0.0
66
+ lossconfig:
67
+ target: torch.nn.Identity
68
+
69
+ cond_stage_config:
70
+ target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
environment-wsl2.yaml ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name: automatic
2
+ channels:
3
+ - pytorch
4
+ - defaults
5
+ dependencies:
6
+ - python=3.10
7
+ - pip=23.0
8
+ - cudatoolkit=11.8
9
+ - pytorch=2.0
10
+ - torchvision=0.15
11
+ - numpy=1.23
extensions-builtin/LDSR/ldsr_model_arch.py ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import gc
3
+ import time
4
+
5
+ import numpy as np
6
+ import torch
7
+ import torchvision
8
+ from PIL import Image
9
+ from einops import rearrange, repeat
10
+ from omegaconf import OmegaConf
11
+ import safetensors.torch
12
+
13
+ from ldm.models.diffusion.ddim import DDIMSampler
14
+ from ldm.util import instantiate_from_config, ismap
15
+ from modules import shared, sd_hijack, devices
16
+
17
+ cached_ldsr_model: torch.nn.Module = None
18
+
19
+
20
+ # Create LDSR Class
21
+ class LDSR:
22
+ def load_model_from_config(self, half_attention):
23
+ global cached_ldsr_model
24
+
25
+ if shared.opts.ldsr_cached and cached_ldsr_model is not None:
26
+ print("Loading model from cache")
27
+ model: torch.nn.Module = cached_ldsr_model
28
+ else:
29
+ print(f"Loading model from {self.modelPath}")
30
+ _, extension = os.path.splitext(self.modelPath)
31
+ if extension.lower() == ".safetensors":
32
+ pl_sd = safetensors.torch.load_file(self.modelPath, device="cpu")
33
+ else:
34
+ pl_sd = torch.load(self.modelPath, map_location="cpu")
35
+ sd = pl_sd["state_dict"] if "state_dict" in pl_sd else pl_sd
36
+ config = OmegaConf.load(self.yamlPath)
37
+ config.model.target = "ldm.models.diffusion.ddpm.LatentDiffusionV1"
38
+ model: torch.nn.Module = instantiate_from_config(config.model)
39
+ model.load_state_dict(sd, strict=False)
40
+ model = model.to(shared.device)
41
+ if half_attention:
42
+ model = model.half()
43
+ if shared.cmd_opts.opt_channelslast:
44
+ model = model.to(memory_format=torch.channels_last)
45
+
46
+ sd_hijack.model_hijack.hijack(model) # apply optimization
47
+ model.eval()
48
+
49
+ if shared.opts.ldsr_cached:
50
+ cached_ldsr_model = model
51
+
52
+ return {"model": model}
53
+
54
+ def __init__(self, model_path, yaml_path):
55
+ self.modelPath = model_path
56
+ self.yamlPath = yaml_path
57
+
58
+ @staticmethod
59
+ def run(model, selected_path, custom_steps, eta):
60
+ example = get_cond(selected_path)
61
+
62
+ n_runs = 1
63
+ guider = None
64
+ ckwargs = None
65
+ ddim_use_x0_pred = False
66
+ temperature = 1.
67
+ eta = eta
68
+ custom_shape = None
69
+
70
+ height, width = example["image"].shape[1:3]
71
+ split_input = height >= 128 and width >= 128
72
+
73
+ if split_input:
74
+ ks = 128
75
+ stride = 64
76
+ vqf = 4 #
77
+ model.split_input_params = {"ks": (ks, ks), "stride": (stride, stride),
78
+ "vqf": vqf,
79
+ "patch_distributed_vq": True,
80
+ "tie_braker": False,
81
+ "clip_max_weight": 0.5,
82
+ "clip_min_weight": 0.01,
83
+ "clip_max_tie_weight": 0.5,
84
+ "clip_min_tie_weight": 0.01}
85
+ else:
86
+ if hasattr(model, "split_input_params"):
87
+ delattr(model, "split_input_params")
88
+
89
+ x_t = None
90
+ logs = None
91
+ for _ in range(n_runs):
92
+ if custom_shape is not None:
93
+ x_t = torch.randn(1, custom_shape[1], custom_shape[2], custom_shape[3]).to(model.device)
94
+ x_t = repeat(x_t, '1 c h w -> b c h w', b=custom_shape[0])
95
+
96
+ logs = make_convolutional_sample(example, model,
97
+ custom_steps=custom_steps,
98
+ eta=eta, quantize_x0=False,
99
+ custom_shape=custom_shape,
100
+ temperature=temperature, noise_dropout=0.,
101
+ corrector=guider, corrector_kwargs=ckwargs, x_T=x_t,
102
+ ddim_use_x0_pred=ddim_use_x0_pred
103
+ )
104
+ return logs
105
+
106
+ def super_resolution(self, image, steps=100, target_scale=2, half_attention=False):
107
+ model = self.load_model_from_config(half_attention)
108
+
109
+ # Run settings
110
+ diffusion_steps = int(steps)
111
+ eta = 1.0
112
+
113
+
114
+ gc.collect()
115
+ devices.torch_gc()
116
+
117
+ im_og = image
118
+ width_og, height_og = im_og.size
119
+ # If we can adjust the max upscale size, then the 4 below should be our variable
120
+ down_sample_rate = target_scale / 4
121
+ wd = width_og * down_sample_rate
122
+ hd = height_og * down_sample_rate
123
+ width_downsampled_pre = int(np.ceil(wd))
124
+ height_downsampled_pre = int(np.ceil(hd))
125
+
126
+ if down_sample_rate != 1:
127
+ print(
128
+ f'Downsampling from [{width_og}, {height_og}] to [{width_downsampled_pre}, {height_downsampled_pre}]')
129
+ im_og = im_og.resize((width_downsampled_pre, height_downsampled_pre), Image.LANCZOS)
130
+ else:
131
+ print(f"Down sample rate is 1 from {target_scale} / 4 (Not downsampling)")
132
+
133
+ # pad width and height to multiples of 64, pads with the edge values of image to avoid artifacts
134
+ pad_w, pad_h = np.max(((2, 2), np.ceil(np.array(im_og.size) / 64).astype(int)), axis=0) * 64 - im_og.size
135
+ im_padded = Image.fromarray(np.pad(np.array(im_og), ((0, pad_h), (0, pad_w), (0, 0)), mode='edge'))
136
+
137
+ logs = self.run(model["model"], im_padded, diffusion_steps, eta)
138
+
139
+ sample = logs["sample"]
140
+ sample = sample.detach().cpu()
141
+ sample = torch.clamp(sample, -1., 1.)
142
+ sample = (sample + 1.) / 2. * 255
143
+ sample = sample.numpy().astype(np.uint8)
144
+ sample = np.transpose(sample, (0, 2, 3, 1))
145
+ a = Image.fromarray(sample[0])
146
+
147
+ # remove padding
148
+ a = a.crop((0, 0) + tuple(np.array(im_og.size) * 4))
149
+
150
+ del model
151
+ gc.collect()
152
+ devices.torch_gc()
153
+
154
+ return a
155
+
156
+
157
+ def get_cond(selected_path):
158
+ example = {}
159
+ up_f = 4
160
+ c = selected_path.convert('RGB')
161
+ c = torch.unsqueeze(torchvision.transforms.ToTensor()(c), 0)
162
+ c_up = torchvision.transforms.functional.resize(c, size=[up_f * c.shape[2], up_f * c.shape[3]],
163
+ antialias=True)
164
+ c_up = rearrange(c_up, '1 c h w -> 1 h w c')
165
+ c = rearrange(c, '1 c h w -> 1 h w c')
166
+ c = 2. * c - 1.
167
+
168
+ c = c.to(shared.device)
169
+ example["LR_image"] = c
170
+ example["image"] = c_up
171
+
172
+ return example
173
+
174
+
175
+ @torch.no_grad()
176
+ def convsample_ddim(model, cond, steps, shape, eta=1.0, callback=None, normals_sequence=None,
177
+ mask=None, x0=None, quantize_x0=False, temperature=1., score_corrector=None,
178
+ corrector_kwargs=None, x_t=None
179
+ ):
180
+ ddim = DDIMSampler(model)
181
+ bs = shape[0]
182
+ shape = shape[1:]
183
+ print(f"Sampling with eta = {eta}; steps: {steps}")
184
+ samples, intermediates = ddim.sample(steps, batch_size=bs, shape=shape, conditioning=cond, callback=callback,
185
+ normals_sequence=normals_sequence, quantize_x0=quantize_x0, eta=eta,
186
+ mask=mask, x0=x0, temperature=temperature, verbose=False,
187
+ score_corrector=score_corrector,
188
+ corrector_kwargs=corrector_kwargs, x_t=x_t)
189
+
190
+ return samples, intermediates
191
+
192
+
193
+ @torch.no_grad()
194
+ def make_convolutional_sample(batch, model, custom_steps=None, eta=1.0, quantize_x0=False, custom_shape=None, temperature=1., noise_dropout=0., corrector=None,
195
+ corrector_kwargs=None, x_T=None, ddim_use_x0_pred=False):
196
+ log = {}
197
+
198
+ z, c, x, xrec, xc = model.get_input(batch, model.first_stage_key,
199
+ return_first_stage_outputs=True,
200
+ force_c_encode=not (hasattr(model, 'split_input_params')
201
+ and model.cond_stage_key == 'coordinates_bbox'),
202
+ return_original_cond=True)
203
+
204
+ if custom_shape is not None:
205
+ z = torch.randn(custom_shape)
206
+ print(f"Generating {custom_shape[0]} samples of shape {custom_shape[1:]}")
207
+
208
+ z0 = None
209
+
210
+ log["input"] = x
211
+ log["reconstruction"] = xrec
212
+
213
+ if ismap(xc):
214
+ log["original_conditioning"] = model.to_rgb(xc)
215
+ if hasattr(model, 'cond_stage_key'):
216
+ log[model.cond_stage_key] = model.to_rgb(xc)
217
+
218
+ else:
219
+ log["original_conditioning"] = xc if xc is not None else torch.zeros_like(x)
220
+ if model.cond_stage_model:
221
+ log[model.cond_stage_key] = xc if xc is not None else torch.zeros_like(x)
222
+ if model.cond_stage_key == 'class_label':
223
+ log[model.cond_stage_key] = xc[model.cond_stage_key]
224
+
225
+ with model.ema_scope("Plotting"):
226
+ t0 = time.time()
227
+
228
+ sample, intermediates = convsample_ddim(model, c, steps=custom_steps, shape=z.shape,
229
+ eta=eta,
230
+ quantize_x0=quantize_x0, mask=None, x0=z0,
231
+ temperature=temperature, score_corrector=corrector, corrector_kwargs=corrector_kwargs,
232
+ x_t=x_T)
233
+ t1 = time.time()
234
+
235
+ if ddim_use_x0_pred:
236
+ sample = intermediates['pred_x0'][-1]
237
+
238
+ x_sample = model.decode_first_stage(sample)
239
+
240
+ try:
241
+ x_sample_noquant = model.decode_first_stage(sample, force_not_quantize=True)
242
+ log["sample_noquant"] = x_sample_noquant
243
+ log["sample_diff"] = torch.abs(x_sample_noquant - x_sample)
244
+ except Exception:
245
+ pass
246
+
247
+ log["sample"] = x_sample
248
+ log["time"] = t1 - t0
249
+
250
+ return log
extensions-builtin/LDSR/preload.py ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ import os
2
+ from modules import paths
3
+
4
+
5
+ def preload(parser):
6
+ parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with LDSR model file(s).", default=os.path.join(paths.models_path, 'LDSR'))
extensions-builtin/LDSR/scripts/ldsr_model.py ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+
3
+ from modules.modelloader import load_file_from_url
4
+ from modules.upscaler import Upscaler, UpscalerData
5
+ from modules_forge.forge_util import prepare_free_memory
6
+ from ldsr_model_arch import LDSR
7
+ from modules import shared, script_callbacks, errors
8
+ import sd_hijack_autoencoder # noqa: F401
9
+ import sd_hijack_ddpm_v1 # noqa: F401
10
+
11
+
12
+ class UpscalerLDSR(Upscaler):
13
+ def __init__(self, user_path):
14
+ self.name = "LDSR"
15
+ self.user_path = user_path
16
+ self.model_url = "https://heibox.uni-heidelberg.de/f/578df07c8fc04ffbadf3/?dl=1"
17
+ self.yaml_url = "https://heibox.uni-heidelberg.de/f/31a76b13ea27482981b4/?dl=1"
18
+ super().__init__()
19
+ scaler_data = UpscalerData("LDSR", None, self)
20
+ self.scalers = [scaler_data]
21
+
22
+ def load_model(self, path: str):
23
+ # Remove incorrect project.yaml file if too big
24
+ yaml_path = os.path.join(self.model_path, "project.yaml")
25
+ old_model_path = os.path.join(self.model_path, "model.pth")
26
+ new_model_path = os.path.join(self.model_path, "model.ckpt")
27
+
28
+ local_model_paths = self.find_models(ext_filter=[".ckpt", ".safetensors"])
29
+ local_ckpt_path = next(iter([local_model for local_model in local_model_paths if local_model.endswith("model.ckpt")]), None)
30
+ local_safetensors_path = next(iter([local_model for local_model in local_model_paths if local_model.endswith("model.safetensors")]), None)
31
+ local_yaml_path = next(iter([local_model for local_model in local_model_paths if local_model.endswith("project.yaml")]), None)
32
+
33
+ if os.path.exists(yaml_path):
34
+ statinfo = os.stat(yaml_path)
35
+ if statinfo.st_size >= 10485760:
36
+ print("Removing invalid LDSR YAML file.")
37
+ os.remove(yaml_path)
38
+
39
+ if os.path.exists(old_model_path):
40
+ print("Renaming model from model.pth to model.ckpt")
41
+ os.rename(old_model_path, new_model_path)
42
+
43
+ if local_safetensors_path is not None and os.path.exists(local_safetensors_path):
44
+ model = local_safetensors_path
45
+ else:
46
+ model = local_ckpt_path or load_file_from_url(self.model_url, model_dir=self.model_download_path, file_name="model.ckpt")
47
+
48
+ yaml = local_yaml_path or load_file_from_url(self.yaml_url, model_dir=self.model_download_path, file_name="project.yaml")
49
+
50
+ return LDSR(model, yaml)
51
+
52
+ def do_upscale(self, img, path):
53
+ prepare_free_memory(aggressive=True)
54
+ try:
55
+ ldsr = self.load_model(path)
56
+ except Exception:
57
+ errors.report(f"Failed loading LDSR model {path}", exc_info=True)
58
+ return img
59
+ ddim_steps = shared.opts.ldsr_steps
60
+ return ldsr.super_resolution(img, ddim_steps, self.scale)
61
+
62
+
63
+ def on_ui_settings():
64
+ import gradio as gr
65
+
66
+ shared.opts.add_option("ldsr_steps", shared.OptionInfo(100, "LDSR processing steps. Lower = faster", gr.Slider, {"minimum": 1, "maximum": 200, "step": 1}, section=('upscaling', "Upscaling")))
67
+ shared.opts.add_option("ldsr_cached", shared.OptionInfo(False, "Cache LDSR model in memory", gr.Checkbox, {"interactive": True}, section=('upscaling', "Upscaling")))
68
+
69
+
70
+ script_callbacks.on_ui_settings(on_ui_settings)
extensions-builtin/LDSR/sd_hijack_autoencoder.py ADDED
@@ -0,0 +1,293 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # The content of this file comes from the ldm/models/autoencoder.py file of the compvis/stable-diffusion repo
2
+ # The VQModel & VQModelInterface were subsequently removed from ldm/models/autoencoder.py when we moved to the stability-ai/stablediffusion repo
3
+ # As the LDSR upscaler relies on VQModel & VQModelInterface, the hijack aims to put them back into the ldm.models.autoencoder
4
+ import numpy as np
5
+ import torch
6
+ import pytorch_lightning as pl
7
+ import torch.nn.functional as F
8
+ from contextlib import contextmanager
9
+
10
+ from torch.optim.lr_scheduler import LambdaLR
11
+
12
+ from ldm.modules.ema import LitEma
13
+ from vqvae_quantize import VectorQuantizer2 as VectorQuantizer
14
+ from ldm.modules.diffusionmodules.model import Encoder, Decoder
15
+ from ldm.util import instantiate_from_config
16
+
17
+ import ldm.models.autoencoder
18
+ from packaging import version
19
+
20
+ class VQModel(pl.LightningModule):
21
+ def __init__(self,
22
+ ddconfig,
23
+ lossconfig,
24
+ n_embed,
25
+ embed_dim,
26
+ ckpt_path=None,
27
+ ignore_keys=None,
28
+ image_key="image",
29
+ colorize_nlabels=None,
30
+ monitor=None,
31
+ batch_resize_range=None,
32
+ scheduler_config=None,
33
+ lr_g_factor=1.0,
34
+ remap=None,
35
+ sane_index_shape=False, # tell vector quantizer to return indices as bhw
36
+ use_ema=False
37
+ ):
38
+ super().__init__()
39
+ self.embed_dim = embed_dim
40
+ self.n_embed = n_embed
41
+ self.image_key = image_key
42
+ self.encoder = Encoder(**ddconfig)
43
+ self.decoder = Decoder(**ddconfig)
44
+ self.loss = instantiate_from_config(lossconfig)
45
+ self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25,
46
+ remap=remap,
47
+ sane_index_shape=sane_index_shape)
48
+ self.quant_conv = torch.nn.Conv2d(ddconfig["z_channels"], embed_dim, 1)
49
+ self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
50
+ if colorize_nlabels is not None:
51
+ assert type(colorize_nlabels)==int
52
+ self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
53
+ if monitor is not None:
54
+ self.monitor = monitor
55
+ self.batch_resize_range = batch_resize_range
56
+ if self.batch_resize_range is not None:
57
+ print(f"{self.__class__.__name__}: Using per-batch resizing in range {batch_resize_range}.")
58
+
59
+ self.use_ema = use_ema
60
+ if self.use_ema:
61
+ self.model_ema = LitEma(self)
62
+ print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
63
+
64
+ if ckpt_path is not None:
65
+ self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys or [])
66
+ self.scheduler_config = scheduler_config
67
+ self.lr_g_factor = lr_g_factor
68
+
69
+ @contextmanager
70
+ def ema_scope(self, context=None):
71
+ if self.use_ema:
72
+ self.model_ema.store(self.parameters())
73
+ self.model_ema.copy_to(self)
74
+ if context is not None:
75
+ print(f"{context}: Switched to EMA weights")
76
+ try:
77
+ yield None
78
+ finally:
79
+ if self.use_ema:
80
+ self.model_ema.restore(self.parameters())
81
+ if context is not None:
82
+ print(f"{context}: Restored training weights")
83
+
84
+ def init_from_ckpt(self, path, ignore_keys=None):
85
+ sd = torch.load(path, map_location="cpu")["state_dict"]
86
+ keys = list(sd.keys())
87
+ for k in keys:
88
+ for ik in ignore_keys or []:
89
+ if k.startswith(ik):
90
+ print("Deleting key {} from state_dict.".format(k))
91
+ del sd[k]
92
+ missing, unexpected = self.load_state_dict(sd, strict=False)
93
+ print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
94
+ if missing:
95
+ print(f"Missing Keys: {missing}")
96
+ if unexpected:
97
+ print(f"Unexpected Keys: {unexpected}")
98
+
99
+ def on_train_batch_end(self, *args, **kwargs):
100
+ if self.use_ema:
101
+ self.model_ema(self)
102
+
103
+ def encode(self, x):
104
+ h = self.encoder(x)
105
+ h = self.quant_conv(h)
106
+ quant, emb_loss, info = self.quantize(h)
107
+ return quant, emb_loss, info
108
+
109
+ def encode_to_prequant(self, x):
110
+ h = self.encoder(x)
111
+ h = self.quant_conv(h)
112
+ return h
113
+
114
+ def decode(self, quant):
115
+ quant = self.post_quant_conv(quant)
116
+ dec = self.decoder(quant)
117
+ return dec
118
+
119
+ def decode_code(self, code_b):
120
+ quant_b = self.quantize.embed_code(code_b)
121
+ dec = self.decode(quant_b)
122
+ return dec
123
+
124
+ def forward(self, input, return_pred_indices=False):
125
+ quant, diff, (_,_,ind) = self.encode(input)
126
+ dec = self.decode(quant)
127
+ if return_pred_indices:
128
+ return dec, diff, ind
129
+ return dec, diff
130
+
131
+ def get_input(self, batch, k):
132
+ x = batch[k]
133
+ if len(x.shape) == 3:
134
+ x = x[..., None]
135
+ x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
136
+ if self.batch_resize_range is not None:
137
+ lower_size = self.batch_resize_range[0]
138
+ upper_size = self.batch_resize_range[1]
139
+ if self.global_step <= 4:
140
+ # do the first few batches with max size to avoid later oom
141
+ new_resize = upper_size
142
+ else:
143
+ new_resize = np.random.choice(np.arange(lower_size, upper_size+16, 16))
144
+ if new_resize != x.shape[2]:
145
+ x = F.interpolate(x, size=new_resize, mode="bicubic")
146
+ x = x.detach()
147
+ return x
148
+
149
+ def training_step(self, batch, batch_idx, optimizer_idx):
150
+ # https://github.com/pytorch/pytorch/issues/37142
151
+ # try not to fool the heuristics
152
+ x = self.get_input(batch, self.image_key)
153
+ xrec, qloss, ind = self(x, return_pred_indices=True)
154
+
155
+ if optimizer_idx == 0:
156
+ # autoencode
157
+ aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
158
+ last_layer=self.get_last_layer(), split="train",
159
+ predicted_indices=ind)
160
+
161
+ self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True)
162
+ return aeloss
163
+
164
+ if optimizer_idx == 1:
165
+ # discriminator
166
+ discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
167
+ last_layer=self.get_last_layer(), split="train")
168
+ self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True)
169
+ return discloss
170
+
171
+ def validation_step(self, batch, batch_idx):
172
+ log_dict = self._validation_step(batch, batch_idx)
173
+ with self.ema_scope():
174
+ self._validation_step(batch, batch_idx, suffix="_ema")
175
+ return log_dict
176
+
177
+ def _validation_step(self, batch, batch_idx, suffix=""):
178
+ x = self.get_input(batch, self.image_key)
179
+ xrec, qloss, ind = self(x, return_pred_indices=True)
180
+ aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0,
181
+ self.global_step,
182
+ last_layer=self.get_last_layer(),
183
+ split="val"+suffix,
184
+ predicted_indices=ind
185
+ )
186
+
187
+ discloss, log_dict_disc = self.loss(qloss, x, xrec, 1,
188
+ self.global_step,
189
+ last_layer=self.get_last_layer(),
190
+ split="val"+suffix,
191
+ predicted_indices=ind
192
+ )
193
+ rec_loss = log_dict_ae[f"val{suffix}/rec_loss"]
194
+ self.log(f"val{suffix}/rec_loss", rec_loss,
195
+ prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
196
+ self.log(f"val{suffix}/aeloss", aeloss,
197
+ prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
198
+ if version.parse(pl.__version__) >= version.parse('1.4.0'):
199
+ del log_dict_ae[f"val{suffix}/rec_loss"]
200
+ self.log_dict(log_dict_ae)
201
+ self.log_dict(log_dict_disc)
202
+ return self.log_dict
203
+
204
+ def configure_optimizers(self):
205
+ lr_d = self.learning_rate
206
+ lr_g = self.lr_g_factor*self.learning_rate
207
+ print("lr_d", lr_d)
208
+ print("lr_g", lr_g)
209
+ opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
210
+ list(self.decoder.parameters())+
211
+ list(self.quantize.parameters())+
212
+ list(self.quant_conv.parameters())+
213
+ list(self.post_quant_conv.parameters()),
214
+ lr=lr_g, betas=(0.5, 0.9))
215
+ opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
216
+ lr=lr_d, betas=(0.5, 0.9))
217
+
218
+ if self.scheduler_config is not None:
219
+ scheduler = instantiate_from_config(self.scheduler_config)
220
+
221
+ print("Setting up LambdaLR scheduler...")
222
+ scheduler = [
223
+ {
224
+ 'scheduler': LambdaLR(opt_ae, lr_lambda=scheduler.schedule),
225
+ 'interval': 'step',
226
+ 'frequency': 1
227
+ },
228
+ {
229
+ 'scheduler': LambdaLR(opt_disc, lr_lambda=scheduler.schedule),
230
+ 'interval': 'step',
231
+ 'frequency': 1
232
+ },
233
+ ]
234
+ return [opt_ae, opt_disc], scheduler
235
+ return [opt_ae, opt_disc], []
236
+
237
+ def get_last_layer(self):
238
+ return self.decoder.conv_out.weight
239
+
240
+ def log_images(self, batch, only_inputs=False, plot_ema=False, **kwargs):
241
+ log = {}
242
+ x = self.get_input(batch, self.image_key)
243
+ x = x.to(self.device)
244
+ if only_inputs:
245
+ log["inputs"] = x
246
+ return log
247
+ xrec, _ = self(x)
248
+ if x.shape[1] > 3:
249
+ # colorize with random projection
250
+ assert xrec.shape[1] > 3
251
+ x = self.to_rgb(x)
252
+ xrec = self.to_rgb(xrec)
253
+ log["inputs"] = x
254
+ log["reconstructions"] = xrec
255
+ if plot_ema:
256
+ with self.ema_scope():
257
+ xrec_ema, _ = self(x)
258
+ if x.shape[1] > 3:
259
+ xrec_ema = self.to_rgb(xrec_ema)
260
+ log["reconstructions_ema"] = xrec_ema
261
+ return log
262
+
263
+ def to_rgb(self, x):
264
+ assert self.image_key == "segmentation"
265
+ if not hasattr(self, "colorize"):
266
+ self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
267
+ x = F.conv2d(x, weight=self.colorize)
268
+ x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
269
+ return x
270
+
271
+
272
+ class VQModelInterface(VQModel):
273
+ def __init__(self, embed_dim, *args, **kwargs):
274
+ super().__init__(*args, embed_dim=embed_dim, **kwargs)
275
+ self.embed_dim = embed_dim
276
+
277
+ def encode(self, x):
278
+ h = self.encoder(x)
279
+ h = self.quant_conv(h)
280
+ return h
281
+
282
+ def decode(self, h, force_not_quantize=False):
283
+ # also go through quantization layer
284
+ if not force_not_quantize:
285
+ quant, emb_loss, info = self.quantize(h)
286
+ else:
287
+ quant = h
288
+ quant = self.post_quant_conv(quant)
289
+ dec = self.decoder(quant)
290
+ return dec
291
+
292
+ ldm.models.autoencoder.VQModel = VQModel
293
+ ldm.models.autoencoder.VQModelInterface = VQModelInterface
extensions-builtin/LDSR/sd_hijack_ddpm_v1.py ADDED
@@ -0,0 +1,1443 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # This script is copied from the compvis/stable-diffusion repo (aka the SD V1 repo)
2
+ # Original filename: ldm/models/diffusion/ddpm.py
3
+ # The purpose to reinstate the old DDPM logic which works with VQ, whereas the V2 one doesn't
4
+ # Some models such as LDSR require VQ to work correctly
5
+ # The classes are suffixed with "V1" and added back to the "ldm.models.diffusion.ddpm" module
6
+
7
+ import torch
8
+ import torch.nn as nn
9
+ import numpy as np
10
+ import pytorch_lightning as pl
11
+ from torch.optim.lr_scheduler import LambdaLR
12
+ from einops import rearrange, repeat
13
+ from contextlib import contextmanager
14
+ from functools import partial
15
+ from tqdm import tqdm
16
+ from torchvision.utils import make_grid
17
+ from pytorch_lightning.utilities.distributed import rank_zero_only
18
+
19
+ from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config
20
+ from ldm.modules.ema import LitEma
21
+ from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution
22
+ from ldm.models.autoencoder import VQModelInterface, IdentityFirstStage, AutoencoderKL
23
+ from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
24
+ from ldm.models.diffusion.ddim import DDIMSampler
25
+
26
+ import ldm.models.diffusion.ddpm
27
+
28
+ __conditioning_keys__ = {'concat': 'c_concat',
29
+ 'crossattn': 'c_crossattn',
30
+ 'adm': 'y'}
31
+
32
+
33
+ def disabled_train(self, mode=True):
34
+ """Overwrite model.train with this function to make sure train/eval mode
35
+ does not change anymore."""
36
+ return self
37
+
38
+
39
+ def uniform_on_device(r1, r2, shape, device):
40
+ return (r1 - r2) * torch.rand(*shape, device=device) + r2
41
+
42
+
43
+ class DDPMV1(pl.LightningModule):
44
+ # classic DDPM with Gaussian diffusion, in image space
45
+ def __init__(self,
46
+ unet_config,
47
+ timesteps=1000,
48
+ beta_schedule="linear",
49
+ loss_type="l2",
50
+ ckpt_path=None,
51
+ ignore_keys=None,
52
+ load_only_unet=False,
53
+ monitor="val/loss",
54
+ use_ema=True,
55
+ first_stage_key="image",
56
+ image_size=256,
57
+ channels=3,
58
+ log_every_t=100,
59
+ clip_denoised=True,
60
+ linear_start=1e-4,
61
+ linear_end=2e-2,
62
+ cosine_s=8e-3,
63
+ given_betas=None,
64
+ original_elbo_weight=0.,
65
+ v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
66
+ l_simple_weight=1.,
67
+ conditioning_key=None,
68
+ parameterization="eps", # all assuming fixed variance schedules
69
+ scheduler_config=None,
70
+ use_positional_encodings=False,
71
+ learn_logvar=False,
72
+ logvar_init=0.,
73
+ ):
74
+ super().__init__()
75
+ assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"'
76
+ self.parameterization = parameterization
77
+ print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
78
+ self.cond_stage_model = None
79
+ self.clip_denoised = clip_denoised
80
+ self.log_every_t = log_every_t
81
+ self.first_stage_key = first_stage_key
82
+ self.image_size = image_size # try conv?
83
+ self.channels = channels
84
+ self.use_positional_encodings = use_positional_encodings
85
+ self.model = DiffusionWrapperV1(unet_config, conditioning_key)
86
+ count_params(self.model, verbose=True)
87
+ self.use_ema = use_ema
88
+ if self.use_ema:
89
+ self.model_ema = LitEma(self.model)
90
+ print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
91
+
92
+ self.use_scheduler = scheduler_config is not None
93
+ if self.use_scheduler:
94
+ self.scheduler_config = scheduler_config
95
+
96
+ self.v_posterior = v_posterior
97
+ self.original_elbo_weight = original_elbo_weight
98
+ self.l_simple_weight = l_simple_weight
99
+
100
+ if monitor is not None:
101
+ self.monitor = monitor
102
+ if ckpt_path is not None:
103
+ self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys or [], only_model=load_only_unet)
104
+
105
+ self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
106
+ linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
107
+
108
+ self.loss_type = loss_type
109
+
110
+ self.learn_logvar = learn_logvar
111
+ self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
112
+ if self.learn_logvar:
113
+ self.logvar = nn.Parameter(self.logvar, requires_grad=True)
114
+
115
+
116
+ def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
117
+ linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
118
+ if exists(given_betas):
119
+ betas = given_betas
120
+ else:
121
+ betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
122
+ cosine_s=cosine_s)
123
+ alphas = 1. - betas
124
+ alphas_cumprod = np.cumprod(alphas, axis=0)
125
+ alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
126
+
127
+ timesteps, = betas.shape
128
+ self.num_timesteps = int(timesteps)
129
+ self.linear_start = linear_start
130
+ self.linear_end = linear_end
131
+ assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
132
+
133
+ to_torch = partial(torch.tensor, dtype=torch.float32)
134
+
135
+ self.register_buffer('betas', to_torch(betas))
136
+ self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
137
+ self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
138
+
139
+ # calculations for diffusion q(x_t | x_{t-1}) and others
140
+ self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
141
+ self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
142
+ self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
143
+ self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
144
+ self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
145
+
146
+ # calculations for posterior q(x_{t-1} | x_t, x_0)
147
+ posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (
148
+ 1. - alphas_cumprod) + self.v_posterior * betas
149
+ # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
150
+ self.register_buffer('posterior_variance', to_torch(posterior_variance))
151
+ # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
152
+ self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
153
+ self.register_buffer('posterior_mean_coef1', to_torch(
154
+ betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
155
+ self.register_buffer('posterior_mean_coef2', to_torch(
156
+ (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
157
+
158
+ if self.parameterization == "eps":
159
+ lvlb_weights = self.betas ** 2 / (
160
+ 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
161
+ elif self.parameterization == "x0":
162
+ lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod))
163
+ else:
164
+ raise NotImplementedError("mu not supported")
165
+ # TODO how to choose this term
166
+ lvlb_weights[0] = lvlb_weights[1]
167
+ self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
168
+ assert not torch.isnan(self.lvlb_weights).all()
169
+
170
+ @contextmanager
171
+ def ema_scope(self, context=None):
172
+ if self.use_ema:
173
+ self.model_ema.store(self.model.parameters())
174
+ self.model_ema.copy_to(self.model)
175
+ if context is not None:
176
+ print(f"{context}: Switched to EMA weights")
177
+ try:
178
+ yield None
179
+ finally:
180
+ if self.use_ema:
181
+ self.model_ema.restore(self.model.parameters())
182
+ if context is not None:
183
+ print(f"{context}: Restored training weights")
184
+
185
+ def init_from_ckpt(self, path, ignore_keys=None, only_model=False):
186
+ sd = torch.load(path, map_location="cpu")
187
+ if "state_dict" in list(sd.keys()):
188
+ sd = sd["state_dict"]
189
+ keys = list(sd.keys())
190
+ for k in keys:
191
+ for ik in ignore_keys or []:
192
+ if k.startswith(ik):
193
+ print("Deleting key {} from state_dict.".format(k))
194
+ del sd[k]
195
+ missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
196
+ sd, strict=False)
197
+ print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
198
+ if missing:
199
+ print(f"Missing Keys: {missing}")
200
+ if unexpected:
201
+ print(f"Unexpected Keys: {unexpected}")
202
+
203
+ def q_mean_variance(self, x_start, t):
204
+ """
205
+ Get the distribution q(x_t | x_0).
206
+ :param x_start: the [N x C x ...] tensor of noiseless inputs.
207
+ :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
208
+ :return: A tuple (mean, variance, log_variance), all of x_start's shape.
209
+ """
210
+ mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
211
+ variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
212
+ log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
213
+ return mean, variance, log_variance
214
+
215
+ def predict_start_from_noise(self, x_t, t, noise):
216
+ return (
217
+ extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
218
+ extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
219
+ )
220
+
221
+ def q_posterior(self, x_start, x_t, t):
222
+ posterior_mean = (
223
+ extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
224
+ extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
225
+ )
226
+ posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
227
+ posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
228
+ return posterior_mean, posterior_variance, posterior_log_variance_clipped
229
+
230
+ def p_mean_variance(self, x, t, clip_denoised: bool):
231
+ model_out = self.model(x, t)
232
+ if self.parameterization == "eps":
233
+ x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
234
+ elif self.parameterization == "x0":
235
+ x_recon = model_out
236
+ if clip_denoised:
237
+ x_recon.clamp_(-1., 1.)
238
+
239
+ model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
240
+ return model_mean, posterior_variance, posterior_log_variance
241
+
242
+ @torch.no_grad()
243
+ def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
244
+ b, *_, device = *x.shape, x.device
245
+ model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
246
+ noise = noise_like(x.shape, device, repeat_noise)
247
+ # no noise when t == 0
248
+ nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
249
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
250
+
251
+ @torch.no_grad()
252
+ def p_sample_loop(self, shape, return_intermediates=False):
253
+ device = self.betas.device
254
+ b = shape[0]
255
+ img = torch.randn(shape, device=device)
256
+ intermediates = [img]
257
+ for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
258
+ img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
259
+ clip_denoised=self.clip_denoised)
260
+ if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
261
+ intermediates.append(img)
262
+ if return_intermediates:
263
+ return img, intermediates
264
+ return img
265
+
266
+ @torch.no_grad()
267
+ def sample(self, batch_size=16, return_intermediates=False):
268
+ image_size = self.image_size
269
+ channels = self.channels
270
+ return self.p_sample_loop((batch_size, channels, image_size, image_size),
271
+ return_intermediates=return_intermediates)
272
+
273
+ def q_sample(self, x_start, t, noise=None):
274
+ noise = default(noise, lambda: torch.randn_like(x_start))
275
+ return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
276
+ extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
277
+
278
+ def get_loss(self, pred, target, mean=True):
279
+ if self.loss_type == 'l1':
280
+ loss = (target - pred).abs()
281
+ if mean:
282
+ loss = loss.mean()
283
+ elif self.loss_type == 'l2':
284
+ if mean:
285
+ loss = torch.nn.functional.mse_loss(target, pred)
286
+ else:
287
+ loss = torch.nn.functional.mse_loss(target, pred, reduction='none')
288
+ else:
289
+ raise NotImplementedError("unknown loss type '{loss_type}'")
290
+
291
+ return loss
292
+
293
+ def p_losses(self, x_start, t, noise=None):
294
+ noise = default(noise, lambda: torch.randn_like(x_start))
295
+ x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
296
+ model_out = self.model(x_noisy, t)
297
+
298
+ loss_dict = {}
299
+ if self.parameterization == "eps":
300
+ target = noise
301
+ elif self.parameterization == "x0":
302
+ target = x_start
303
+ else:
304
+ raise NotImplementedError(f"Paramterization {self.parameterization} not yet supported")
305
+
306
+ loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
307
+
308
+ log_prefix = 'train' if self.training else 'val'
309
+
310
+ loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()})
311
+ loss_simple = loss.mean() * self.l_simple_weight
312
+
313
+ loss_vlb = (self.lvlb_weights[t] * loss).mean()
314
+ loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb})
315
+
316
+ loss = loss_simple + self.original_elbo_weight * loss_vlb
317
+
318
+ loss_dict.update({f'{log_prefix}/loss': loss})
319
+
320
+ return loss, loss_dict
321
+
322
+ def forward(self, x, *args, **kwargs):
323
+ # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
324
+ # assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
325
+ t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
326
+ return self.p_losses(x, t, *args, **kwargs)
327
+
328
+ def get_input(self, batch, k):
329
+ x = batch[k]
330
+ if len(x.shape) == 3:
331
+ x = x[..., None]
332
+ x = rearrange(x, 'b h w c -> b c h w')
333
+ x = x.to(memory_format=torch.contiguous_format).float()
334
+ return x
335
+
336
+ def shared_step(self, batch):
337
+ x = self.get_input(batch, self.first_stage_key)
338
+ loss, loss_dict = self(x)
339
+ return loss, loss_dict
340
+
341
+ def training_step(self, batch, batch_idx):
342
+ loss, loss_dict = self.shared_step(batch)
343
+
344
+ self.log_dict(loss_dict, prog_bar=True,
345
+ logger=True, on_step=True, on_epoch=True)
346
+
347
+ self.log("global_step", self.global_step,
348
+ prog_bar=True, logger=True, on_step=True, on_epoch=False)
349
+
350
+ if self.use_scheduler:
351
+ lr = self.optimizers().param_groups[0]['lr']
352
+ self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
353
+
354
+ return loss
355
+
356
+ @torch.no_grad()
357
+ def validation_step(self, batch, batch_idx):
358
+ _, loss_dict_no_ema = self.shared_step(batch)
359
+ with self.ema_scope():
360
+ _, loss_dict_ema = self.shared_step(batch)
361
+ loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema}
362
+ self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
363
+ self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
364
+
365
+ def on_train_batch_end(self, *args, **kwargs):
366
+ if self.use_ema:
367
+ self.model_ema(self.model)
368
+
369
+ def _get_rows_from_list(self, samples):
370
+ n_imgs_per_row = len(samples)
371
+ denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
372
+ denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
373
+ denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
374
+ return denoise_grid
375
+
376
+ @torch.no_grad()
377
+ def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
378
+ log = {}
379
+ x = self.get_input(batch, self.first_stage_key)
380
+ N = min(x.shape[0], N)
381
+ n_row = min(x.shape[0], n_row)
382
+ x = x.to(self.device)[:N]
383
+ log["inputs"] = x
384
+
385
+ # get diffusion row
386
+ diffusion_row = []
387
+ x_start = x[:n_row]
388
+
389
+ for t in range(self.num_timesteps):
390
+ if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
391
+ t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
392
+ t = t.to(self.device).long()
393
+ noise = torch.randn_like(x_start)
394
+ x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
395
+ diffusion_row.append(x_noisy)
396
+
397
+ log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
398
+
399
+ if sample:
400
+ # get denoise row
401
+ with self.ema_scope("Plotting"):
402
+ samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
403
+
404
+ log["samples"] = samples
405
+ log["denoise_row"] = self._get_rows_from_list(denoise_row)
406
+
407
+ if return_keys:
408
+ if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
409
+ return log
410
+ else:
411
+ return {key: log[key] for key in return_keys}
412
+ return log
413
+
414
+ def configure_optimizers(self):
415
+ lr = self.learning_rate
416
+ params = list(self.model.parameters())
417
+ if self.learn_logvar:
418
+ params = params + [self.logvar]
419
+ opt = torch.optim.AdamW(params, lr=lr)
420
+ return opt
421
+
422
+
423
+ class LatentDiffusionV1(DDPMV1):
424
+ """main class"""
425
+ def __init__(self,
426
+ first_stage_config,
427
+ cond_stage_config,
428
+ num_timesteps_cond=None,
429
+ cond_stage_key="image",
430
+ cond_stage_trainable=False,
431
+ concat_mode=True,
432
+ cond_stage_forward=None,
433
+ conditioning_key=None,
434
+ scale_factor=1.0,
435
+ scale_by_std=False,
436
+ *args, **kwargs):
437
+ self.num_timesteps_cond = default(num_timesteps_cond, 1)
438
+ self.scale_by_std = scale_by_std
439
+ assert self.num_timesteps_cond <= kwargs['timesteps']
440
+ # for backwards compatibility after implementation of DiffusionWrapper
441
+ if conditioning_key is None:
442
+ conditioning_key = 'concat' if concat_mode else 'crossattn'
443
+ if cond_stage_config == '__is_unconditional__':
444
+ conditioning_key = None
445
+ ckpt_path = kwargs.pop("ckpt_path", None)
446
+ ignore_keys = kwargs.pop("ignore_keys", [])
447
+ super().__init__(*args, conditioning_key=conditioning_key, **kwargs)
448
+ self.concat_mode = concat_mode
449
+ self.cond_stage_trainable = cond_stage_trainable
450
+ self.cond_stage_key = cond_stage_key
451
+ try:
452
+ self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
453
+ except Exception:
454
+ self.num_downs = 0
455
+ if not scale_by_std:
456
+ self.scale_factor = scale_factor
457
+ else:
458
+ self.register_buffer('scale_factor', torch.tensor(scale_factor))
459
+ self.instantiate_first_stage(first_stage_config)
460
+ self.instantiate_cond_stage(cond_stage_config)
461
+ self.cond_stage_forward = cond_stage_forward
462
+ self.clip_denoised = False
463
+ self.bbox_tokenizer = None
464
+
465
+ self.restarted_from_ckpt = False
466
+ if ckpt_path is not None:
467
+ self.init_from_ckpt(ckpt_path, ignore_keys)
468
+ self.restarted_from_ckpt = True
469
+
470
+ def make_cond_schedule(self, ):
471
+ self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
472
+ ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
473
+ self.cond_ids[:self.num_timesteps_cond] = ids
474
+
475
+ @rank_zero_only
476
+ @torch.no_grad()
477
+ def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
478
+ # only for very first batch
479
+ if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
480
+ assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
481
+ # set rescale weight to 1./std of encodings
482
+ print("### USING STD-RESCALING ###")
483
+ x = super().get_input(batch, self.first_stage_key)
484
+ x = x.to(self.device)
485
+ encoder_posterior = self.encode_first_stage(x)
486
+ z = self.get_first_stage_encoding(encoder_posterior).detach()
487
+ del self.scale_factor
488
+ self.register_buffer('scale_factor', 1. / z.flatten().std())
489
+ print(f"setting self.scale_factor to {self.scale_factor}")
490
+ print("### USING STD-RESCALING ###")
491
+
492
+ def register_schedule(self,
493
+ given_betas=None, beta_schedule="linear", timesteps=1000,
494
+ linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
495
+ super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
496
+
497
+ self.shorten_cond_schedule = self.num_timesteps_cond > 1
498
+ if self.shorten_cond_schedule:
499
+ self.make_cond_schedule()
500
+
501
+ def instantiate_first_stage(self, config):
502
+ model = instantiate_from_config(config)
503
+ self.first_stage_model = model.eval()
504
+ self.first_stage_model.train = disabled_train
505
+ for param in self.first_stage_model.parameters():
506
+ param.requires_grad = False
507
+
508
+ def instantiate_cond_stage(self, config):
509
+ if not self.cond_stage_trainable:
510
+ if config == "__is_first_stage__":
511
+ print("Using first stage also as cond stage.")
512
+ self.cond_stage_model = self.first_stage_model
513
+ elif config == "__is_unconditional__":
514
+ print(f"Training {self.__class__.__name__} as an unconditional model.")
515
+ self.cond_stage_model = None
516
+ # self.be_unconditional = True
517
+ else:
518
+ model = instantiate_from_config(config)
519
+ self.cond_stage_model = model.eval()
520
+ self.cond_stage_model.train = disabled_train
521
+ for param in self.cond_stage_model.parameters():
522
+ param.requires_grad = False
523
+ else:
524
+ assert config != '__is_first_stage__'
525
+ assert config != '__is_unconditional__'
526
+ model = instantiate_from_config(config)
527
+ self.cond_stage_model = model
528
+
529
+ def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False):
530
+ denoise_row = []
531
+ for zd in tqdm(samples, desc=desc):
532
+ denoise_row.append(self.decode_first_stage(zd.to(self.device),
533
+ force_not_quantize=force_no_decoder_quantization))
534
+ n_imgs_per_row = len(denoise_row)
535
+ denoise_row = torch.stack(denoise_row) # n_log_step, n_row, C, H, W
536
+ denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
537
+ denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
538
+ denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
539
+ return denoise_grid
540
+
541
+ def get_first_stage_encoding(self, encoder_posterior):
542
+ if isinstance(encoder_posterior, DiagonalGaussianDistribution):
543
+ z = encoder_posterior.sample()
544
+ elif isinstance(encoder_posterior, torch.Tensor):
545
+ z = encoder_posterior
546
+ else:
547
+ raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
548
+ return self.scale_factor * z
549
+
550
+ def get_learned_conditioning(self, c):
551
+ if self.cond_stage_forward is None:
552
+ if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
553
+ c = self.cond_stage_model.encode(c)
554
+ if isinstance(c, DiagonalGaussianDistribution):
555
+ c = c.mode()
556
+ else:
557
+ c = self.cond_stage_model(c)
558
+ else:
559
+ assert hasattr(self.cond_stage_model, self.cond_stage_forward)
560
+ c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
561
+ return c
562
+
563
+ def meshgrid(self, h, w):
564
+ y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
565
+ x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)
566
+
567
+ arr = torch.cat([y, x], dim=-1)
568
+ return arr
569
+
570
+ def delta_border(self, h, w):
571
+ """
572
+ :param h: height
573
+ :param w: width
574
+ :return: normalized distance to image border,
575
+ wtith min distance = 0 at border and max dist = 0.5 at image center
576
+ """
577
+ lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
578
+ arr = self.meshgrid(h, w) / lower_right_corner
579
+ dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
580
+ dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
581
+ edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0]
582
+ return edge_dist
583
+
584
+ def get_weighting(self, h, w, Ly, Lx, device):
585
+ weighting = self.delta_border(h, w)
586
+ weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"],
587
+ self.split_input_params["clip_max_weight"], )
588
+ weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)
589
+
590
+ if self.split_input_params["tie_braker"]:
591
+ L_weighting = self.delta_border(Ly, Lx)
592
+ L_weighting = torch.clip(L_weighting,
593
+ self.split_input_params["clip_min_tie_weight"],
594
+ self.split_input_params["clip_max_tie_weight"])
595
+
596
+ L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
597
+ weighting = weighting * L_weighting
598
+ return weighting
599
+
600
+ def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1): # todo load once not every time, shorten code
601
+ """
602
+ :param x: img of size (bs, c, h, w)
603
+ :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
604
+ """
605
+ bs, nc, h, w = x.shape
606
+
607
+ # number of crops in image
608
+ Ly = (h - kernel_size[0]) // stride[0] + 1
609
+ Lx = (w - kernel_size[1]) // stride[1] + 1
610
+
611
+ if uf == 1 and df == 1:
612
+ fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
613
+ unfold = torch.nn.Unfold(**fold_params)
614
+
615
+ fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)
616
+
617
+ weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype)
618
+ normalization = fold(weighting).view(1, 1, h, w) # normalizes the overlap
619
+ weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))
620
+
621
+ elif uf > 1 and df == 1:
622
+ fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
623
+ unfold = torch.nn.Unfold(**fold_params)
624
+
625
+ fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
626
+ dilation=1, padding=0,
627
+ stride=(stride[0] * uf, stride[1] * uf))
628
+ fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2)
629
+
630
+ weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype)
631
+ normalization = fold(weighting).view(1, 1, h * uf, w * uf) # normalizes the overlap
632
+ weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx))
633
+
634
+ elif df > 1 and uf == 1:
635
+ fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
636
+ unfold = torch.nn.Unfold(**fold_params)
637
+
638
+ fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
639
+ dilation=1, padding=0,
640
+ stride=(stride[0] // df, stride[1] // df))
641
+ fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2)
642
+
643
+ weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype)
644
+ normalization = fold(weighting).view(1, 1, h // df, w // df) # normalizes the overlap
645
+ weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx))
646
+
647
+ else:
648
+ raise NotImplementedError
649
+
650
+ return fold, unfold, normalization, weighting
651
+
652
+ @torch.no_grad()
653
+ def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False,
654
+ cond_key=None, return_original_cond=False, bs=None):
655
+ x = super().get_input(batch, k)
656
+ if bs is not None:
657
+ x = x[:bs]
658
+ x = x.to(self.device)
659
+ encoder_posterior = self.encode_first_stage(x)
660
+ z = self.get_first_stage_encoding(encoder_posterior).detach()
661
+
662
+ if self.model.conditioning_key is not None:
663
+ if cond_key is None:
664
+ cond_key = self.cond_stage_key
665
+ if cond_key != self.first_stage_key:
666
+ if cond_key in ['caption', 'coordinates_bbox']:
667
+ xc = batch[cond_key]
668
+ elif cond_key == 'class_label':
669
+ xc = batch
670
+ else:
671
+ xc = super().get_input(batch, cond_key).to(self.device)
672
+ else:
673
+ xc = x
674
+ if not self.cond_stage_trainable or force_c_encode:
675
+ if isinstance(xc, dict) or isinstance(xc, list):
676
+ # import pudb; pudb.set_trace()
677
+ c = self.get_learned_conditioning(xc)
678
+ else:
679
+ c = self.get_learned_conditioning(xc.to(self.device))
680
+ else:
681
+ c = xc
682
+ if bs is not None:
683
+ c = c[:bs]
684
+
685
+ if self.use_positional_encodings:
686
+ pos_x, pos_y = self.compute_latent_shifts(batch)
687
+ ckey = __conditioning_keys__[self.model.conditioning_key]
688
+ c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y}
689
+
690
+ else:
691
+ c = None
692
+ xc = None
693
+ if self.use_positional_encodings:
694
+ pos_x, pos_y = self.compute_latent_shifts(batch)
695
+ c = {'pos_x': pos_x, 'pos_y': pos_y}
696
+ out = [z, c]
697
+ if return_first_stage_outputs:
698
+ xrec = self.decode_first_stage(z)
699
+ out.extend([x, xrec])
700
+ if return_original_cond:
701
+ out.append(xc)
702
+ return out
703
+
704
+ @torch.no_grad()
705
+ def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
706
+ if predict_cids:
707
+ if z.dim() == 4:
708
+ z = torch.argmax(z.exp(), dim=1).long()
709
+ z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
710
+ z = rearrange(z, 'b h w c -> b c h w').contiguous()
711
+
712
+ z = 1. / self.scale_factor * z
713
+
714
+ if hasattr(self, "split_input_params"):
715
+ if self.split_input_params["patch_distributed_vq"]:
716
+ ks = self.split_input_params["ks"] # eg. (128, 128)
717
+ stride = self.split_input_params["stride"] # eg. (64, 64)
718
+ uf = self.split_input_params["vqf"]
719
+ bs, nc, h, w = z.shape
720
+ if ks[0] > h or ks[1] > w:
721
+ ks = (min(ks[0], h), min(ks[1], w))
722
+ print("reducing Kernel")
723
+
724
+ if stride[0] > h or stride[1] > w:
725
+ stride = (min(stride[0], h), min(stride[1], w))
726
+ print("reducing stride")
727
+
728
+ fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
729
+
730
+ z = unfold(z) # (bn, nc * prod(**ks), L)
731
+ # 1. Reshape to img shape
732
+ z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
733
+
734
+ # 2. apply model loop over last dim
735
+ if isinstance(self.first_stage_model, VQModelInterface):
736
+ output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
737
+ force_not_quantize=predict_cids or force_not_quantize)
738
+ for i in range(z.shape[-1])]
739
+ else:
740
+
741
+ output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
742
+ for i in range(z.shape[-1])]
743
+
744
+ o = torch.stack(output_list, axis=-1) # # (bn, nc, ks[0], ks[1], L)
745
+ o = o * weighting
746
+ # Reverse 1. reshape to img shape
747
+ o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
748
+ # stitch crops together
749
+ decoded = fold(o)
750
+ decoded = decoded / normalization # norm is shape (1, 1, h, w)
751
+ return decoded
752
+ else:
753
+ if isinstance(self.first_stage_model, VQModelInterface):
754
+ return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
755
+ else:
756
+ return self.first_stage_model.decode(z)
757
+
758
+ else:
759
+ if isinstance(self.first_stage_model, VQModelInterface):
760
+ return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
761
+ else:
762
+ return self.first_stage_model.decode(z)
763
+
764
+ # same as above but without decorator
765
+ def differentiable_decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
766
+ if predict_cids:
767
+ if z.dim() == 4:
768
+ z = torch.argmax(z.exp(), dim=1).long()
769
+ z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
770
+ z = rearrange(z, 'b h w c -> b c h w').contiguous()
771
+
772
+ z = 1. / self.scale_factor * z
773
+
774
+ if hasattr(self, "split_input_params"):
775
+ if self.split_input_params["patch_distributed_vq"]:
776
+ ks = self.split_input_params["ks"] # eg. (128, 128)
777
+ stride = self.split_input_params["stride"] # eg. (64, 64)
778
+ uf = self.split_input_params["vqf"]
779
+ bs, nc, h, w = z.shape
780
+ if ks[0] > h or ks[1] > w:
781
+ ks = (min(ks[0], h), min(ks[1], w))
782
+ print("reducing Kernel")
783
+
784
+ if stride[0] > h or stride[1] > w:
785
+ stride = (min(stride[0], h), min(stride[1], w))
786
+ print("reducing stride")
787
+
788
+ fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
789
+
790
+ z = unfold(z) # (bn, nc * prod(**ks), L)
791
+ # 1. Reshape to img shape
792
+ z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
793
+
794
+ # 2. apply model loop over last dim
795
+ if isinstance(self.first_stage_model, VQModelInterface):
796
+ output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
797
+ force_not_quantize=predict_cids or force_not_quantize)
798
+ for i in range(z.shape[-1])]
799
+ else:
800
+
801
+ output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
802
+ for i in range(z.shape[-1])]
803
+
804
+ o = torch.stack(output_list, axis=-1) # # (bn, nc, ks[0], ks[1], L)
805
+ o = o * weighting
806
+ # Reverse 1. reshape to img shape
807
+ o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
808
+ # stitch crops together
809
+ decoded = fold(o)
810
+ decoded = decoded / normalization # norm is shape (1, 1, h, w)
811
+ return decoded
812
+ else:
813
+ if isinstance(self.first_stage_model, VQModelInterface):
814
+ return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
815
+ else:
816
+ return self.first_stage_model.decode(z)
817
+
818
+ else:
819
+ if isinstance(self.first_stage_model, VQModelInterface):
820
+ return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
821
+ else:
822
+ return self.first_stage_model.decode(z)
823
+
824
+ @torch.no_grad()
825
+ def encode_first_stage(self, x):
826
+ if hasattr(self, "split_input_params"):
827
+ if self.split_input_params["patch_distributed_vq"]:
828
+ ks = self.split_input_params["ks"] # eg. (128, 128)
829
+ stride = self.split_input_params["stride"] # eg. (64, 64)
830
+ df = self.split_input_params["vqf"]
831
+ self.split_input_params['original_image_size'] = x.shape[-2:]
832
+ bs, nc, h, w = x.shape
833
+ if ks[0] > h or ks[1] > w:
834
+ ks = (min(ks[0], h), min(ks[1], w))
835
+ print("reducing Kernel")
836
+
837
+ if stride[0] > h or stride[1] > w:
838
+ stride = (min(stride[0], h), min(stride[1], w))
839
+ print("reducing stride")
840
+
841
+ fold, unfold, normalization, weighting = self.get_fold_unfold(x, ks, stride, df=df)
842
+ z = unfold(x) # (bn, nc * prod(**ks), L)
843
+ # Reshape to img shape
844
+ z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
845
+
846
+ output_list = [self.first_stage_model.encode(z[:, :, :, :, i])
847
+ for i in range(z.shape[-1])]
848
+
849
+ o = torch.stack(output_list, axis=-1)
850
+ o = o * weighting
851
+
852
+ # Reverse reshape to img shape
853
+ o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
854
+ # stitch crops together
855
+ decoded = fold(o)
856
+ decoded = decoded / normalization
857
+ return decoded
858
+
859
+ else:
860
+ return self.first_stage_model.encode(x)
861
+ else:
862
+ return self.first_stage_model.encode(x)
863
+
864
+ def shared_step(self, batch, **kwargs):
865
+ x, c = self.get_input(batch, self.first_stage_key)
866
+ loss = self(x, c)
867
+ return loss
868
+
869
+ def forward(self, x, c, *args, **kwargs):
870
+ t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
871
+ if self.model.conditioning_key is not None:
872
+ assert c is not None
873
+ if self.cond_stage_trainable:
874
+ c = self.get_learned_conditioning(c)
875
+ if self.shorten_cond_schedule: # TODO: drop this option
876
+ tc = self.cond_ids[t].to(self.device)
877
+ c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float()))
878
+ return self.p_losses(x, c, t, *args, **kwargs)
879
+
880
+ def apply_model(self, x_noisy, t, cond, return_ids=False):
881
+
882
+ if isinstance(cond, dict):
883
+ # hybrid case, cond is exptected to be a dict
884
+ pass
885
+ else:
886
+ if not isinstance(cond, list):
887
+ cond = [cond]
888
+ key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
889
+ cond = {key: cond}
890
+
891
+ if hasattr(self, "split_input_params"):
892
+ assert len(cond) == 1 # todo can only deal with one conditioning atm
893
+ assert not return_ids
894
+ ks = self.split_input_params["ks"] # eg. (128, 128)
895
+ stride = self.split_input_params["stride"] # eg. (64, 64)
896
+
897
+ h, w = x_noisy.shape[-2:]
898
+
899
+ fold, unfold, normalization, weighting = self.get_fold_unfold(x_noisy, ks, stride)
900
+
901
+ z = unfold(x_noisy) # (bn, nc * prod(**ks), L)
902
+ # Reshape to img shape
903
+ z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
904
+ z_list = [z[:, :, :, :, i] for i in range(z.shape[-1])]
905
+
906
+ if self.cond_stage_key in ["image", "LR_image", "segmentation",
907
+ 'bbox_img'] and self.model.conditioning_key: # todo check for completeness
908
+ c_key = next(iter(cond.keys())) # get key
909
+ c = next(iter(cond.values())) # get value
910
+ assert (len(c) == 1) # todo extend to list with more than one elem
911
+ c = c[0] # get element
912
+
913
+ c = unfold(c)
914
+ c = c.view((c.shape[0], -1, ks[0], ks[1], c.shape[-1])) # (bn, nc, ks[0], ks[1], L )
915
+
916
+ cond_list = [{c_key: [c[:, :, :, :, i]]} for i in range(c.shape[-1])]
917
+
918
+ elif self.cond_stage_key == 'coordinates_bbox':
919
+ assert 'original_image_size' in self.split_input_params, 'BoudingBoxRescaling is missing original_image_size'
920
+
921
+ # assuming padding of unfold is always 0 and its dilation is always 1
922
+ n_patches_per_row = int((w - ks[0]) / stride[0] + 1)
923
+ full_img_h, full_img_w = self.split_input_params['original_image_size']
924
+ # as we are operating on latents, we need the factor from the original image size to the
925
+ # spatial latent size to properly rescale the crops for regenerating the bbox annotations
926
+ num_downs = self.first_stage_model.encoder.num_resolutions - 1
927
+ rescale_latent = 2 ** (num_downs)
928
+
929
+ # get top left postions of patches as conforming for the bbbox tokenizer, therefore we
930
+ # need to rescale the tl patch coordinates to be in between (0,1)
931
+ tl_patch_coordinates = [(rescale_latent * stride[0] * (patch_nr % n_patches_per_row) / full_img_w,
932
+ rescale_latent * stride[1] * (patch_nr // n_patches_per_row) / full_img_h)
933
+ for patch_nr in range(z.shape[-1])]
934
+
935
+ # patch_limits are tl_coord, width and height coordinates as (x_tl, y_tl, h, w)
936
+ patch_limits = [(x_tl, y_tl,
937
+ rescale_latent * ks[0] / full_img_w,
938
+ rescale_latent * ks[1] / full_img_h) for x_tl, y_tl in tl_patch_coordinates]
939
+ # patch_values = [(np.arange(x_tl,min(x_tl+ks, 1.)),np.arange(y_tl,min(y_tl+ks, 1.))) for x_tl, y_tl in tl_patch_coordinates]
940
+
941
+ # tokenize crop coordinates for the bounding boxes of the respective patches
942
+ patch_limits_tknzd = [torch.LongTensor(self.bbox_tokenizer._crop_encoder(bbox))[None].to(self.device)
943
+ for bbox in patch_limits] # list of length l with tensors of shape (1, 2)
944
+ print(patch_limits_tknzd[0].shape)
945
+ # cut tknzd crop position from conditioning
946
+ assert isinstance(cond, dict), 'cond must be dict to be fed into model'
947
+ cut_cond = cond['c_crossattn'][0][..., :-2].to(self.device)
948
+ print(cut_cond.shape)
949
+
950
+ adapted_cond = torch.stack([torch.cat([cut_cond, p], dim=1) for p in patch_limits_tknzd])
951
+ adapted_cond = rearrange(adapted_cond, 'l b n -> (l b) n')
952
+ print(adapted_cond.shape)
953
+ adapted_cond = self.get_learned_conditioning(adapted_cond)
954
+ print(adapted_cond.shape)
955
+ adapted_cond = rearrange(adapted_cond, '(l b) n d -> l b n d', l=z.shape[-1])
956
+ print(adapted_cond.shape)
957
+
958
+ cond_list = [{'c_crossattn': [e]} for e in adapted_cond]
959
+
960
+ else:
961
+ cond_list = [cond for i in range(z.shape[-1])] # Todo make this more efficient
962
+
963
+ # apply model by loop over crops
964
+ output_list = [self.model(z_list[i], t, **cond_list[i]) for i in range(z.shape[-1])]
965
+ assert not isinstance(output_list[0],
966
+ tuple) # todo cant deal with multiple model outputs check this never happens
967
+
968
+ o = torch.stack(output_list, axis=-1)
969
+ o = o * weighting
970
+ # Reverse reshape to img shape
971
+ o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
972
+ # stitch crops together
973
+ x_recon = fold(o) / normalization
974
+
975
+ else:
976
+ x_recon = self.model(x_noisy, t, **cond)
977
+
978
+ if isinstance(x_recon, tuple) and not return_ids:
979
+ return x_recon[0]
980
+ else:
981
+ return x_recon
982
+
983
+ def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
984
+ return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \
985
+ extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
986
+
987
+ def _prior_bpd(self, x_start):
988
+ """
989
+ Get the prior KL term for the variational lower-bound, measured in
990
+ bits-per-dim.
991
+ This term can't be optimized, as it only depends on the encoder.
992
+ :param x_start: the [N x C x ...] tensor of inputs.
993
+ :return: a batch of [N] KL values (in bits), one per batch element.
994
+ """
995
+ batch_size = x_start.shape[0]
996
+ t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
997
+ qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
998
+ kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0)
999
+ return mean_flat(kl_prior) / np.log(2.0)
1000
+
1001
+ def p_losses(self, x_start, cond, t, noise=None):
1002
+ noise = default(noise, lambda: torch.randn_like(x_start))
1003
+ x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
1004
+ model_output = self.apply_model(x_noisy, t, cond)
1005
+
1006
+ loss_dict = {}
1007
+ prefix = 'train' if self.training else 'val'
1008
+
1009
+ if self.parameterization == "x0":
1010
+ target = x_start
1011
+ elif self.parameterization == "eps":
1012
+ target = noise
1013
+ else:
1014
+ raise NotImplementedError()
1015
+
1016
+ loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3])
1017
+ loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
1018
+
1019
+ logvar_t = self.logvar[t].to(self.device)
1020
+ loss = loss_simple / torch.exp(logvar_t) + logvar_t
1021
+ # loss = loss_simple / torch.exp(self.logvar) + self.logvar
1022
+ if self.learn_logvar:
1023
+ loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
1024
+ loss_dict.update({'logvar': self.logvar.data.mean()})
1025
+
1026
+ loss = self.l_simple_weight * loss.mean()
1027
+
1028
+ loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3))
1029
+ loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
1030
+ loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
1031
+ loss += (self.original_elbo_weight * loss_vlb)
1032
+ loss_dict.update({f'{prefix}/loss': loss})
1033
+
1034
+ return loss, loss_dict
1035
+
1036
+ def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False,
1037
+ return_x0=False, score_corrector=None, corrector_kwargs=None):
1038
+ t_in = t
1039
+ model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
1040
+
1041
+ if score_corrector is not None:
1042
+ assert self.parameterization == "eps"
1043
+ model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
1044
+
1045
+ if return_codebook_ids:
1046
+ model_out, logits = model_out
1047
+
1048
+ if self.parameterization == "eps":
1049
+ x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
1050
+ elif self.parameterization == "x0":
1051
+ x_recon = model_out
1052
+ else:
1053
+ raise NotImplementedError()
1054
+
1055
+ if clip_denoised:
1056
+ x_recon.clamp_(-1., 1.)
1057
+ if quantize_denoised:
1058
+ x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
1059
+ model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
1060
+ if return_codebook_ids:
1061
+ return model_mean, posterior_variance, posterior_log_variance, logits
1062
+ elif return_x0:
1063
+ return model_mean, posterior_variance, posterior_log_variance, x_recon
1064
+ else:
1065
+ return model_mean, posterior_variance, posterior_log_variance
1066
+
1067
+ @torch.no_grad()
1068
+ def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False,
1069
+ return_codebook_ids=False, quantize_denoised=False, return_x0=False,
1070
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None):
1071
+ b, *_, device = *x.shape, x.device
1072
+ outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised,
1073
+ return_codebook_ids=return_codebook_ids,
1074
+ quantize_denoised=quantize_denoised,
1075
+ return_x0=return_x0,
1076
+ score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
1077
+ if return_codebook_ids:
1078
+ raise DeprecationWarning("Support dropped.")
1079
+ model_mean, _, model_log_variance, logits = outputs
1080
+ elif return_x0:
1081
+ model_mean, _, model_log_variance, x0 = outputs
1082
+ else:
1083
+ model_mean, _, model_log_variance = outputs
1084
+
1085
+ noise = noise_like(x.shape, device, repeat_noise) * temperature
1086
+ if noise_dropout > 0.:
1087
+ noise = torch.nn.functional.dropout(noise, p=noise_dropout)
1088
+ # no noise when t == 0
1089
+ nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
1090
+
1091
+ if return_codebook_ids:
1092
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1)
1093
+ if return_x0:
1094
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
1095
+ else:
1096
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
1097
+
1098
+ @torch.no_grad()
1099
+ def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False,
1100
+ img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0.,
1101
+ score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None,
1102
+ log_every_t=None):
1103
+ if not log_every_t:
1104
+ log_every_t = self.log_every_t
1105
+ timesteps = self.num_timesteps
1106
+ if batch_size is not None:
1107
+ b = batch_size if batch_size is not None else shape[0]
1108
+ shape = [batch_size] + list(shape)
1109
+ else:
1110
+ b = batch_size = shape[0]
1111
+ if x_T is None:
1112
+ img = torch.randn(shape, device=self.device)
1113
+ else:
1114
+ img = x_T
1115
+ intermediates = []
1116
+ if cond is not None:
1117
+ if isinstance(cond, dict):
1118
+ cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
1119
+ [x[:batch_size] for x in cond[key]] for key in cond}
1120
+ else:
1121
+ cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
1122
+
1123
+ if start_T is not None:
1124
+ timesteps = min(timesteps, start_T)
1125
+ iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation',
1126
+ total=timesteps) if verbose else reversed(
1127
+ range(0, timesteps))
1128
+ if type(temperature) == float:
1129
+ temperature = [temperature] * timesteps
1130
+
1131
+ for i in iterator:
1132
+ ts = torch.full((b,), i, device=self.device, dtype=torch.long)
1133
+ if self.shorten_cond_schedule:
1134
+ assert self.model.conditioning_key != 'hybrid'
1135
+ tc = self.cond_ids[ts].to(cond.device)
1136
+ cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
1137
+
1138
+ img, x0_partial = self.p_sample(img, cond, ts,
1139
+ clip_denoised=self.clip_denoised,
1140
+ quantize_denoised=quantize_denoised, return_x0=True,
1141
+ temperature=temperature[i], noise_dropout=noise_dropout,
1142
+ score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
1143
+ if mask is not None:
1144
+ assert x0 is not None
1145
+ img_orig = self.q_sample(x0, ts)
1146
+ img = img_orig * mask + (1. - mask) * img
1147
+
1148
+ if i % log_every_t == 0 or i == timesteps - 1:
1149
+ intermediates.append(x0_partial)
1150
+ if callback:
1151
+ callback(i)
1152
+ if img_callback:
1153
+ img_callback(img, i)
1154
+ return img, intermediates
1155
+
1156
+ @torch.no_grad()
1157
+ def p_sample_loop(self, cond, shape, return_intermediates=False,
1158
+ x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False,
1159
+ mask=None, x0=None, img_callback=None, start_T=None,
1160
+ log_every_t=None):
1161
+
1162
+ if not log_every_t:
1163
+ log_every_t = self.log_every_t
1164
+ device = self.betas.device
1165
+ b = shape[0]
1166
+ if x_T is None:
1167
+ img = torch.randn(shape, device=device)
1168
+ else:
1169
+ img = x_T
1170
+
1171
+ intermediates = [img]
1172
+ if timesteps is None:
1173
+ timesteps = self.num_timesteps
1174
+
1175
+ if start_T is not None:
1176
+ timesteps = min(timesteps, start_T)
1177
+ iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(
1178
+ range(0, timesteps))
1179
+
1180
+ if mask is not None:
1181
+ assert x0 is not None
1182
+ assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match
1183
+
1184
+ for i in iterator:
1185
+ ts = torch.full((b,), i, device=device, dtype=torch.long)
1186
+ if self.shorten_cond_schedule:
1187
+ assert self.model.conditioning_key != 'hybrid'
1188
+ tc = self.cond_ids[ts].to(cond.device)
1189
+ cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
1190
+
1191
+ img = self.p_sample(img, cond, ts,
1192
+ clip_denoised=self.clip_denoised,
1193
+ quantize_denoised=quantize_denoised)
1194
+ if mask is not None:
1195
+ img_orig = self.q_sample(x0, ts)
1196
+ img = img_orig * mask + (1. - mask) * img
1197
+
1198
+ if i % log_every_t == 0 or i == timesteps - 1:
1199
+ intermediates.append(img)
1200
+ if callback:
1201
+ callback(i)
1202
+ if img_callback:
1203
+ img_callback(img, i)
1204
+
1205
+ if return_intermediates:
1206
+ return img, intermediates
1207
+ return img
1208
+
1209
+ @torch.no_grad()
1210
+ def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None,
1211
+ verbose=True, timesteps=None, quantize_denoised=False,
1212
+ mask=None, x0=None, shape=None,**kwargs):
1213
+ if shape is None:
1214
+ shape = (batch_size, self.channels, self.image_size, self.image_size)
1215
+ if cond is not None:
1216
+ if isinstance(cond, dict):
1217
+ cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
1218
+ [x[:batch_size] for x in cond[key]] for key in cond}
1219
+ else:
1220
+ cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
1221
+ return self.p_sample_loop(cond,
1222
+ shape,
1223
+ return_intermediates=return_intermediates, x_T=x_T,
1224
+ verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised,
1225
+ mask=mask, x0=x0)
1226
+
1227
+ @torch.no_grad()
1228
+ def sample_log(self,cond,batch_size,ddim, ddim_steps,**kwargs):
1229
+
1230
+ if ddim:
1231
+ ddim_sampler = DDIMSampler(self)
1232
+ shape = (self.channels, self.image_size, self.image_size)
1233
+ samples, intermediates =ddim_sampler.sample(ddim_steps,batch_size,
1234
+ shape,cond,verbose=False,**kwargs)
1235
+
1236
+ else:
1237
+ samples, intermediates = self.sample(cond=cond, batch_size=batch_size,
1238
+ return_intermediates=True,**kwargs)
1239
+
1240
+ return samples, intermediates
1241
+
1242
+
1243
+ @torch.no_grad()
1244
+ def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
1245
+ quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
1246
+ plot_diffusion_rows=True, **kwargs):
1247
+
1248
+ use_ddim = ddim_steps is not None
1249
+
1250
+ log = {}
1251
+ z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key,
1252
+ return_first_stage_outputs=True,
1253
+ force_c_encode=True,
1254
+ return_original_cond=True,
1255
+ bs=N)
1256
+ N = min(x.shape[0], N)
1257
+ n_row = min(x.shape[0], n_row)
1258
+ log["inputs"] = x
1259
+ log["reconstruction"] = xrec
1260
+ if self.model.conditioning_key is not None:
1261
+ if hasattr(self.cond_stage_model, "decode"):
1262
+ xc = self.cond_stage_model.decode(c)
1263
+ log["conditioning"] = xc
1264
+ elif self.cond_stage_key in ["caption"]:
1265
+ xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["caption"])
1266
+ log["conditioning"] = xc
1267
+ elif self.cond_stage_key == 'class_label':
1268
+ xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
1269
+ log['conditioning'] = xc
1270
+ elif isimage(xc):
1271
+ log["conditioning"] = xc
1272
+ if ismap(xc):
1273
+ log["original_conditioning"] = self.to_rgb(xc)
1274
+
1275
+ if plot_diffusion_rows:
1276
+ # get diffusion row
1277
+ diffusion_row = []
1278
+ z_start = z[:n_row]
1279
+ for t in range(self.num_timesteps):
1280
+ if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
1281
+ t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
1282
+ t = t.to(self.device).long()
1283
+ noise = torch.randn_like(z_start)
1284
+ z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
1285
+ diffusion_row.append(self.decode_first_stage(z_noisy))
1286
+
1287
+ diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W
1288
+ diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
1289
+ diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
1290
+ diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
1291
+ log["diffusion_row"] = diffusion_grid
1292
+
1293
+ if sample:
1294
+ # get denoise row
1295
+ with self.ema_scope("Plotting"):
1296
+ samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
1297
+ ddim_steps=ddim_steps,eta=ddim_eta)
1298
+ # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
1299
+ x_samples = self.decode_first_stage(samples)
1300
+ log["samples"] = x_samples
1301
+ if plot_denoise_rows:
1302
+ denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
1303
+ log["denoise_row"] = denoise_grid
1304
+
1305
+ if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance(
1306
+ self.first_stage_model, IdentityFirstStage):
1307
+ # also display when quantizing x0 while sampling
1308
+ with self.ema_scope("Plotting Quantized Denoised"):
1309
+ samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
1310
+ ddim_steps=ddim_steps,eta=ddim_eta,
1311
+ quantize_denoised=True)
1312
+ # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True,
1313
+ # quantize_denoised=True)
1314
+ x_samples = self.decode_first_stage(samples.to(self.device))
1315
+ log["samples_x0_quantized"] = x_samples
1316
+
1317
+ if inpaint:
1318
+ # make a simple center square
1319
+ h, w = z.shape[2], z.shape[3]
1320
+ mask = torch.ones(N, h, w).to(self.device)
1321
+ # zeros will be filled in
1322
+ mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0.
1323
+ mask = mask[:, None, ...]
1324
+ with self.ema_scope("Plotting Inpaint"):
1325
+
1326
+ samples, _ = self.sample_log(cond=c,batch_size=N,ddim=use_ddim, eta=ddim_eta,
1327
+ ddim_steps=ddim_steps, x0=z[:N], mask=mask)
1328
+ x_samples = self.decode_first_stage(samples.to(self.device))
1329
+ log["samples_inpainting"] = x_samples
1330
+ log["mask"] = mask
1331
+
1332
+ # outpaint
1333
+ with self.ema_scope("Plotting Outpaint"):
1334
+ samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,eta=ddim_eta,
1335
+ ddim_steps=ddim_steps, x0=z[:N], mask=mask)
1336
+ x_samples = self.decode_first_stage(samples.to(self.device))
1337
+ log["samples_outpainting"] = x_samples
1338
+
1339
+ if plot_progressive_rows:
1340
+ with self.ema_scope("Plotting Progressives"):
1341
+ img, progressives = self.progressive_denoising(c,
1342
+ shape=(self.channels, self.image_size, self.image_size),
1343
+ batch_size=N)
1344
+ prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
1345
+ log["progressive_row"] = prog_row
1346
+
1347
+ if return_keys:
1348
+ if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
1349
+ return log
1350
+ else:
1351
+ return {key: log[key] for key in return_keys}
1352
+ return log
1353
+
1354
+ def configure_optimizers(self):
1355
+ lr = self.learning_rate
1356
+ params = list(self.model.parameters())
1357
+ if self.cond_stage_trainable:
1358
+ print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
1359
+ params = params + list(self.cond_stage_model.parameters())
1360
+ if self.learn_logvar:
1361
+ print('Diffusion model optimizing logvar')
1362
+ params.append(self.logvar)
1363
+ opt = torch.optim.AdamW(params, lr=lr)
1364
+ if self.use_scheduler:
1365
+ assert 'target' in self.scheduler_config
1366
+ scheduler = instantiate_from_config(self.scheduler_config)
1367
+
1368
+ print("Setting up LambdaLR scheduler...")
1369
+ scheduler = [
1370
+ {
1371
+ 'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule),
1372
+ 'interval': 'step',
1373
+ 'frequency': 1
1374
+ }]
1375
+ return [opt], scheduler
1376
+ return opt
1377
+
1378
+ @torch.no_grad()
1379
+ def to_rgb(self, x):
1380
+ x = x.float()
1381
+ if not hasattr(self, "colorize"):
1382
+ self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x)
1383
+ x = nn.functional.conv2d(x, weight=self.colorize)
1384
+ x = 2. * (x - x.min()) / (x.max() - x.min()) - 1.
1385
+ return x
1386
+
1387
+
1388
+ class DiffusionWrapperV1(pl.LightningModule):
1389
+ def __init__(self, diff_model_config, conditioning_key):
1390
+ super().__init__()
1391
+ self.diffusion_model = instantiate_from_config(diff_model_config)
1392
+ self.conditioning_key = conditioning_key
1393
+ assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm']
1394
+
1395
+ def forward(self, x, t, c_concat: list = None, c_crossattn: list = None):
1396
+ if self.conditioning_key is None:
1397
+ out = self.diffusion_model(x, t)
1398
+ elif self.conditioning_key == 'concat':
1399
+ xc = torch.cat([x] + c_concat, dim=1)
1400
+ out = self.diffusion_model(xc, t)
1401
+ elif self.conditioning_key == 'crossattn':
1402
+ cc = torch.cat(c_crossattn, 1)
1403
+ out = self.diffusion_model(x, t, context=cc)
1404
+ elif self.conditioning_key == 'hybrid':
1405
+ xc = torch.cat([x] + c_concat, dim=1)
1406
+ cc = torch.cat(c_crossattn, 1)
1407
+ out = self.diffusion_model(xc, t, context=cc)
1408
+ elif self.conditioning_key == 'adm':
1409
+ cc = c_crossattn[0]
1410
+ out = self.diffusion_model(x, t, y=cc)
1411
+ else:
1412
+ raise NotImplementedError()
1413
+
1414
+ return out
1415
+
1416
+
1417
+ class Layout2ImgDiffusionV1(LatentDiffusionV1):
1418
+ # TODO: move all layout-specific hacks to this class
1419
+ def __init__(self, cond_stage_key, *args, **kwargs):
1420
+ assert cond_stage_key == 'coordinates_bbox', 'Layout2ImgDiffusion only for cond_stage_key="coordinates_bbox"'
1421
+ super().__init__(*args, cond_stage_key=cond_stage_key, **kwargs)
1422
+
1423
+ def log_images(self, batch, N=8, *args, **kwargs):
1424
+ logs = super().log_images(*args, batch=batch, N=N, **kwargs)
1425
+
1426
+ key = 'train' if self.training else 'validation'
1427
+ dset = self.trainer.datamodule.datasets[key]
1428
+ mapper = dset.conditional_builders[self.cond_stage_key]
1429
+
1430
+ bbox_imgs = []
1431
+ map_fn = lambda catno: dset.get_textual_label(dset.get_category_id(catno))
1432
+ for tknzd_bbox in batch[self.cond_stage_key][:N]:
1433
+ bboximg = mapper.plot(tknzd_bbox.detach().cpu(), map_fn, (256, 256))
1434
+ bbox_imgs.append(bboximg)
1435
+
1436
+ cond_img = torch.stack(bbox_imgs, dim=0)
1437
+ logs['bbox_image'] = cond_img
1438
+ return logs
1439
+
1440
+ ldm.models.diffusion.ddpm.DDPMV1 = DDPMV1
1441
+ ldm.models.diffusion.ddpm.LatentDiffusionV1 = LatentDiffusionV1
1442
+ ldm.models.diffusion.ddpm.DiffusionWrapperV1 = DiffusionWrapperV1
1443
+ ldm.models.diffusion.ddpm.Layout2ImgDiffusionV1 = Layout2ImgDiffusionV1
extensions-builtin/LDSR/vqvae_quantize.py ADDED
@@ -0,0 +1,147 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Vendored from https://raw.githubusercontent.com/CompVis/taming-transformers/24268930bf1dce879235a7fddd0b2355b84d7ea6/taming/modules/vqvae/quantize.py,
2
+ # where the license is as follows:
3
+ #
4
+ # Copyright (c) 2020 Patrick Esser and Robin Rombach and Björn Ommer
5
+ #
6
+ # Permission is hereby granted, free of charge, to any person obtaining a copy
7
+ # of this software and associated documentation files (the "Software"), to deal
8
+ # in the Software without restriction, including without limitation the rights
9
+ # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10
+ # copies of the Software, and to permit persons to whom the Software is
11
+ # furnished to do so, subject to the following conditions:
12
+ #
13
+ # The above copyright notice and this permission notice shall be included in all
14
+ # copies or substantial portions of the Software.
15
+ #
16
+ # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
17
+ # EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
18
+ # MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
19
+ # IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
20
+ # DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21
+ # OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE
22
+ # OR OTHER DEALINGS IN THE SOFTWARE./
23
+
24
+ import torch
25
+ import torch.nn as nn
26
+ import numpy as np
27
+ from einops import rearrange
28
+
29
+
30
+ class VectorQuantizer2(nn.Module):
31
+ """
32
+ Improved version over VectorQuantizer, can be used as a drop-in replacement. Mostly
33
+ avoids costly matrix multiplications and allows for post-hoc remapping of indices.
34
+ """
35
+
36
+ # NOTE: due to a bug the beta term was applied to the wrong term. for
37
+ # backwards compatibility we use the buggy version by default, but you can
38
+ # specify legacy=False to fix it.
39
+ def __init__(self, n_e, e_dim, beta, remap=None, unknown_index="random",
40
+ sane_index_shape=False, legacy=True):
41
+ super().__init__()
42
+ self.n_e = n_e
43
+ self.e_dim = e_dim
44
+ self.beta = beta
45
+ self.legacy = legacy
46
+
47
+ self.embedding = nn.Embedding(self.n_e, self.e_dim)
48
+ self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)
49
+
50
+ self.remap = remap
51
+ if self.remap is not None:
52
+ self.register_buffer("used", torch.tensor(np.load(self.remap)))
53
+ self.re_embed = self.used.shape[0]
54
+ self.unknown_index = unknown_index # "random" or "extra" or integer
55
+ if self.unknown_index == "extra":
56
+ self.unknown_index = self.re_embed
57
+ self.re_embed = self.re_embed + 1
58
+ print(f"Remapping {self.n_e} indices to {self.re_embed} indices. "
59
+ f"Using {self.unknown_index} for unknown indices.")
60
+ else:
61
+ self.re_embed = n_e
62
+
63
+ self.sane_index_shape = sane_index_shape
64
+
65
+ def remap_to_used(self, inds):
66
+ ishape = inds.shape
67
+ assert len(ishape) > 1
68
+ inds = inds.reshape(ishape[0], -1)
69
+ used = self.used.to(inds)
70
+ match = (inds[:, :, None] == used[None, None, ...]).long()
71
+ new = match.argmax(-1)
72
+ unknown = match.sum(2) < 1
73
+ if self.unknown_index == "random":
74
+ new[unknown] = torch.randint(0, self.re_embed, size=new[unknown].shape).to(device=new.device)
75
+ else:
76
+ new[unknown] = self.unknown_index
77
+ return new.reshape(ishape)
78
+
79
+ def unmap_to_all(self, inds):
80
+ ishape = inds.shape
81
+ assert len(ishape) > 1
82
+ inds = inds.reshape(ishape[0], -1)
83
+ used = self.used.to(inds)
84
+ if self.re_embed > self.used.shape[0]: # extra token
85
+ inds[inds >= self.used.shape[0]] = 0 # simply set to zero
86
+ back = torch.gather(used[None, :][inds.shape[0] * [0], :], 1, inds)
87
+ return back.reshape(ishape)
88
+
89
+ def forward(self, z, temp=None, rescale_logits=False, return_logits=False):
90
+ assert temp is None or temp == 1.0, "Only for interface compatible with Gumbel"
91
+ assert rescale_logits is False, "Only for interface compatible with Gumbel"
92
+ assert return_logits is False, "Only for interface compatible with Gumbel"
93
+ # reshape z -> (batch, height, width, channel) and flatten
94
+ z = rearrange(z, 'b c h w -> b h w c').contiguous()
95
+ z_flattened = z.view(-1, self.e_dim)
96
+ # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
97
+
98
+ d = torch.sum(z_flattened ** 2, dim=1, keepdim=True) + \
99
+ torch.sum(self.embedding.weight ** 2, dim=1) - 2 * \
100
+ torch.einsum('bd,dn->bn', z_flattened, rearrange(self.embedding.weight, 'n d -> d n'))
101
+
102
+ min_encoding_indices = torch.argmin(d, dim=1)
103
+ z_q = self.embedding(min_encoding_indices).view(z.shape)
104
+ perplexity = None
105
+ min_encodings = None
106
+
107
+ # compute loss for embedding
108
+ if not self.legacy:
109
+ loss = self.beta * torch.mean((z_q.detach() - z) ** 2) + \
110
+ torch.mean((z_q - z.detach()) ** 2)
111
+ else:
112
+ loss = torch.mean((z_q.detach() - z) ** 2) + self.beta * \
113
+ torch.mean((z_q - z.detach()) ** 2)
114
+
115
+ # preserve gradients
116
+ z_q = z + (z_q - z).detach()
117
+
118
+ # reshape back to match original input shape
119
+ z_q = rearrange(z_q, 'b h w c -> b c h w').contiguous()
120
+
121
+ if self.remap is not None:
122
+ min_encoding_indices = min_encoding_indices.reshape(z.shape[0], -1) # add batch axis
123
+ min_encoding_indices = self.remap_to_used(min_encoding_indices)
124
+ min_encoding_indices = min_encoding_indices.reshape(-1, 1) # flatten
125
+
126
+ if self.sane_index_shape:
127
+ min_encoding_indices = min_encoding_indices.reshape(
128
+ z_q.shape[0], z_q.shape[2], z_q.shape[3])
129
+
130
+ return z_q, loss, (perplexity, min_encodings, min_encoding_indices)
131
+
132
+ def get_codebook_entry(self, indices, shape):
133
+ # shape specifying (batch, height, width, channel)
134
+ if self.remap is not None:
135
+ indices = indices.reshape(shape[0], -1) # add batch axis
136
+ indices = self.unmap_to_all(indices)
137
+ indices = indices.reshape(-1) # flatten again
138
+
139
+ # get quantized latent vectors
140
+ z_q = self.embedding(indices)
141
+
142
+ if shape is not None:
143
+ z_q = z_q.view(shape)
144
+ # reshape back to match original input shape
145
+ z_q = z_q.permute(0, 3, 1, 2).contiguous()
146
+
147
+ return z_q
extensions-builtin/Lora/extra_networks_lora.py ADDED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from modules import extra_networks, shared
2
+ import networks
3
+
4
+
5
+ class ExtraNetworkLora(extra_networks.ExtraNetwork):
6
+ def __init__(self):
7
+ super().__init__('lora')
8
+
9
+ self.errors = {}
10
+ """mapping of network names to the number of errors the network had during operation"""
11
+
12
+ def activate(self, p, params_list):
13
+ additional = shared.opts.sd_lora
14
+
15
+ self.errors.clear()
16
+
17
+ if additional != "None" and additional in networks.available_networks and not any(x for x in params_list if x.items[0] == additional):
18
+ p.all_prompts = [x + f"<lora:{additional}:{shared.opts.extra_networks_default_multiplier}>" for x in p.all_prompts]
19
+ params_list.append(extra_networks.ExtraNetworkParams(items=[additional, shared.opts.extra_networks_default_multiplier]))
20
+
21
+ names = []
22
+ te_multipliers = []
23
+ unet_multipliers = []
24
+ dyn_dims = []
25
+ for params in params_list:
26
+ assert params.items
27
+
28
+ names.append(params.positional[0])
29
+
30
+ te_multiplier = float(params.positional[1]) if len(params.positional) > 1 else 1.0
31
+ te_multiplier = float(params.named.get("te", te_multiplier))
32
+
33
+ unet_multiplier = float(params.positional[2]) if len(params.positional) > 2 else te_multiplier
34
+ unet_multiplier = float(params.named.get("unet", unet_multiplier))
35
+
36
+ dyn_dim = int(params.positional[3]) if len(params.positional) > 3 else None
37
+ dyn_dim = int(params.named["dyn"]) if "dyn" in params.named else dyn_dim
38
+
39
+ te_multipliers.append(te_multiplier)
40
+ unet_multipliers.append(unet_multiplier)
41
+ dyn_dims.append(dyn_dim)
42
+
43
+ networks.load_networks(names, te_multipliers, unet_multipliers, dyn_dims)
44
+
45
+ if shared.opts.lora_add_hashes_to_infotext:
46
+ network_hashes = []
47
+ for item in networks.loaded_networks:
48
+ shorthash = item.network_on_disk.shorthash
49
+ if not shorthash:
50
+ continue
51
+
52
+ alias = item.mentioned_name
53
+ if not alias:
54
+ continue
55
+
56
+ alias = alias.replace(":", "").replace(",", "")
57
+
58
+ network_hashes.append(f"{alias}: {shorthash}")
59
+
60
+ if network_hashes:
61
+ p.extra_generation_params["Lora hashes"] = ", ".join(network_hashes)
62
+
63
+ def deactivate(self, p):
64
+ if self.errors:
65
+ p.comment("Networks with errors: " + ", ".join(f"{k} ({v})" for k, v in self.errors.items()))
66
+
67
+ self.errors.clear()
extensions-builtin/Lora/lora.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import networks
2
+
3
+ list_available_loras = networks.list_available_networks
4
+
5
+ available_loras = networks.available_networks
6
+ available_lora_aliases = networks.available_network_aliases
7
+ available_lora_hash_lookup = networks.available_network_hash_lookup
8
+ forbidden_lora_aliases = networks.forbidden_network_aliases
9
+ loaded_loras = networks.loaded_networks
extensions-builtin/Lora/lora_logger.py ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import sys
2
+ import copy
3
+ import logging
4
+
5
+
6
+ class ColoredFormatter(logging.Formatter):
7
+ COLORS = {
8
+ "DEBUG": "\033[0;36m", # CYAN
9
+ "INFO": "\033[0;32m", # GREEN
10
+ "WARNING": "\033[0;33m", # YELLOW
11
+ "ERROR": "\033[0;31m", # RED
12
+ "CRITICAL": "\033[0;37;41m", # WHITE ON RED
13
+ "RESET": "\033[0m", # RESET COLOR
14
+ }
15
+
16
+ def format(self, record):
17
+ colored_record = copy.copy(record)
18
+ levelname = colored_record.levelname
19
+ seq = self.COLORS.get(levelname, self.COLORS["RESET"])
20
+ colored_record.levelname = f"{seq}{levelname}{self.COLORS['RESET']}"
21
+ return super().format(colored_record)
22
+
23
+
24
+ logger = logging.getLogger("lora")
25
+ logger.propagate = False
26
+
27
+
28
+ if not logger.handlers:
29
+ handler = logging.StreamHandler(sys.stdout)
30
+ handler.setFormatter(
31
+ ColoredFormatter("[%(name)s]-%(levelname)s: %(message)s")
32
+ )
33
+ logger.addHandler(handler)
extensions-builtin/Lora/lora_patches.py ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ class LoraPatches:
2
+ def __init__(self):
3
+ pass
4
+
5
+ def undo(self):
6
+ pass
extensions-builtin/Lora/network.py ADDED
@@ -0,0 +1,190 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from __future__ import annotations
2
+ import os
3
+ from collections import namedtuple
4
+ import enum
5
+
6
+ import torch.nn as nn
7
+ import torch.nn.functional as F
8
+
9
+ from modules import sd_models, cache, errors, hashes, shared
10
+
11
+ NetworkWeights = namedtuple('NetworkWeights', ['network_key', 'sd_key', 'w', 'sd_module'])
12
+
13
+ metadata_tags_order = {"ss_sd_model_name": 1, "ss_resolution": 2, "ss_clip_skip": 3, "ss_num_train_images": 10, "ss_tag_frequency": 20}
14
+
15
+
16
+ class SdVersion(enum.Enum):
17
+ Unknown = 1
18
+ SD1 = 2
19
+ SD2 = 3
20
+ SDXL = 4
21
+
22
+
23
+ class NetworkOnDisk:
24
+ def __init__(self, name, filename):
25
+ self.name = name
26
+ self.filename = filename
27
+ self.metadata = {}
28
+ self.is_safetensors = os.path.splitext(filename)[1].lower() == ".safetensors"
29
+
30
+ def read_metadata():
31
+ metadata = sd_models.read_metadata_from_safetensors(filename)
32
+ metadata.pop('ssmd_cover_images', None) # those are cover images, and they are too big to display in UI as text
33
+
34
+ return metadata
35
+
36
+ if self.is_safetensors:
37
+ try:
38
+ self.metadata = cache.cached_data_for_file('safetensors-metadata', "lora/" + self.name, filename, read_metadata)
39
+ except Exception as e:
40
+ errors.display(e, f"reading lora {filename}")
41
+
42
+ if self.metadata:
43
+ m = {}
44
+ for k, v in sorted(self.metadata.items(), key=lambda x: metadata_tags_order.get(x[0], 999)):
45
+ m[k] = v
46
+
47
+ self.metadata = m
48
+
49
+ self.alias = self.metadata.get('ss_output_name', self.name)
50
+
51
+ self.hash = None
52
+ self.shorthash = None
53
+ self.set_hash(
54
+ self.metadata.get('sshs_model_hash') or
55
+ hashes.sha256_from_cache(self.filename, "lora/" + self.name, use_addnet_hash=self.is_safetensors) or
56
+ ''
57
+ )
58
+
59
+ self.sd_version = self.detect_version()
60
+
61
+ def detect_version(self):
62
+ if str(self.metadata.get('ss_base_model_version', "")).startswith("sdxl_"):
63
+ return SdVersion.SDXL
64
+ elif str(self.metadata.get('ss_v2', "")) == "True":
65
+ return SdVersion.SD2
66
+ elif len(self.metadata):
67
+ return SdVersion.SD1
68
+
69
+ return SdVersion.Unknown
70
+
71
+ def set_hash(self, v):
72
+ self.hash = v
73
+ self.shorthash = self.hash[0:12]
74
+
75
+ if self.shorthash:
76
+ import networks
77
+ networks.available_network_hash_lookup[self.shorthash] = self
78
+
79
+ def read_hash(self):
80
+ if not self.hash:
81
+ self.set_hash(hashes.sha256(self.filename, "lora/" + self.name, use_addnet_hash=self.is_safetensors) or '')
82
+
83
+ def get_alias(self):
84
+ import networks
85
+ if shared.opts.lora_preferred_name == "Filename" or self.alias.lower() in networks.forbidden_network_aliases:
86
+ return self.name
87
+ else:
88
+ return self.alias
89
+
90
+
91
+ class Network: # LoraModule
92
+ def __init__(self, name, network_on_disk: NetworkOnDisk):
93
+ self.name = name
94
+ self.network_on_disk = network_on_disk
95
+ self.te_multiplier = 1.0
96
+ self.unet_multiplier = 1.0
97
+ self.dyn_dim = None
98
+ self.modules = {}
99
+ self.bundle_embeddings = {}
100
+ self.mtime = None
101
+
102
+ self.mentioned_name = None
103
+ """the text that was used to add the network to prompt - can be either name or an alias"""
104
+
105
+
106
+ class ModuleType:
107
+ def create_module(self, net: Network, weights: NetworkWeights) -> Network | None:
108
+ return None
109
+
110
+
111
+ class NetworkModule:
112
+ def __init__(self, net: Network, weights: NetworkWeights):
113
+ self.network = net
114
+ self.network_key = weights.network_key
115
+ self.sd_key = weights.sd_key
116
+ self.sd_module = weights.sd_module
117
+
118
+ if hasattr(self.sd_module, 'weight'):
119
+ self.shape = self.sd_module.weight.shape
120
+
121
+ self.ops = None
122
+ self.extra_kwargs = {}
123
+ if isinstance(self.sd_module, nn.Conv2d):
124
+ self.ops = F.conv2d
125
+ self.extra_kwargs = {
126
+ 'stride': self.sd_module.stride,
127
+ 'padding': self.sd_module.padding
128
+ }
129
+ elif isinstance(self.sd_module, nn.Linear):
130
+ self.ops = F.linear
131
+ elif isinstance(self.sd_module, nn.LayerNorm):
132
+ self.ops = F.layer_norm
133
+ self.extra_kwargs = {
134
+ 'normalized_shape': self.sd_module.normalized_shape,
135
+ 'eps': self.sd_module.eps
136
+ }
137
+ elif isinstance(self.sd_module, nn.GroupNorm):
138
+ self.ops = F.group_norm
139
+ self.extra_kwargs = {
140
+ 'num_groups': self.sd_module.num_groups,
141
+ 'eps': self.sd_module.eps
142
+ }
143
+
144
+ self.dim = None
145
+ self.bias = weights.w.get("bias")
146
+ self.alpha = weights.w["alpha"].item() if "alpha" in weights.w else None
147
+ self.scale = weights.w["scale"].item() if "scale" in weights.w else None
148
+
149
+ def multiplier(self):
150
+ if 'transformer' in self.sd_key[:20]:
151
+ return self.network.te_multiplier
152
+ else:
153
+ return self.network.unet_multiplier
154
+
155
+ def calc_scale(self):
156
+ if self.scale is not None:
157
+ return self.scale
158
+ if self.dim is not None and self.alpha is not None:
159
+ return self.alpha / self.dim
160
+
161
+ return 1.0
162
+
163
+ def finalize_updown(self, updown, orig_weight, output_shape, ex_bias=None):
164
+ if self.bias is not None:
165
+ updown = updown.reshape(self.bias.shape)
166
+ updown += self.bias.to(orig_weight.device, dtype=updown.dtype)
167
+ updown = updown.reshape(output_shape)
168
+
169
+ if len(output_shape) == 4:
170
+ updown = updown.reshape(output_shape)
171
+
172
+ if orig_weight.size().numel() == updown.size().numel():
173
+ updown = updown.reshape(orig_weight.shape)
174
+
175
+ if ex_bias is not None:
176
+ ex_bias = ex_bias * self.multiplier()
177
+
178
+ return updown * self.calc_scale() * self.multiplier(), ex_bias
179
+
180
+ def calc_updown(self, target):
181
+ raise NotImplementedError()
182
+
183
+ def forward(self, x, y):
184
+ """A general forward implementation for all modules"""
185
+ if self.ops is None:
186
+ raise NotImplementedError()
187
+ else:
188
+ updown, ex_bias = self.calc_updown(self.sd_module.weight)
189
+ return y + self.ops(x, weight=updown, bias=ex_bias, **self.extra_kwargs)
190
+
extensions-builtin/Lora/networks.py ADDED
@@ -0,0 +1,219 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import re
3
+
4
+ import lora_patches
5
+ import functools
6
+ import network
7
+
8
+ import torch
9
+ from typing import Union
10
+
11
+ from modules import shared, sd_models, errors, scripts
12
+ from ldm_patched.modules.utils import load_torch_file
13
+ from ldm_patched.modules.sd import load_lora_for_models
14
+
15
+
16
+ @functools.lru_cache(maxsize=5)
17
+ def load_lora_state_dict(filename):
18
+ return load_torch_file(filename, safe_load=True)
19
+
20
+
21
+ def convert_diffusers_name_to_compvis(key, is_sd2):
22
+ pass
23
+
24
+
25
+ def assign_network_names_to_compvis_modules(sd_model):
26
+ pass
27
+
28
+
29
+ def load_network(name, network_on_disk):
30
+ net = network.Network(name, network_on_disk)
31
+ net.mtime = os.path.getmtime(network_on_disk.filename)
32
+
33
+ return net
34
+
35
+
36
+ def purge_networks_from_memory():
37
+ pass
38
+
39
+
40
+ def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=None):
41
+ global lora_state_dict_cache
42
+
43
+ current_sd = sd_models.model_data.get_sd_model()
44
+ if current_sd is None:
45
+ return
46
+
47
+ loaded_networks.clear()
48
+
49
+ networks_on_disk = [available_networks.get(name, None) if name.lower() in forbidden_network_aliases else available_network_aliases.get(name, None) for name in names]
50
+ if any(x is None for x in networks_on_disk):
51
+ list_available_networks()
52
+ networks_on_disk = [available_networks.get(name, None) if name.lower() in forbidden_network_aliases else available_network_aliases.get(name, None) for name in names]
53
+
54
+ for i, (network_on_disk, name) in enumerate(zip(networks_on_disk, names)):
55
+ try:
56
+ net = load_network(name, network_on_disk)
57
+ except Exception as e:
58
+ errors.display(e, f"loading network {network_on_disk.filename}")
59
+ continue
60
+ net.mentioned_name = name
61
+ network_on_disk.read_hash()
62
+ loaded_networks.append(net)
63
+
64
+ compiled_lora_targets = []
65
+ for a, b, c in zip(networks_on_disk, unet_multipliers, te_multipliers):
66
+ compiled_lora_targets.append([a.filename, b, c])
67
+
68
+ compiled_lora_targets_hash = str(compiled_lora_targets)
69
+
70
+ if current_sd.current_lora_hash == compiled_lora_targets_hash:
71
+ return
72
+
73
+ current_sd.current_lora_hash = compiled_lora_targets_hash
74
+ current_sd.forge_objects.unet = current_sd.forge_objects_original.unet
75
+ current_sd.forge_objects.clip = current_sd.forge_objects_original.clip
76
+
77
+ for filename, strength_model, strength_clip in compiled_lora_targets:
78
+ lora_sd = load_lora_state_dict(filename)
79
+ current_sd.forge_objects.unet, current_sd.forge_objects.clip = load_lora_for_models(
80
+ current_sd.forge_objects.unet, current_sd.forge_objects.clip, lora_sd, strength_model, strength_clip,
81
+ filename=filename)
82
+
83
+ current_sd.forge_objects_after_applying_lora = current_sd.forge_objects.shallow_copy()
84
+ return
85
+
86
+
87
+ def network_restore_weights_from_backup(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn.GroupNorm, torch.nn.LayerNorm, torch.nn.MultiheadAttention]):
88
+ pass
89
+
90
+
91
+ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn.GroupNorm, torch.nn.LayerNorm, torch.nn.MultiheadAttention]):
92
+ pass
93
+
94
+
95
+ def network_forward(org_module, input, original_forward):
96
+ pass
97
+
98
+
99
+ def network_reset_cached_weight(self: Union[torch.nn.Conv2d, torch.nn.Linear]):
100
+ pass
101
+
102
+
103
+ def network_Linear_forward(self, input):
104
+ pass
105
+
106
+
107
+ def network_Linear_load_state_dict(self, *args, **kwargs):
108
+ pass
109
+
110
+
111
+ def network_Conv2d_forward(self, input):
112
+ pass
113
+
114
+
115
+ def network_Conv2d_load_state_dict(self, *args, **kwargs):
116
+ pass
117
+
118
+
119
+ def network_GroupNorm_forward(self, input):
120
+ pass
121
+
122
+
123
+ def network_GroupNorm_load_state_dict(self, *args, **kwargs):
124
+ pass
125
+
126
+
127
+ def network_LayerNorm_forward(self, input):
128
+ pass
129
+
130
+
131
+ def network_LayerNorm_load_state_dict(self, *args, **kwargs):
132
+ pass
133
+
134
+
135
+ def network_MultiheadAttention_forward(self, *args, **kwargs):
136
+ pass
137
+
138
+
139
+ def network_MultiheadAttention_load_state_dict(self, *args, **kwargs):
140
+ pass
141
+
142
+
143
+ def list_available_networks():
144
+ available_networks.clear()
145
+ available_network_aliases.clear()
146
+ forbidden_network_aliases.clear()
147
+ available_network_hash_lookup.clear()
148
+ forbidden_network_aliases.update({"none": 1, "Addams": 1})
149
+
150
+ os.makedirs(shared.cmd_opts.lora_dir, exist_ok=True)
151
+
152
+ candidates = list(shared.walk_files(shared.cmd_opts.lora_dir, allowed_extensions=[".pt", ".ckpt", ".safetensors"]))
153
+ for filename in candidates:
154
+ if os.path.isdir(filename):
155
+ continue
156
+
157
+ name = os.path.splitext(os.path.basename(filename))[0]
158
+ try:
159
+ entry = network.NetworkOnDisk(name, filename)
160
+ except OSError: # should catch FileNotFoundError and PermissionError etc.
161
+ errors.report(f"Failed to load network {name} from {filename}", exc_info=True)
162
+ continue
163
+
164
+ available_networks[name] = entry
165
+
166
+ if entry.alias in available_network_aliases:
167
+ forbidden_network_aliases[entry.alias.lower()] = 1
168
+
169
+ available_network_aliases[name] = entry
170
+ available_network_aliases[entry.alias] = entry
171
+
172
+
173
+ re_network_name = re.compile(r"(.*)\s*\([0-9a-fA-F]+\)")
174
+
175
+
176
+ def infotext_pasted(infotext, params):
177
+ if "AddNet Module 1" in [x[1] for x in scripts.scripts_txt2img.infotext_fields]:
178
+ return # if the other extension is active, it will handle those fields, no need to do anything
179
+
180
+ added = []
181
+
182
+ for k in params:
183
+ if not k.startswith("AddNet Model "):
184
+ continue
185
+
186
+ num = k[13:]
187
+
188
+ if params.get("AddNet Module " + num) != "LoRA":
189
+ continue
190
+
191
+ name = params.get("AddNet Model " + num)
192
+ if name is None:
193
+ continue
194
+
195
+ m = re_network_name.match(name)
196
+ if m:
197
+ name = m.group(1)
198
+
199
+ multiplier = params.get("AddNet Weight A " + num, "1.0")
200
+
201
+ added.append(f"<lora:{name}:{multiplier}>")
202
+
203
+ if added:
204
+ params["Prompt"] += "\n" + "".join(added)
205
+
206
+
207
+ originals: lora_patches.LoraPatches = None
208
+
209
+ extra_network_lora = None
210
+
211
+ available_networks = {}
212
+ available_network_aliases = {}
213
+ loaded_networks = []
214
+ loaded_bundle_embeddings = {}
215
+ networks_in_memory = {}
216
+ available_network_hash_lookup = {}
217
+ forbidden_network_aliases = {}
218
+
219
+ list_available_networks()
extensions-builtin/Lora/preload.py ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ from modules import paths
3
+ from modules.paths_internal import normalized_filepath
4
+
5
+
6
+ def preload(parser):
7
+ parser.add_argument("--lora-dir", type=normalized_filepath, help="Path to directory with Lora networks.", default=os.path.join(paths.models_path, 'Lora'))
8
+ parser.add_argument("--lyco-dir-backcompat", type=normalized_filepath, help="Path to directory with LyCORIS networks (for backawards compatibility; can also use --lyco-dir).", default=os.path.join(paths.models_path, 'LyCORIS'))
extensions-builtin/Lora/scripts/lora_script.py ADDED
@@ -0,0 +1,100 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+
3
+ import gradio as gr
4
+ from fastapi import FastAPI
5
+
6
+ import network
7
+ import networks
8
+ import lora # noqa:F401
9
+ import lora_patches
10
+ import extra_networks_lora
11
+ import ui_extra_networks_lora
12
+ from modules import script_callbacks, ui_extra_networks, extra_networks, shared
13
+
14
+
15
+ def unload():
16
+ networks.originals.undo()
17
+
18
+
19
+ def before_ui():
20
+ ui_extra_networks.register_page(ui_extra_networks_lora.ExtraNetworksPageLora())
21
+
22
+ networks.extra_network_lora = extra_networks_lora.ExtraNetworkLora()
23
+ extra_networks.register_extra_network(networks.extra_network_lora)
24
+
25
+
26
+ networks.originals = lora_patches.LoraPatches()
27
+
28
+ script_callbacks.on_model_loaded(networks.assign_network_names_to_compvis_modules)
29
+ script_callbacks.on_script_unloaded(unload)
30
+ script_callbacks.on_before_ui(before_ui)
31
+ script_callbacks.on_infotext_pasted(networks.infotext_pasted)
32
+
33
+
34
+ shared.options_templates.update(shared.options_section(('extra_networks', "Extra Networks"), {
35
+ "sd_lora": shared.OptionInfo("None", "Add network to prompt", gr.Dropdown, lambda: {"choices": ["None", *networks.available_networks]}, refresh=networks.list_available_networks),
36
+ "lora_preferred_name": shared.OptionInfo("Alias from file", "When adding to prompt, refer to Lora by", gr.Radio, {"choices": ["Alias from file", "Filename"]}),
37
+ "lora_add_hashes_to_infotext": shared.OptionInfo(True, "Add Lora hashes to infotext"),
38
+ "lora_show_all": shared.OptionInfo(False, "Always show all networks on the Lora page").info("otherwise, those detected as for incompatible version of Stable Diffusion will be hidden"),
39
+ "lora_hide_unknown_for_versions": shared.OptionInfo([], "Hide networks of unknown versions for model versions", gr.CheckboxGroup, {"choices": ["SD1", "SD2", "SDXL"]}),
40
+ "lora_in_memory_limit": shared.OptionInfo(0, "Number of Lora networks to keep cached in memory", gr.Number, {"precision": 0}),
41
+ "lora_not_found_warning_console": shared.OptionInfo(False, "Lora not found warning in console"),
42
+ "lora_not_found_gradio_warning": shared.OptionInfo(False, "Lora not found warning popup in webui"),
43
+ }))
44
+
45
+
46
+ shared.options_templates.update(shared.options_section(('compatibility', "Compatibility"), {
47
+ "lora_functional": shared.OptionInfo(False, "Lora/Networks: use old method that takes longer when you have multiple Loras active and produces same results as kohya-ss/sd-webui-additional-networks extension"),
48
+ }))
49
+
50
+
51
+ def create_lora_json(obj: network.NetworkOnDisk):
52
+ return {
53
+ "name": obj.name,
54
+ "alias": obj.alias,
55
+ "path": obj.filename,
56
+ "metadata": obj.metadata,
57
+ }
58
+
59
+
60
+ def api_networks(_: gr.Blocks, app: FastAPI):
61
+ @app.get("/sdapi/v1/loras")
62
+ async def get_loras():
63
+ return [create_lora_json(obj) for obj in networks.available_networks.values()]
64
+
65
+ @app.post("/sdapi/v1/refresh-loras")
66
+ async def refresh_loras():
67
+ return networks.list_available_networks()
68
+
69
+
70
+ script_callbacks.on_app_started(api_networks)
71
+
72
+ re_lora = re.compile("<lora:([^:]+):")
73
+
74
+
75
+ def infotext_pasted(infotext, d):
76
+ hashes = d.get("Lora hashes")
77
+ if not hashes:
78
+ return
79
+
80
+ hashes = [x.strip().split(':', 1) for x in hashes.split(",")]
81
+ hashes = {x[0].strip().replace(",", ""): x[1].strip() for x in hashes}
82
+
83
+ def network_replacement(m):
84
+ alias = m.group(1)
85
+ shorthash = hashes.get(alias)
86
+ if shorthash is None:
87
+ return m.group(0)
88
+
89
+ network_on_disk = networks.available_network_hash_lookup.get(shorthash)
90
+ if network_on_disk is None:
91
+ return m.group(0)
92
+
93
+ return f'<lora:{network_on_disk.get_alias()}:'
94
+
95
+ d["Prompt"] = re.sub(re_lora, network_replacement, d["Prompt"])
96
+
97
+
98
+ script_callbacks.on_infotext_pasted(infotext_pasted)
99
+
100
+ shared.opts.onchange("lora_in_memory_limit", networks.purge_networks_from_memory)
extensions-builtin/Lora/ui_edit_user_metadata.py ADDED
@@ -0,0 +1,222 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import datetime
2
+ import html
3
+ import random
4
+
5
+ import gradio as gr
6
+ import re
7
+
8
+ from modules import ui_extra_networks_user_metadata
9
+
10
+
11
+ def is_non_comma_tagset(tags):
12
+ average_tag_length = sum(len(x) for x in tags.keys()) / len(tags)
13
+
14
+ return average_tag_length >= 16
15
+
16
+
17
+ re_word = re.compile(r"[-_\w']+")
18
+ re_comma = re.compile(r" *, *")
19
+
20
+
21
+ def build_tags(metadata):
22
+ tags = {}
23
+
24
+ for _, tags_dict in metadata.get("ss_tag_frequency", {}).items():
25
+ for tag, tag_count in tags_dict.items():
26
+ tag = tag.strip()
27
+ tags[tag] = tags.get(tag, 0) + int(tag_count)
28
+
29
+ if tags and is_non_comma_tagset(tags):
30
+ new_tags = {}
31
+
32
+ for text, text_count in tags.items():
33
+ for word in re.findall(re_word, text):
34
+ if len(word) < 3:
35
+ continue
36
+
37
+ new_tags[word] = new_tags.get(word, 0) + text_count
38
+
39
+ tags = new_tags
40
+
41
+ ordered_tags = sorted(tags.keys(), key=tags.get, reverse=True)
42
+
43
+ return [(tag, tags[tag]) for tag in ordered_tags]
44
+
45
+
46
+ class LoraUserMetadataEditor(ui_extra_networks_user_metadata.UserMetadataEditor):
47
+ def __init__(self, ui, tabname, page):
48
+ super().__init__(ui, tabname, page)
49
+
50
+ self.select_sd_version = None
51
+
52
+ self.taginfo = None
53
+ self.edit_activation_text = None
54
+ self.slider_preferred_weight = None
55
+ self.edit_notes = None
56
+
57
+ def save_lora_user_metadata(self, name, desc, sd_version, activation_text, preferred_weight, negative_text, notes):
58
+ user_metadata = self.get_user_metadata(name)
59
+ user_metadata["description"] = desc
60
+ user_metadata["sd version"] = sd_version
61
+ user_metadata["activation text"] = activation_text
62
+ user_metadata["preferred weight"] = preferred_weight
63
+ user_metadata["negative text"] = negative_text
64
+ user_metadata["notes"] = notes
65
+
66
+ self.write_user_metadata(name, user_metadata)
67
+
68
+ def get_metadata_table(self, name):
69
+ table = super().get_metadata_table(name)
70
+ item = self.page.items.get(name, {})
71
+ metadata = item.get("metadata") or {}
72
+
73
+ keys = {
74
+ 'ss_output_name': "Output name:",
75
+ 'ss_sd_model_name': "Model:",
76
+ 'ss_clip_skip': "Clip skip:",
77
+ 'ss_network_module': "Kohya module:",
78
+ }
79
+
80
+ for key, label in keys.items():
81
+ value = metadata.get(key, None)
82
+ if value is not None and str(value) != "None":
83
+ table.append((label, html.escape(value)))
84
+
85
+ ss_training_started_at = metadata.get('ss_training_started_at')
86
+ if ss_training_started_at:
87
+ table.append(("Date trained:", datetime.datetime.utcfromtimestamp(float(ss_training_started_at)).strftime('%Y-%m-%d %H:%M')))
88
+
89
+ ss_bucket_info = metadata.get("ss_bucket_info")
90
+ if ss_bucket_info and "buckets" in ss_bucket_info:
91
+ resolutions = {}
92
+ for _, bucket in ss_bucket_info["buckets"].items():
93
+ resolution = bucket["resolution"]
94
+ resolution = f'{resolution[1]}x{resolution[0]}'
95
+
96
+ resolutions[resolution] = resolutions.get(resolution, 0) + int(bucket["count"])
97
+
98
+ resolutions_list = sorted(resolutions.keys(), key=resolutions.get, reverse=True)
99
+ resolutions_text = html.escape(", ".join(resolutions_list[0:4]))
100
+ if len(resolutions) > 4:
101
+ resolutions_text += ", ..."
102
+ resolutions_text = f"<span title='{html.escape(', '.join(resolutions_list))}'>{resolutions_text}</span>"
103
+
104
+ table.append(('Resolutions:' if len(resolutions_list) > 1 else 'Resolution:', resolutions_text))
105
+
106
+ image_count = 0
107
+ for _, params in metadata.get("ss_dataset_dirs", {}).items():
108
+ image_count += int(params.get("img_count", 0))
109
+
110
+ if image_count:
111
+ table.append(("Dataset size:", image_count))
112
+
113
+ return table
114
+
115
+ def put_values_into_components(self, name):
116
+ user_metadata = self.get_user_metadata(name)
117
+ values = super().put_values_into_components(name)
118
+
119
+ item = self.page.items.get(name, {})
120
+ metadata = item.get("metadata") or {}
121
+
122
+ tags = build_tags(metadata)
123
+ gradio_tags = [(tag, str(count)) for tag, count in tags[0:24]]
124
+
125
+ return [
126
+ *values[0:5],
127
+ item.get("sd_version", "Unknown"),
128
+ gr.HighlightedText.update(value=gradio_tags, visible=True if tags else False),
129
+ user_metadata.get('activation text', ''),
130
+ float(user_metadata.get('preferred weight', 0.0)),
131
+ user_metadata.get('negative text', ''),
132
+ gr.update(visible=True if tags else False),
133
+ gr.update(value=self.generate_random_prompt_from_tags(tags), visible=True if tags else False),
134
+ ]
135
+
136
+ def generate_random_prompt(self, name):
137
+ item = self.page.items.get(name, {})
138
+ metadata = item.get("metadata") or {}
139
+ tags = build_tags(metadata)
140
+
141
+ return self.generate_random_prompt_from_tags(tags)
142
+
143
+ def generate_random_prompt_from_tags(self, tags):
144
+ max_count = None
145
+ res = []
146
+ for tag, count in tags:
147
+ if not max_count:
148
+ max_count = count
149
+
150
+ v = random.random() * max_count
151
+ if count > v:
152
+ res.append(tag)
153
+
154
+ return ", ".join(sorted(res))
155
+
156
+ def create_extra_default_items_in_left_column(self):
157
+
158
+ # this would be a lot better as gr.Radio but I can't make it work
159
+ self.select_sd_version = gr.Dropdown(['SD1', 'SD2', 'SDXL', 'Unknown'], value='Unknown', label='Stable Diffusion version', interactive=True)
160
+
161
+ def create_editor(self):
162
+ self.create_default_editor_elems()
163
+
164
+ self.taginfo = gr.HighlightedText(label="Training dataset tags")
165
+ self.edit_activation_text = gr.Text(label='Activation text', info="Will be added to prompt along with Lora")
166
+ self.slider_preferred_weight = gr.Slider(label='Preferred weight', info="Set to 0 to disable", minimum=0.0, maximum=2.0, step=0.01)
167
+ self.edit_negative_text = gr.Text(label='Negative prompt', info="Will be added to negative prompts")
168
+ with gr.Row() as row_random_prompt:
169
+ with gr.Column(scale=8):
170
+ random_prompt = gr.Textbox(label='Random prompt', lines=4, max_lines=4, interactive=False)
171
+
172
+ with gr.Column(scale=1, min_width=120):
173
+ generate_random_prompt = gr.Button('Generate', size="lg", scale=1)
174
+
175
+ self.edit_notes = gr.TextArea(label='Notes', lines=4)
176
+
177
+ generate_random_prompt.click(fn=self.generate_random_prompt, inputs=[self.edit_name_input], outputs=[random_prompt], show_progress=False)
178
+
179
+ def select_tag(activation_text, evt: gr.SelectData):
180
+ tag = evt.value[0]
181
+
182
+ words = re.split(re_comma, activation_text)
183
+ if tag in words:
184
+ words = [x for x in words if x != tag and x.strip()]
185
+ return ", ".join(words)
186
+
187
+ return activation_text + ", " + tag if activation_text else tag
188
+
189
+ self.taginfo.select(fn=select_tag, inputs=[self.edit_activation_text], outputs=[self.edit_activation_text], show_progress=False)
190
+
191
+ self.create_default_buttons()
192
+
193
+ viewed_components = [
194
+ self.edit_name,
195
+ self.edit_description,
196
+ self.html_filedata,
197
+ self.html_preview,
198
+ self.edit_notes,
199
+ self.select_sd_version,
200
+ self.taginfo,
201
+ self.edit_activation_text,
202
+ self.slider_preferred_weight,
203
+ self.edit_negative_text,
204
+ row_random_prompt,
205
+ random_prompt,
206
+ ]
207
+
208
+ self.button_edit\
209
+ .click(fn=self.put_values_into_components, inputs=[self.edit_name_input], outputs=viewed_components)\
210
+ .then(fn=lambda: gr.update(visible=True), inputs=[], outputs=[self.box])
211
+
212
+ edited_components = [
213
+ self.edit_description,
214
+ self.select_sd_version,
215
+ self.edit_activation_text,
216
+ self.slider_preferred_weight,
217
+ self.edit_negative_text,
218
+ self.edit_notes,
219
+ ]
220
+
221
+
222
+ self.setup_save_handler(self.button_save, self.save_lora_user_metadata, edited_components)
extensions-builtin/Lora/ui_extra_networks_lora.py ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+
3
+ import network
4
+ import networks
5
+
6
+ from modules import shared, ui_extra_networks
7
+ from modules.ui_extra_networks import quote_js
8
+ from ui_edit_user_metadata import LoraUserMetadataEditor
9
+
10
+
11
+ class ExtraNetworksPageLora(ui_extra_networks.ExtraNetworksPage):
12
+ def __init__(self):
13
+ super().__init__('Lora')
14
+
15
+ def refresh(self):
16
+ networks.list_available_networks()
17
+
18
+ def create_item(self, name, index=None, enable_filter=True):
19
+ lora_on_disk = networks.available_networks.get(name)
20
+ if lora_on_disk is None:
21
+ return
22
+
23
+ path, ext = os.path.splitext(lora_on_disk.filename)
24
+
25
+ alias = lora_on_disk.get_alias()
26
+
27
+ search_terms = [self.search_terms_from_path(lora_on_disk.filename)]
28
+ if lora_on_disk.hash:
29
+ search_terms.append(lora_on_disk.hash)
30
+ item = {
31
+ "name": name,
32
+ "filename": lora_on_disk.filename,
33
+ "shorthash": lora_on_disk.shorthash,
34
+ "preview": self.find_preview(path),
35
+ "description": self.find_description(path),
36
+ "search_terms": search_terms,
37
+ "local_preview": f"{path}.{shared.opts.samples_format}",
38
+ "metadata": lora_on_disk.metadata,
39
+ "sort_keys": {'default': index, **self.get_sort_keys(lora_on_disk.filename)},
40
+ "sd_version": lora_on_disk.sd_version.name,
41
+ }
42
+
43
+ self.read_user_metadata(item)
44
+ activation_text = item["user_metadata"].get("activation text")
45
+ preferred_weight = item["user_metadata"].get("preferred weight", 0.0)
46
+ item["prompt"] = quote_js(f"<lora:{alias}:") + " + " + (str(preferred_weight) if preferred_weight else "opts.extra_networks_default_multiplier") + " + " + quote_js(">")
47
+
48
+ if activation_text:
49
+ item["prompt"] += " + " + quote_js(" " + activation_text)
50
+
51
+ negative_prompt = item["user_metadata"].get("negative text")
52
+ item["negative_prompt"] = quote_js("")
53
+ if negative_prompt:
54
+ item["negative_prompt"] = quote_js('(' + negative_prompt + ':1)')
55
+
56
+ sd_version = item["user_metadata"].get("sd version")
57
+ if sd_version in network.SdVersion.__members__:
58
+ item["sd_version"] = sd_version
59
+ sd_version = network.SdVersion[sd_version]
60
+ else:
61
+ sd_version = lora_on_disk.sd_version
62
+
63
+ if shared.opts.lora_show_all or not enable_filter:
64
+ pass
65
+ elif sd_version == network.SdVersion.Unknown:
66
+ model_version = network.SdVersion.SDXL if shared.sd_model.is_sdxl else network.SdVersion.SD2 if shared.sd_model.is_sd2 else network.SdVersion.SD1
67
+ if model_version.name in shared.opts.lora_hide_unknown_for_versions:
68
+ return None
69
+ elif shared.sd_model.is_sdxl and sd_version != network.SdVersion.SDXL:
70
+ return None
71
+ elif shared.sd_model.is_sd2 and sd_version != network.SdVersion.SD2:
72
+ return None
73
+ elif shared.sd_model.is_sd1 and sd_version != network.SdVersion.SD1:
74
+ return None
75
+
76
+ return item
77
+
78
+ def list_items(self):
79
+ # instantiate a list to protect against concurrent modification
80
+ names = list(networks.available_networks)
81
+ for index, name in enumerate(names):
82
+ item = self.create_item(name, index)
83
+ if item is not None:
84
+ yield item
85
+
86
+ def allowed_directories_for_previews(self):
87
+ return [shared.cmd_opts.lora_dir]
88
+
89
+ def create_user_metadata_editor(self, ui, tabname):
90
+ return LoraUserMetadataEditor(ui, tabname, self)
extensions-builtin/ScuNET/preload.py ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ import os
2
+ from modules import paths
3
+
4
+
5
+ def preload(parser):
6
+ parser.add_argument("--scunet-models-path", type=str, help="Path to directory with ScuNET model file(s).", default=os.path.join(paths.models_path, 'ScuNET'))
extensions-builtin/ScuNET/scripts/scunet_model.py ADDED
@@ -0,0 +1,74 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import sys
2
+
3
+ import PIL.Image
4
+
5
+ import modules.upscaler
6
+ from modules import devices, errors, modelloader, script_callbacks, shared, upscaler_utils
7
+
8
+
9
+ class UpscalerScuNET(modules.upscaler.Upscaler):
10
+ def __init__(self, dirname):
11
+ self.name = "ScuNET"
12
+ self.model_name = "ScuNET GAN"
13
+ self.model_name2 = "ScuNET PSNR"
14
+ self.model_url = "https://github.com/cszn/KAIR/releases/download/v1.0/scunet_color_real_gan.pth"
15
+ self.model_url2 = "https://github.com/cszn/KAIR/releases/download/v1.0/scunet_color_real_psnr.pth"
16
+ self.user_path = dirname
17
+ super().__init__()
18
+ model_paths = self.find_models(ext_filter=[".pth"])
19
+ scalers = []
20
+ add_model2 = True
21
+ for file in model_paths:
22
+ if file.startswith("http"):
23
+ name = self.model_name
24
+ else:
25
+ name = modelloader.friendly_name(file)
26
+ if name == self.model_name2 or file == self.model_url2:
27
+ add_model2 = False
28
+ try:
29
+ scaler_data = modules.upscaler.UpscalerData(name, file, self, 4)
30
+ scalers.append(scaler_data)
31
+ except Exception:
32
+ errors.report(f"Error loading ScuNET model: {file}", exc_info=True)
33
+ if add_model2:
34
+ scaler_data2 = modules.upscaler.UpscalerData(self.model_name2, self.model_url2, self)
35
+ scalers.append(scaler_data2)
36
+ self.scalers = scalers
37
+
38
+ def do_upscale(self, img: PIL.Image.Image, selected_file):
39
+ devices.torch_gc()
40
+ try:
41
+ model = self.load_model(selected_file)
42
+ except Exception as e:
43
+ print(f"ScuNET: Unable to load model from {selected_file}: {e}", file=sys.stderr)
44
+ return img
45
+
46
+ img = upscaler_utils.upscale_2(
47
+ img,
48
+ model,
49
+ tile_size=shared.opts.SCUNET_tile,
50
+ tile_overlap=shared.opts.SCUNET_tile_overlap,
51
+ scale=1, # ScuNET is a denoising model, not an upscaler
52
+ desc='ScuNET',
53
+ )
54
+ devices.torch_gc()
55
+ return img
56
+
57
+ def load_model(self, path: str):
58
+ device = devices.get_device_for('scunet')
59
+ if path.startswith("http"):
60
+ # TODO: this doesn't use `path` at all?
61
+ filename = modelloader.load_file_from_url(self.model_url, model_dir=self.model_download_path, file_name=f"{self.name}.pth")
62
+ else:
63
+ filename = path
64
+ return modelloader.load_spandrel_model(filename, device=device, expected_architecture='SCUNet')
65
+
66
+
67
+ def on_ui_settings():
68
+ import gradio as gr
69
+
70
+ shared.opts.add_option("SCUNET_tile", shared.OptionInfo(256, "Tile size for SCUNET upscalers.", gr.Slider, {"minimum": 0, "maximum": 512, "step": 16}, section=('upscaling', "Upscaling")).info("0 = no tiling"))
71
+ shared.opts.add_option("SCUNET_tile_overlap", shared.OptionInfo(8, "Tile overlap for SCUNET upscalers.", gr.Slider, {"minimum": 0, "maximum": 64, "step": 1}, section=('upscaling', "Upscaling")).info("Low values = visible seam"))
72
+
73
+
74
+ script_callbacks.on_ui_settings(on_ui_settings)
extensions-builtin/SwinIR/preload.py ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ import os
2
+ from modules import paths
3
+
4
+
5
+ def preload(parser):
6
+ parser.add_argument("--swinir-models-path", type=str, help="Path to directory with SwinIR model file(s).", default=os.path.join(paths.models_path, 'SwinIR'))
extensions-builtin/SwinIR/scripts/swinir_model.py ADDED
@@ -0,0 +1,98 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import logging
2
+ import sys
3
+
4
+ import torch
5
+ from PIL import Image
6
+
7
+ from modules import devices, modelloader, script_callbacks, shared, upscaler_utils
8
+ from modules.upscaler import Upscaler, UpscalerData
9
+ from modules_forge.forge_util import prepare_free_memory
10
+
11
+ SWINIR_MODEL_URL = "https://github.com/JingyunLiang/SwinIR/releases/download/v0.0/003_realSR_BSRGAN_DFOWMFC_s64w8_SwinIR-L_x4_GAN.pth"
12
+
13
+ logger = logging.getLogger(__name__)
14
+
15
+
16
+ class UpscalerSwinIR(Upscaler):
17
+ def __init__(self, dirname):
18
+ self._cached_model = None # keep the model when SWIN_torch_compile is on to prevent re-compile every runs
19
+ self._cached_model_config = None # to clear '_cached_model' when changing model (v1/v2) or settings
20
+ self.name = "SwinIR"
21
+ self.model_url = SWINIR_MODEL_URL
22
+ self.model_name = "SwinIR 4x"
23
+ self.user_path = dirname
24
+ super().__init__()
25
+ scalers = []
26
+ model_files = self.find_models(ext_filter=[".pt", ".pth"])
27
+ for model in model_files:
28
+ if model.startswith("http"):
29
+ name = self.model_name
30
+ else:
31
+ name = modelloader.friendly_name(model)
32
+ model_data = UpscalerData(name, model, self)
33
+ scalers.append(model_data)
34
+ self.scalers = scalers
35
+
36
+ def do_upscale(self, img: Image.Image, model_file: str) -> Image.Image:
37
+ prepare_free_memory()
38
+
39
+ current_config = (model_file, shared.opts.SWIN_tile)
40
+
41
+ if self._cached_model_config == current_config:
42
+ model = self._cached_model
43
+ else:
44
+ try:
45
+ model = self.load_model(model_file)
46
+ except Exception as e:
47
+ print(f"Failed loading SwinIR model {model_file}: {e}", file=sys.stderr)
48
+ return img
49
+ self._cached_model = model
50
+ self._cached_model_config = current_config
51
+
52
+ img = upscaler_utils.upscale_2(
53
+ img,
54
+ model,
55
+ tile_size=shared.opts.SWIN_tile,
56
+ tile_overlap=shared.opts.SWIN_tile_overlap,
57
+ scale=model.scale,
58
+ desc="SwinIR",
59
+ )
60
+ devices.torch_gc()
61
+ return img
62
+
63
+ def load_model(self, path, scale=4):
64
+ if path.startswith("http"):
65
+ filename = modelloader.load_file_from_url(
66
+ url=path,
67
+ model_dir=self.model_download_path,
68
+ file_name=f"{self.model_name.replace(' ', '_')}.pth",
69
+ )
70
+ else:
71
+ filename = path
72
+
73
+ model_descriptor = modelloader.load_spandrel_model(
74
+ filename,
75
+ device=self._get_device(),
76
+ prefer_half=(devices.dtype == torch.float16),
77
+ expected_architecture="SwinIR",
78
+ )
79
+ if getattr(shared.opts, 'SWIN_torch_compile', False):
80
+ try:
81
+ model_descriptor.model.compile()
82
+ except Exception:
83
+ logger.warning("Failed to compile SwinIR model, fallback to JIT", exc_info=True)
84
+ return model_descriptor
85
+
86
+ def _get_device(self):
87
+ return devices.get_device_for('swinir')
88
+
89
+
90
+ def on_ui_settings():
91
+ import gradio as gr
92
+
93
+ shared.opts.add_option("SWIN_tile", shared.OptionInfo(192, "Tile size for all SwinIR.", gr.Slider, {"minimum": 16, "maximum": 512, "step": 16}, section=('upscaling', "Upscaling")))
94
+ shared.opts.add_option("SWIN_tile_overlap", shared.OptionInfo(8, "Tile overlap, in pixels for SwinIR. Low values = visible seam.", gr.Slider, {"minimum": 0, "maximum": 48, "step": 1}, section=('upscaling', "Upscaling")))
95
+ shared.opts.add_option("SWIN_torch_compile", shared.OptionInfo(False, "Use torch.compile to accelerate SwinIR.", gr.Checkbox, {"interactive": True}, section=('upscaling', "Upscaling")).info("Takes longer on first run"))
96
+
97
+
98
+ script_callbacks.on_ui_settings(on_ui_settings)
extensions-builtin/canvas-zoom-and-pan/javascript/zoom.js ADDED
@@ -0,0 +1,968 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ onUiLoaded(async() => {
2
+ const elementIDs = {
3
+ img2imgTabs: "#mode_img2img .tab-nav",
4
+ inpaint: "#img2maskimg",
5
+ inpaintSketch: "#inpaint_sketch",
6
+ rangeGroup: "#img2img_column_size",
7
+ sketch: "#img2img_sketch"
8
+ };
9
+ const tabNameToElementId = {
10
+ "Inpaint sketch": elementIDs.inpaintSketch,
11
+ "Inpaint": elementIDs.inpaint,
12
+ "Sketch": elementIDs.sketch
13
+ };
14
+
15
+
16
+ // Helper functions
17
+ // Get active tab
18
+
19
+ /**
20
+ * Waits for an element to be present in the DOM.
21
+ */
22
+ const waitForElement = (id) => new Promise(resolve => {
23
+ const checkForElement = () => {
24
+ const element = document.querySelector(id);
25
+ if (element) return resolve(element);
26
+ setTimeout(checkForElement, 100);
27
+ };
28
+ checkForElement();
29
+ });
30
+
31
+ function getActiveTab(elements, all = false) {
32
+ const tabs = elements.img2imgTabs.querySelectorAll("button");
33
+
34
+ if (all) return tabs;
35
+
36
+ for (let tab of tabs) {
37
+ if (tab.classList.contains("selected")) {
38
+ return tab;
39
+ }
40
+ }
41
+ }
42
+
43
+ // Get tab ID
44
+ function getTabId(elements) {
45
+ const activeTab = getActiveTab(elements);
46
+ return tabNameToElementId[activeTab.innerText];
47
+ }
48
+
49
+ // Wait until opts loaded
50
+ async function waitForOpts() {
51
+ for (; ;) {
52
+ if (window.opts && Object.keys(window.opts).length) {
53
+ return window.opts;
54
+ }
55
+ await new Promise(resolve => setTimeout(resolve, 100));
56
+ }
57
+ }
58
+
59
+ // Detect whether the element has a horizontal scroll bar
60
+ function hasHorizontalScrollbar(element) {
61
+ return element.scrollWidth > element.clientWidth;
62
+ }
63
+
64
+ // Function for defining the "Ctrl", "Shift" and "Alt" keys
65
+ function isModifierKey(event, key) {
66
+ switch (key) {
67
+ case "Ctrl":
68
+ return event.ctrlKey;
69
+ case "Shift":
70
+ return event.shiftKey;
71
+ case "Alt":
72
+ return event.altKey;
73
+ default:
74
+ return false;
75
+ }
76
+ }
77
+
78
+ // Check if hotkey is valid
79
+ function isValidHotkey(value) {
80
+ const specialKeys = ["Ctrl", "Alt", "Shift", "Disable"];
81
+ return (
82
+ (typeof value === "string" &&
83
+ value.length === 1 &&
84
+ /[a-z]/i.test(value)) ||
85
+ specialKeys.includes(value)
86
+ );
87
+ }
88
+
89
+ // Normalize hotkey
90
+ function normalizeHotkey(hotkey) {
91
+ return hotkey.length === 1 ? "Key" + hotkey.toUpperCase() : hotkey;
92
+ }
93
+
94
+ // Format hotkey for display
95
+ function formatHotkeyForDisplay(hotkey) {
96
+ return hotkey.startsWith("Key") ? hotkey.slice(3) : hotkey;
97
+ }
98
+
99
+ // Create hotkey configuration with the provided options
100
+ function createHotkeyConfig(defaultHotkeysConfig, hotkeysConfigOpts) {
101
+ const result = {}; // Resulting hotkey configuration
102
+ const usedKeys = new Set(); // Set of used hotkeys
103
+
104
+ // Iterate through defaultHotkeysConfig keys
105
+ for (const key in defaultHotkeysConfig) {
106
+ const userValue = hotkeysConfigOpts[key]; // User-provided hotkey value
107
+ const defaultValue = defaultHotkeysConfig[key]; // Default hotkey value
108
+
109
+ // Apply appropriate value for undefined, boolean, or object userValue
110
+ if (
111
+ userValue === undefined ||
112
+ typeof userValue === "boolean" ||
113
+ typeof userValue === "object" ||
114
+ userValue === "disable"
115
+ ) {
116
+ result[key] =
117
+ userValue === undefined ? defaultValue : userValue;
118
+ } else if (isValidHotkey(userValue)) {
119
+ const normalizedUserValue = normalizeHotkey(userValue);
120
+
121
+ // Check for conflicting hotkeys
122
+ if (!usedKeys.has(normalizedUserValue)) {
123
+ usedKeys.add(normalizedUserValue);
124
+ result[key] = normalizedUserValue;
125
+ } else {
126
+ console.error(
127
+ `Hotkey: ${formatHotkeyForDisplay(
128
+ userValue
129
+ )} for ${key} is repeated and conflicts with another hotkey. The default hotkey is used: ${formatHotkeyForDisplay(
130
+ defaultValue
131
+ )}`
132
+ );
133
+ result[key] = defaultValue;
134
+ }
135
+ } else {
136
+ console.error(
137
+ `Hotkey: ${formatHotkeyForDisplay(
138
+ userValue
139
+ )} for ${key} is not valid. The default hotkey is used: ${formatHotkeyForDisplay(
140
+ defaultValue
141
+ )}`
142
+ );
143
+ result[key] = defaultValue;
144
+ }
145
+ }
146
+
147
+ return result;
148
+ }
149
+
150
+ // Disables functions in the config object based on the provided list of function names
151
+ function disableFunctions(config, disabledFunctions) {
152
+ // Bind the hasOwnProperty method to the functionMap object to avoid errors
153
+ const hasOwnProperty =
154
+ Object.prototype.hasOwnProperty.bind(functionMap);
155
+
156
+ // Loop through the disabledFunctions array and disable the corresponding functions in the config object
157
+ disabledFunctions.forEach(funcName => {
158
+ if (hasOwnProperty(funcName)) {
159
+ const key = functionMap[funcName];
160
+ config[key] = "disable";
161
+ }
162
+ });
163
+
164
+ // Return the updated config object
165
+ return config;
166
+ }
167
+
168
+ /**
169
+ * The restoreImgRedMask function displays a red mask around an image to indicate the aspect ratio.
170
+ * If the image display property is set to 'none', the mask breaks. To fix this, the function
171
+ * temporarily sets the display property to 'block' and then hides the mask again after 300 milliseconds
172
+ * to avoid breaking the canvas. Additionally, the function adjusts the mask to work correctly on
173
+ * very long images.
174
+ */
175
+ function restoreImgRedMask(elements) {
176
+ const mainTabId = getTabId(elements);
177
+
178
+ if (!mainTabId) return;
179
+
180
+ const mainTab = gradioApp().querySelector(mainTabId);
181
+ const img = mainTab.querySelector("img");
182
+ const imageARPreview = gradioApp().querySelector("#imageARPreview");
183
+
184
+ if (!img || !imageARPreview) return;
185
+
186
+ imageARPreview.style.transform = "";
187
+ if (parseFloat(mainTab.style.width) > 865) {
188
+ const transformString = mainTab.style.transform;
189
+ const scaleMatch = transformString.match(
190
+ /scale\(([-+]?[0-9]*\.?[0-9]+)\)/
191
+ );
192
+ let zoom = 1; // default zoom
193
+
194
+ if (scaleMatch && scaleMatch[1]) {
195
+ zoom = Number(scaleMatch[1]);
196
+ }
197
+
198
+ imageARPreview.style.transformOrigin = "0 0";
199
+ imageARPreview.style.transform = `scale(${zoom})`;
200
+ }
201
+
202
+ if (img.style.display !== "none") return;
203
+
204
+ img.style.display = "block";
205
+
206
+ setTimeout(() => {
207
+ img.style.display = "none";
208
+ }, 400);
209
+ }
210
+
211
+ const hotkeysConfigOpts = await waitForOpts();
212
+
213
+ // Default config
214
+ const defaultHotkeysConfig = {
215
+ canvas_hotkey_zoom: "Alt",
216
+ canvas_hotkey_adjust: "Ctrl",
217
+ canvas_hotkey_reset: "KeyR",
218
+ canvas_hotkey_fullscreen: "KeyS",
219
+ canvas_hotkey_move: "KeyF",
220
+ canvas_hotkey_overlap: "KeyO",
221
+ canvas_hotkey_shrink_brush: "KeyQ",
222
+ canvas_hotkey_grow_brush: "KeyW",
223
+ canvas_disabled_functions: [],
224
+ canvas_show_tooltip: true,
225
+ canvas_auto_expand: true,
226
+ canvas_blur_prompt: false,
227
+ };
228
+
229
+ const functionMap = {
230
+ "Zoom": "canvas_hotkey_zoom",
231
+ "Adjust brush size": "canvas_hotkey_adjust",
232
+ "Hotkey shrink brush": "canvas_hotkey_shrink_brush",
233
+ "Hotkey enlarge brush": "canvas_hotkey_grow_brush",
234
+ "Moving canvas": "canvas_hotkey_move",
235
+ "Fullscreen": "canvas_hotkey_fullscreen",
236
+ "Reset Zoom": "canvas_hotkey_reset",
237
+ "Overlap": "canvas_hotkey_overlap"
238
+ };
239
+
240
+ // Loading the configuration from opts
241
+ const preHotkeysConfig = createHotkeyConfig(
242
+ defaultHotkeysConfig,
243
+ hotkeysConfigOpts
244
+ );
245
+
246
+ // Disable functions that are not needed by the user
247
+ const hotkeysConfig = disableFunctions(
248
+ preHotkeysConfig,
249
+ preHotkeysConfig.canvas_disabled_functions
250
+ );
251
+
252
+ let isMoving = false;
253
+ let mouseX, mouseY;
254
+ let activeElement;
255
+
256
+ const elements = Object.fromEntries(
257
+ Object.keys(elementIDs).map(id => [
258
+ id,
259
+ gradioApp().querySelector(elementIDs[id])
260
+ ])
261
+ );
262
+ const elemData = {};
263
+
264
+ // Apply functionality to the range inputs. Restore redmask and correct for long images.
265
+ const rangeInputs = elements.rangeGroup ?
266
+ Array.from(elements.rangeGroup.querySelectorAll("input")) :
267
+ [
268
+ gradioApp().querySelector("#img2img_width input[type='range']"),
269
+ gradioApp().querySelector("#img2img_height input[type='range']")
270
+ ];
271
+
272
+ for (const input of rangeInputs) {
273
+ input?.addEventListener("input", () => restoreImgRedMask(elements));
274
+ }
275
+
276
+ function applyZoomAndPan(elemId, isExtension = true) {
277
+ const targetElement = gradioApp().querySelector(elemId);
278
+
279
+ if (!targetElement) {
280
+ console.log("Element not found");
281
+ return;
282
+ }
283
+
284
+ targetElement.style.transformOrigin = "0 0";
285
+
286
+ elemData[elemId] = {
287
+ zoom: 1,
288
+ panX: 0,
289
+ panY: 0
290
+ };
291
+ let fullScreenMode = false;
292
+
293
+ // Create tooltip
294
+ function createTooltip() {
295
+ const toolTipElemnt =
296
+ targetElement.querySelector(".image-container");
297
+ const tooltip = document.createElement("div");
298
+ tooltip.className = "canvas-tooltip";
299
+
300
+ // Creating an item of information
301
+ const info = document.createElement("i");
302
+ info.className = "canvas-tooltip-info";
303
+ info.textContent = "";
304
+
305
+ // Create a container for the contents of the tooltip
306
+ const tooltipContent = document.createElement("div");
307
+ tooltipContent.className = "canvas-tooltip-content";
308
+
309
+ // Define an array with hotkey information and their actions
310
+ const hotkeysInfo = [
311
+ {
312
+ configKey: "canvas_hotkey_zoom",
313
+ action: "Zoom canvas",
314
+ keySuffix: " + wheel"
315
+ },
316
+ {
317
+ configKey: "canvas_hotkey_adjust",
318
+ action: "Adjust brush size",
319
+ keySuffix: " + wheel"
320
+ },
321
+ {configKey: "canvas_hotkey_reset", action: "Reset zoom"},
322
+ {
323
+ configKey: "canvas_hotkey_fullscreen",
324
+ action: "Fullscreen mode"
325
+ },
326
+ {configKey: "canvas_hotkey_move", action: "Move canvas"},
327
+ {configKey: "canvas_hotkey_overlap", action: "Overlap"}
328
+ ];
329
+
330
+ // Create hotkeys array with disabled property based on the config values
331
+ const hotkeys = hotkeysInfo.map(info => {
332
+ const configValue = hotkeysConfig[info.configKey];
333
+ const key = info.keySuffix ?
334
+ `${configValue}${info.keySuffix}` :
335
+ configValue.charAt(configValue.length - 1);
336
+ return {
337
+ key,
338
+ action: info.action,
339
+ disabled: configValue === "disable"
340
+ };
341
+ });
342
+
343
+ for (const hotkey of hotkeys) {
344
+ if (hotkey.disabled) {
345
+ continue;
346
+ }
347
+
348
+ const p = document.createElement("p");
349
+ p.innerHTML = `<b>${hotkey.key}</b> - ${hotkey.action}`;
350
+ tooltipContent.appendChild(p);
351
+ }
352
+
353
+ // Add information and content elements to the tooltip element
354
+ tooltip.appendChild(info);
355
+ tooltip.appendChild(tooltipContent);
356
+
357
+ // Add a hint element to the target element
358
+ toolTipElemnt.appendChild(tooltip);
359
+ }
360
+
361
+ //Show tool tip if setting enable
362
+ if (hotkeysConfig.canvas_show_tooltip) {
363
+ createTooltip();
364
+ }
365
+
366
+ // In the course of research, it was found that the tag img is very harmful when zooming and creates white canvases. This hack allows you to almost never think about this problem, it has no effect on webui.
367
+ function fixCanvas() {
368
+ const activeTab = getActiveTab(elements).textContent.trim();
369
+
370
+ if (activeTab !== "img2img") {
371
+ const img = targetElement.querySelector(`${elemId} img`);
372
+
373
+ if (img && img.style.display !== "none") {
374
+ img.style.display = "none";
375
+ img.style.visibility = "hidden";
376
+ }
377
+ }
378
+ }
379
+
380
+ // Reset the zoom level and pan position of the target element to their initial values
381
+ function resetZoom() {
382
+ elemData[elemId] = {
383
+ zoomLevel: 1,
384
+ panX: 0,
385
+ panY: 0
386
+ };
387
+
388
+ if (isExtension) {
389
+ targetElement.style.overflow = "hidden";
390
+ }
391
+
392
+ targetElement.isZoomed = false;
393
+
394
+ fixCanvas();
395
+ targetElement.style.transform = `scale(${elemData[elemId].zoomLevel}) translate(${elemData[elemId].panX}px, ${elemData[elemId].panY}px)`;
396
+
397
+ const canvas = gradioApp().querySelector(
398
+ `${elemId} canvas[key="interface"]`
399
+ );
400
+
401
+ toggleOverlap("off");
402
+ fullScreenMode = false;
403
+
404
+ const closeBtn = targetElement.querySelector("button[aria-label='Remove Image']");
405
+ if (closeBtn) {
406
+ closeBtn.addEventListener("click", resetZoom);
407
+ }
408
+
409
+ if (canvas && isExtension) {
410
+ const parentElement = targetElement.closest('[id^="component-"]');
411
+ if (
412
+ canvas &&
413
+ parseFloat(canvas.style.width) > parentElement.offsetWidth &&
414
+ parseFloat(targetElement.style.width) > parentElement.offsetWidth
415
+ ) {
416
+ fitToElement();
417
+ return;
418
+ }
419
+
420
+ }
421
+
422
+ if (
423
+ canvas &&
424
+ !isExtension &&
425
+ parseFloat(canvas.style.width) > 865 &&
426
+ parseFloat(targetElement.style.width) > 865
427
+ ) {
428
+ fitToElement();
429
+ return;
430
+ }
431
+
432
+ targetElement.style.width = "";
433
+ }
434
+
435
+ // Toggle the zIndex of the target element between two values, allowing it to overlap or be overlapped by other elements
436
+ function toggleOverlap(forced = "") {
437
+ const zIndex1 = "0";
438
+ const zIndex2 = "998";
439
+
440
+ targetElement.style.zIndex =
441
+ targetElement.style.zIndex !== zIndex2 ? zIndex2 : zIndex1;
442
+
443
+ if (forced === "off") {
444
+ targetElement.style.zIndex = zIndex1;
445
+ } else if (forced === "on") {
446
+ targetElement.style.zIndex = zIndex2;
447
+ }
448
+ }
449
+
450
+ // Adjust the brush size based on the deltaY value from a mouse wheel event
451
+ function adjustBrushSize(
452
+ elemId,
453
+ deltaY,
454
+ withoutValue = false,
455
+ percentage = 5
456
+ ) {
457
+ const input =
458
+ gradioApp().querySelector(
459
+ `${elemId} input[aria-label='Brush radius']`
460
+ ) ||
461
+ gradioApp().querySelector(
462
+ `${elemId} button[aria-label="Use brush"]`
463
+ );
464
+
465
+ if (input) {
466
+ input.click();
467
+ if (!withoutValue) {
468
+ const maxValue =
469
+ parseFloat(input.getAttribute("max")) || 100;
470
+ const changeAmount = maxValue * (percentage / 100);
471
+ const newValue =
472
+ parseFloat(input.value) +
473
+ (deltaY > 0 ? -changeAmount : changeAmount);
474
+ input.value = Math.min(Math.max(newValue, 0), maxValue);
475
+ input.dispatchEvent(new Event("change"));
476
+ }
477
+ }
478
+ }
479
+
480
+ // Reset zoom when uploading a new image
481
+ const fileInput = gradioApp().querySelector(
482
+ `${elemId} input[type="file"][accept="image/*"].svelte-116rqfv`
483
+ );
484
+ fileInput.addEventListener("click", resetZoom);
485
+
486
+ // Update the zoom level and pan position of the target element based on the values of the zoomLevel, panX and panY variables
487
+ function updateZoom(newZoomLevel, mouseX, mouseY) {
488
+ newZoomLevel = Math.max(0.1, Math.min(newZoomLevel, 15));
489
+
490
+ elemData[elemId].panX +=
491
+ mouseX - (mouseX * newZoomLevel) / elemData[elemId].zoomLevel;
492
+ elemData[elemId].panY +=
493
+ mouseY - (mouseY * newZoomLevel) / elemData[elemId].zoomLevel;
494
+
495
+ targetElement.style.transformOrigin = "0 0";
496
+ targetElement.style.transform = `translate(${elemData[elemId].panX}px, ${elemData[elemId].panY}px) scale(${newZoomLevel})`;
497
+
498
+ toggleOverlap("on");
499
+ if (isExtension) {
500
+ targetElement.style.overflow = "visible";
501
+ }
502
+
503
+ return newZoomLevel;
504
+ }
505
+
506
+ // Change the zoom level based on user interaction
507
+ function changeZoomLevel(operation, e) {
508
+ if (isModifierKey(e, hotkeysConfig.canvas_hotkey_zoom)) {
509
+ e.preventDefault();
510
+
511
+ let zoomPosX, zoomPosY;
512
+ let delta = 0.2;
513
+ if (elemData[elemId].zoomLevel > 7) {
514
+ delta = 0.9;
515
+ } else if (elemData[elemId].zoomLevel > 2) {
516
+ delta = 0.6;
517
+ }
518
+
519
+ zoomPosX = e.clientX;
520
+ zoomPosY = e.clientY;
521
+
522
+ fullScreenMode = false;
523
+ elemData[elemId].zoomLevel = updateZoom(
524
+ elemData[elemId].zoomLevel +
525
+ (operation === "+" ? delta : -delta),
526
+ zoomPosX - targetElement.getBoundingClientRect().left,
527
+ zoomPosY - targetElement.getBoundingClientRect().top
528
+ );
529
+
530
+ targetElement.isZoomed = true;
531
+ }
532
+ }
533
+
534
+ /**
535
+ * This function fits the target element to the screen by calculating
536
+ * the required scale and offsets. It also updates the global variables
537
+ * zoomLevel, panX, and panY to reflect the new state.
538
+ */
539
+
540
+ function fitToElement() {
541
+ //Reset Zoom
542
+ targetElement.style.transform = `translate(${0}px, ${0}px) scale(${1})`;
543
+
544
+ let parentElement;
545
+
546
+ if (isExtension) {
547
+ parentElement = targetElement.closest('[id^="component-"]');
548
+ } else {
549
+ parentElement = targetElement.parentElement;
550
+ }
551
+
552
+
553
+ // Get element and screen dimensions
554
+ const elementWidth = targetElement.offsetWidth;
555
+ const elementHeight = targetElement.offsetHeight;
556
+
557
+ const screenWidth = parentElement.clientWidth;
558
+ const screenHeight = parentElement.clientHeight;
559
+
560
+ // Get element's coordinates relative to the parent element
561
+ const elementRect = targetElement.getBoundingClientRect();
562
+ const parentRect = parentElement.getBoundingClientRect();
563
+ const elementX = elementRect.x - parentRect.x;
564
+
565
+ // Calculate scale and offsets
566
+ const scaleX = screenWidth / elementWidth;
567
+ const scaleY = screenHeight / elementHeight;
568
+ const scale = Math.min(scaleX, scaleY);
569
+
570
+ const transformOrigin =
571
+ window.getComputedStyle(targetElement).transformOrigin;
572
+ const [originX, originY] = transformOrigin.split(" ");
573
+ const originXValue = parseFloat(originX);
574
+ const originYValue = parseFloat(originY);
575
+
576
+ const offsetX =
577
+ (screenWidth - elementWidth * scale) / 2 -
578
+ originXValue * (1 - scale);
579
+ const offsetY =
580
+ (screenHeight - elementHeight * scale) / 2.5 -
581
+ originYValue * (1 - scale);
582
+
583
+ // Apply scale and offsets to the element
584
+ targetElement.style.transform = `translate(${offsetX}px, ${offsetY}px) scale(${scale})`;
585
+
586
+ // Update global variables
587
+ elemData[elemId].zoomLevel = scale;
588
+ elemData[elemId].panX = offsetX;
589
+ elemData[elemId].panY = offsetY;
590
+
591
+ fullScreenMode = false;
592
+ toggleOverlap("off");
593
+ }
594
+
595
+ /**
596
+ * This function fits the target element to the screen by calculating
597
+ * the required scale and offsets. It also updates the global variables
598
+ * zoomLevel, panX, and panY to reflect the new state.
599
+ */
600
+
601
+ // Fullscreen mode
602
+ function fitToScreen() {
603
+ const canvas = gradioApp().querySelector(
604
+ `${elemId} canvas[key="interface"]`
605
+ );
606
+
607
+ if (!canvas) return;
608
+
609
+ if (canvas.offsetWidth > 862 || isExtension) {
610
+ targetElement.style.width = (canvas.offsetWidth + 2) + "px";
611
+ }
612
+
613
+ if (isExtension) {
614
+ targetElement.style.overflow = "visible";
615
+ }
616
+
617
+ if (fullScreenMode) {
618
+ resetZoom();
619
+ fullScreenMode = false;
620
+ return;
621
+ }
622
+
623
+ //Reset Zoom
624
+ targetElement.style.transform = `translate(${0}px, ${0}px) scale(${1})`;
625
+
626
+ // Get scrollbar width to right-align the image
627
+ const scrollbarWidth =
628
+ window.innerWidth - document.documentElement.clientWidth;
629
+
630
+ // Get element and screen dimensions
631
+ const elementWidth = targetElement.offsetWidth;
632
+ const elementHeight = targetElement.offsetHeight;
633
+ const screenWidth = window.innerWidth - scrollbarWidth;
634
+ const screenHeight = window.innerHeight;
635
+
636
+ // Get element's coordinates relative to the page
637
+ const elementRect = targetElement.getBoundingClientRect();
638
+ const elementY = elementRect.y;
639
+ const elementX = elementRect.x;
640
+
641
+ // Calculate scale and offsets
642
+ const scaleX = screenWidth / elementWidth;
643
+ const scaleY = screenHeight / elementHeight;
644
+ const scale = Math.min(scaleX, scaleY);
645
+
646
+ // Get the current transformOrigin
647
+ const computedStyle = window.getComputedStyle(targetElement);
648
+ const transformOrigin = computedStyle.transformOrigin;
649
+ const [originX, originY] = transformOrigin.split(" ");
650
+ const originXValue = parseFloat(originX);
651
+ const originYValue = parseFloat(originY);
652
+
653
+ // Calculate offsets with respect to the transformOrigin
654
+ const offsetX =
655
+ (screenWidth - elementWidth * scale) / 2 -
656
+ elementX -
657
+ originXValue * (1 - scale);
658
+ const offsetY =
659
+ (screenHeight - elementHeight * scale) / 2 -
660
+ elementY -
661
+ originYValue * (1 - scale);
662
+
663
+ // Apply scale and offsets to the element
664
+ targetElement.style.transform = `translate(${offsetX}px, ${offsetY}px) scale(${scale})`;
665
+
666
+ // Update global variables
667
+ elemData[elemId].zoomLevel = scale;
668
+ elemData[elemId].panX = offsetX;
669
+ elemData[elemId].panY = offsetY;
670
+
671
+ fullScreenMode = true;
672
+ toggleOverlap("on");
673
+ }
674
+
675
+ // Handle keydown events
676
+ function handleKeyDown(event) {
677
+ // Disable key locks to make pasting from the buffer work correctly
678
+ if ((event.ctrlKey && event.code === 'KeyV') || (event.ctrlKey && event.code === 'KeyC') || event.code === "F5") {
679
+ return;
680
+ }
681
+
682
+ // before activating shortcut, ensure user is not actively typing in an input field
683
+ if (!hotkeysConfig.canvas_blur_prompt) {
684
+ if (event.target.nodeName === 'TEXTAREA' || event.target.nodeName === 'INPUT') {
685
+ return;
686
+ }
687
+ }
688
+
689
+
690
+ const hotkeyActions = {
691
+ [hotkeysConfig.canvas_hotkey_reset]: resetZoom,
692
+ [hotkeysConfig.canvas_hotkey_overlap]: toggleOverlap,
693
+ [hotkeysConfig.canvas_hotkey_fullscreen]: fitToScreen,
694
+ [hotkeysConfig.canvas_hotkey_shrink_brush]: () => adjustBrushSize(elemId, 10),
695
+ [hotkeysConfig.canvas_hotkey_grow_brush]: () => adjustBrushSize(elemId, -10)
696
+ };
697
+
698
+ const action = hotkeyActions[event.code];
699
+ if (action) {
700
+ event.preventDefault();
701
+ action(event);
702
+ }
703
+
704
+ if (
705
+ isModifierKey(event, hotkeysConfig.canvas_hotkey_zoom) ||
706
+ isModifierKey(event, hotkeysConfig.canvas_hotkey_adjust)
707
+ ) {
708
+ event.preventDefault();
709
+ }
710
+ }
711
+
712
+ // Get Mouse position
713
+ function getMousePosition(e) {
714
+ mouseX = e.offsetX;
715
+ mouseY = e.offsetY;
716
+ }
717
+
718
+ // Simulation of the function to put a long image into the screen.
719
+ // We detect if an image has a scroll bar or not, make a fullscreen to reveal the image, then reduce it to fit into the element.
720
+ // We hide the image and show it to the user when it is ready.
721
+
722
+ targetElement.isExpanded = false;
723
+ function autoExpand() {
724
+ const canvas = document.querySelector(`${elemId} canvas[key="interface"]`);
725
+ if (canvas) {
726
+ if (hasHorizontalScrollbar(targetElement) && targetElement.isExpanded === false) {
727
+ targetElement.style.visibility = "hidden";
728
+ setTimeout(() => {
729
+ fitToScreen();
730
+ resetZoom();
731
+ targetElement.style.visibility = "visible";
732
+ targetElement.isExpanded = true;
733
+ }, 10);
734
+ }
735
+ }
736
+ }
737
+
738
+ targetElement.addEventListener("mousemove", getMousePosition);
739
+
740
+ //observers
741
+ // Creating an observer with a callback function to handle DOM changes
742
+ const observer = new MutationObserver((mutationsList, observer) => {
743
+ for (let mutation of mutationsList) {
744
+ // If the style attribute of the canvas has changed, by observation it happens only when the picture changes
745
+ if (mutation.type === 'attributes' && mutation.attributeName === 'style' &&
746
+ mutation.target.tagName.toLowerCase() === 'canvas') {
747
+ targetElement.isExpanded = false;
748
+ setTimeout(resetZoom, 10);
749
+ }
750
+ }
751
+ });
752
+
753
+ // Apply auto expand if enabled
754
+ if (hotkeysConfig.canvas_auto_expand) {
755
+ targetElement.addEventListener("mousemove", autoExpand);
756
+ // Set up an observer to track attribute changes
757
+ observer.observe(targetElement, {attributes: true, childList: true, subtree: true});
758
+ }
759
+
760
+ // Handle events only inside the targetElement
761
+ let isKeyDownHandlerAttached = false;
762
+
763
+ function handleMouseMove() {
764
+ if (!isKeyDownHandlerAttached) {
765
+ document.addEventListener("keydown", handleKeyDown);
766
+ isKeyDownHandlerAttached = true;
767
+
768
+ activeElement = elemId;
769
+ }
770
+ }
771
+
772
+ function handleMouseLeave() {
773
+ if (isKeyDownHandlerAttached) {
774
+ document.removeEventListener("keydown", handleKeyDown);
775
+ isKeyDownHandlerAttached = false;
776
+
777
+ activeElement = null;
778
+ }
779
+ }
780
+
781
+ // Add mouse event handlers
782
+ targetElement.addEventListener("mousemove", handleMouseMove);
783
+ targetElement.addEventListener("mouseleave", handleMouseLeave);
784
+
785
+ // Reset zoom when click on another tab
786
+ elements.img2imgTabs.addEventListener("click", resetZoom);
787
+ elements.img2imgTabs.addEventListener("click", () => {
788
+ // targetElement.style.width = "";
789
+ if (parseInt(targetElement.style.width) > 865) {
790
+ setTimeout(fitToElement, 0);
791
+ }
792
+ });
793
+
794
+ targetElement.addEventListener("wheel", e => {
795
+ // change zoom level
796
+ const operation = e.deltaY > 0 ? "-" : "+";
797
+ changeZoomLevel(operation, e);
798
+
799
+ // Handle brush size adjustment with ctrl key pressed
800
+ if (isModifierKey(e, hotkeysConfig.canvas_hotkey_adjust)) {
801
+ e.preventDefault();
802
+
803
+ // Increase or decrease brush size based on scroll direction
804
+ adjustBrushSize(elemId, e.deltaY);
805
+ }
806
+ });
807
+
808
+ // Handle the move event for pan functionality. Updates the panX and panY variables and applies the new transform to the target element.
809
+ function handleMoveKeyDown(e) {
810
+
811
+ // Disable key locks to make pasting from the buffer work correctly
812
+ if ((e.ctrlKey && e.code === 'KeyV') || (e.ctrlKey && event.code === 'KeyC') || e.code === "F5") {
813
+ return;
814
+ }
815
+
816
+ // before activating shortcut, ensure user is not actively typing in an input field
817
+ if (!hotkeysConfig.canvas_blur_prompt) {
818
+ if (e.target.nodeName === 'TEXTAREA' || e.target.nodeName === 'INPUT') {
819
+ return;
820
+ }
821
+ }
822
+
823
+
824
+ if (e.code === hotkeysConfig.canvas_hotkey_move) {
825
+ if (!e.ctrlKey && !e.metaKey && isKeyDownHandlerAttached) {
826
+ e.preventDefault();
827
+ document.activeElement.blur();
828
+ isMoving = true;
829
+ }
830
+ }
831
+ }
832
+
833
+ function handleMoveKeyUp(e) {
834
+ if (e.code === hotkeysConfig.canvas_hotkey_move) {
835
+ isMoving = false;
836
+ }
837
+ }
838
+
839
+ document.addEventListener("keydown", handleMoveKeyDown);
840
+ document.addEventListener("keyup", handleMoveKeyUp);
841
+
842
+ // Detect zoom level and update the pan speed.
843
+ function updatePanPosition(movementX, movementY) {
844
+ let panSpeed = 2;
845
+
846
+ if (elemData[elemId].zoomLevel > 8) {
847
+ panSpeed = 3.5;
848
+ }
849
+
850
+ elemData[elemId].panX += movementX * panSpeed;
851
+ elemData[elemId].panY += movementY * panSpeed;
852
+
853
+ // Delayed redraw of an element
854
+ requestAnimationFrame(() => {
855
+ targetElement.style.transform = `translate(${elemData[elemId].panX}px, ${elemData[elemId].panY}px) scale(${elemData[elemId].zoomLevel})`;
856
+ toggleOverlap("on");
857
+ });
858
+ }
859
+
860
+ function handleMoveByKey(e) {
861
+ if (isMoving && elemId === activeElement) {
862
+ updatePanPosition(e.movementX, e.movementY);
863
+ targetElement.style.pointerEvents = "none";
864
+
865
+ if (isExtension) {
866
+ targetElement.style.overflow = "visible";
867
+ }
868
+
869
+ } else {
870
+ targetElement.style.pointerEvents = "auto";
871
+ }
872
+ }
873
+
874
+ // Prevents sticking to the mouse
875
+ window.onblur = function() {
876
+ isMoving = false;
877
+ };
878
+
879
+ // Checks for extension
880
+ function checkForOutBox() {
881
+ const parentElement = targetElement.closest('[id^="component-"]');
882
+ if (parentElement.offsetWidth < targetElement.offsetWidth && !targetElement.isExpanded) {
883
+ resetZoom();
884
+ targetElement.isExpanded = true;
885
+ }
886
+
887
+ if (parentElement.offsetWidth < targetElement.offsetWidth && elemData[elemId].zoomLevel == 1) {
888
+ resetZoom();
889
+ }
890
+
891
+ if (parentElement.offsetWidth < targetElement.offsetWidth && targetElement.offsetWidth * elemData[elemId].zoomLevel > parentElement.offsetWidth && elemData[elemId].zoomLevel < 1 && !targetElement.isZoomed) {
892
+ resetZoom();
893
+ }
894
+ }
895
+
896
+ if (isExtension) {
897
+ targetElement.addEventListener("mousemove", checkForOutBox);
898
+ }
899
+
900
+
901
+ window.addEventListener('resize', (e) => {
902
+ resetZoom();
903
+
904
+ if (isExtension) {
905
+ targetElement.isExpanded = false;
906
+ targetElement.isZoomed = false;
907
+ }
908
+ });
909
+
910
+ gradioApp().addEventListener("mousemove", handleMoveByKey);
911
+
912
+
913
+ }
914
+
915
+ applyZoomAndPan(elementIDs.sketch, false);
916
+ applyZoomAndPan(elementIDs.inpaint, false);
917
+ applyZoomAndPan(elementIDs.inpaintSketch, false);
918
+
919
+ // Make the function global so that other extensions can take advantage of this solution
920
+ const applyZoomAndPanIntegration = async(id, elementIDs) => {
921
+ const mainEl = document.querySelector(id);
922
+ if (id.toLocaleLowerCase() === "none") {
923
+ for (const elementID of elementIDs) {
924
+ const el = await waitForElement(elementID);
925
+ if (!el) break;
926
+ applyZoomAndPan(elementID);
927
+ }
928
+ return;
929
+ }
930
+
931
+ if (!mainEl) return;
932
+ mainEl.addEventListener("click", async() => {
933
+ for (const elementID of elementIDs) {
934
+ const el = await waitForElement(elementID);
935
+ if (!el) break;
936
+ applyZoomAndPan(elementID);
937
+ }
938
+ }, {once: true});
939
+ };
940
+
941
+ window.applyZoomAndPan = applyZoomAndPan; // Only 1 elements, argument elementID, for example applyZoomAndPan("#txt2img_controlnet_ControlNet_input_image")
942
+
943
+ window.applyZoomAndPanIntegration = applyZoomAndPanIntegration; // for any extension
944
+
945
+ /*
946
+ The function `applyZoomAndPanIntegration` takes two arguments:
947
+
948
+ 1. `id`: A string identifier for the element to which zoom and pan functionality will be applied on click.
949
+ If the `id` value is "none", the functionality will be applied to all elements specified in the second argument without a click event.
950
+
951
+ 2. `elementIDs`: An array of string identifiers for elements. Zoom and pan functionality will be applied to each of these elements on click of the element specified by the first argument.
952
+ If "none" is specified in the first argument, the functionality will be applied to each of these elements without a click event.
953
+
954
+ Example usage:
955
+ applyZoomAndPanIntegration("#txt2img_controlnet", ["#txt2img_controlnet_ControlNet_input_image"]);
956
+ In this example, zoom and pan functionality will be applied to the element with the identifier "txt2img_controlnet_ControlNet_input_image" upon clicking the element with the identifier "txt2img_controlnet".
957
+ */
958
+
959
+ // More examples
960
+ // Add integration with ControlNet txt2img One TAB
961
+ // applyZoomAndPanIntegration("#txt2img_controlnet", ["#txt2img_controlnet_ControlNet_input_image"]);
962
+
963
+ // Add integration with ControlNet txt2img Tabs
964
+ // applyZoomAndPanIntegration("#txt2img_controlnet",Array.from({ length: 10 }, (_, i) => `#txt2img_controlnet_ControlNet-${i}_input_image`));
965
+
966
+ // Add integration with Inpaint Anything
967
+ // applyZoomAndPanIntegration("None", ["#ia_sam_image", "#ia_sel_mask"]);
968
+ });
extensions-builtin/canvas-zoom-and-pan/scripts/hotkey_config.py ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ from modules import shared
3
+
4
+ shared.options_templates.update(shared.options_section(('canvas_hotkey', "Canvas Hotkeys"), {
5
+ "canvas_hotkey_zoom": shared.OptionInfo("Alt", "Zoom canvas", gr.Radio, {"choices": ["Shift","Ctrl", "Alt"]}).info("If you choose 'Shift' you cannot scroll horizontally, 'Alt' can cause a little trouble in firefox"),
6
+ "canvas_hotkey_adjust": shared.OptionInfo("Ctrl", "Adjust brush size", gr.Radio, {"choices": ["Shift","Ctrl", "Alt"]}).info("If you choose 'Shift' you cannot scroll horizontally, 'Alt' can cause a little trouble in firefox"),
7
+ "canvas_hotkey_shrink_brush": shared.OptionInfo("Q", "Shrink the brush size"),
8
+ "canvas_hotkey_grow_brush": shared.OptionInfo("W", "Enlarge the brush size"),
9
+ "canvas_hotkey_move": shared.OptionInfo("F", "Moving the canvas").info("To work correctly in firefox, turn off 'Automatically search the page text when typing' in the browser settings"),
10
+ "canvas_hotkey_fullscreen": shared.OptionInfo("S", "Fullscreen Mode, maximizes the picture so that it fits into the screen and stretches it to its full width "),
11
+ "canvas_hotkey_reset": shared.OptionInfo("R", "Reset zoom and canvas positon"),
12
+ "canvas_hotkey_overlap": shared.OptionInfo("O", "Toggle overlap").info("Technical button, neededs for testing"),
13
+ "canvas_show_tooltip": shared.OptionInfo(True, "Enable tooltip on the canvas"),
14
+ "canvas_auto_expand": shared.OptionInfo(True, "Automatically expands an image that does not fit completely in the canvas area, similar to manually pressing the S and R buttons"),
15
+ "canvas_blur_prompt": shared.OptionInfo(False, "Take the focus off the prompt when working with a canvas"),
16
+ "canvas_disabled_functions": shared.OptionInfo(["Overlap"], "Disable function that you don't use", gr.CheckboxGroup, {"choices": ["Zoom","Adjust brush size","Hotkey enlarge brush","Hotkey shrink brush","Moving canvas","Fullscreen","Reset Zoom","Overlap"]}),
17
+ }))
extensions-builtin/canvas-zoom-and-pan/style.css ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ .canvas-tooltip-info {
2
+ position: absolute;
3
+ top: 10px;
4
+ left: 10px;
5
+ cursor: help;
6
+ background-color: rgba(0, 0, 0, 0.3);
7
+ width: 20px;
8
+ height: 20px;
9
+ border-radius: 50%;
10
+ display: flex;
11
+ align-items: center;
12
+ justify-content: center;
13
+ flex-direction: column;
14
+
15
+ z-index: 100;
16
+ }
17
+
18
+ .canvas-tooltip-info::after {
19
+ content: '';
20
+ display: block;
21
+ width: 2px;
22
+ height: 7px;
23
+ background-color: white;
24
+ margin-top: 2px;
25
+ }
26
+
27
+ .canvas-tooltip-info::before {
28
+ content: '';
29
+ display: block;
30
+ width: 2px;
31
+ height: 2px;
32
+ background-color: white;
33
+ }
34
+
35
+ .canvas-tooltip-content {
36
+ display: none;
37
+ background-color: #f9f9f9;
38
+ color: #333;
39
+ border: 1px solid #ddd;
40
+ padding: 15px;
41
+ position: absolute;
42
+ top: 40px;
43
+ left: 10px;
44
+ width: 250px;
45
+ font-size: 16px;
46
+ opacity: 0;
47
+ border-radius: 8px;
48
+ box-shadow: 0px 8px 16px 0px rgba(0,0,0,0.2);
49
+
50
+ z-index: 100;
51
+ }
52
+
53
+ .canvas-tooltip:hover .canvas-tooltip-content {
54
+ display: block;
55
+ animation: fadeIn 0.5s;
56
+ opacity: 1;
57
+ }
58
+
59
+ @keyframes fadeIn {
60
+ from {opacity: 0;}
61
+ to {opacity: 1;}
62
+ }
63
+
64
+ .styler {
65
+ overflow:inherit !important;
66
+ }
extensions-builtin/extra-options-section/scripts/extra_options_section.py ADDED
@@ -0,0 +1,78 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+
3
+ import gradio as gr
4
+ from modules import scripts, shared, ui_components, ui_settings, infotext_utils
5
+ from modules.ui_components import FormColumn
6
+
7
+
8
+ class ExtraOptionsSection(scripts.Script):
9
+ section = "extra_options"
10
+
11
+ def __init__(self):
12
+ self.comps = None
13
+ self.setting_names = None
14
+
15
+ def title(self):
16
+ return "Extra options"
17
+
18
+ def show(self, is_img2img):
19
+ return scripts.AlwaysVisible
20
+
21
+ def ui(self, is_img2img):
22
+ self.comps = []
23
+ self.setting_names = []
24
+ self.infotext_fields = []
25
+ extra_options = shared.opts.extra_options_img2img if is_img2img else shared.opts.extra_options_txt2img
26
+ elem_id_tabname = "extra_options_" + ("img2img" if is_img2img else "txt2img")
27
+
28
+ mapping = {k: v for v, k in infotext_utils.infotext_to_setting_name_mapping}
29
+
30
+ with gr.Blocks() as interface:
31
+ with gr.Accordion("Options", open=False, elem_id=elem_id_tabname) if shared.opts.extra_options_accordion and extra_options else gr.Group(elem_id=elem_id_tabname):
32
+
33
+ row_count = math.ceil(len(extra_options) / shared.opts.extra_options_cols)
34
+
35
+ for row in range(row_count):
36
+ with gr.Row():
37
+ for col in range(shared.opts.extra_options_cols):
38
+ index = row * shared.opts.extra_options_cols + col
39
+ if index >= len(extra_options):
40
+ break
41
+
42
+ setting_name = extra_options[index]
43
+
44
+ with FormColumn():
45
+ comp = ui_settings.create_setting_component(setting_name)
46
+
47
+ self.comps.append(comp)
48
+ self.setting_names.append(setting_name)
49
+
50
+ setting_infotext_name = mapping.get(setting_name)
51
+ if setting_infotext_name is not None:
52
+ self.infotext_fields.append((comp, setting_infotext_name))
53
+
54
+ def get_settings_values():
55
+ res = [ui_settings.get_value_for_setting(key) for key in self.setting_names]
56
+ return res[0] if len(res) == 1 else res
57
+
58
+ interface.load(fn=get_settings_values, inputs=[], outputs=self.comps, queue=False, show_progress=False)
59
+
60
+ return self.comps
61
+
62
+ def before_process(self, p, *args):
63
+ for name, value in zip(self.setting_names, args):
64
+ if name not in p.override_settings:
65
+ p.override_settings[name] = value
66
+
67
+
68
+ shared.options_templates.update(shared.options_section(('settings_in_ui', "Settings in UI", "ui"), {
69
+ "settings_in_ui": shared.OptionHTML("""
70
+ This page allows you to add some settings to the main interface of txt2img and img2img tabs.
71
+ """),
72
+ "extra_options_txt2img": shared.OptionInfo([], "Settings for txt2img", ui_components.DropdownMulti, lambda: {"choices": list(shared.opts.data_labels.keys())}).js("info", "settingsHintsShowQuicksettings").info("setting entries that also appear in txt2img interfaces").needs_reload_ui(),
73
+ "extra_options_img2img": shared.OptionInfo([], "Settings for img2img", ui_components.DropdownMulti, lambda: {"choices": list(shared.opts.data_labels.keys())}).js("info", "settingsHintsShowQuicksettings").info("setting entries that also appear in img2img interfaces").needs_reload_ui(),
74
+ "extra_options_cols": shared.OptionInfo(1, "Number of columns for added settings", gr.Slider, {"step": 1, "minimum": 1, "maximum": 20}).info("displayed amount will depend on the actual browser window width").needs_reload_ui(),
75
+ "extra_options_accordion": shared.OptionInfo(False, "Place added settings into an accordion").needs_reload_ui()
76
+ }))
77
+
78
+
extensions-builtin/forge_legacy_preprocessors/.gitignore ADDED
@@ -0,0 +1,185 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Byte-compiled / optimized / DLL files
2
+ __pycache__/
3
+ *.py[cod]
4
+ *$py.class
5
+
6
+ # C extensions
7
+ *.so
8
+
9
+ # Distribution / packaging
10
+ .Python
11
+ build/
12
+ develop-eggs/
13
+ dist/
14
+ downloads/
15
+ eggs/
16
+ .eggs/
17
+ lib/
18
+ lib64/
19
+ parts/
20
+ sdist/
21
+ var/
22
+ wheels/
23
+ share/python-wheels/
24
+ *.egg-info/
25
+ .installed.cfg
26
+ *.egg
27
+ MANIFEST
28
+
29
+ # PyInstaller
30
+ # Usually these files are written by a python script from a template
31
+ # before PyInstaller builds the exe, so as to inject date/other infos into it.
32
+ *.manifest
33
+ *.spec
34
+
35
+ # Installer logs
36
+ pip-log.txt
37
+ pip-delete-this-directory.txt
38
+
39
+ # Unit test / coverage reports
40
+ htmlcov/
41
+ .tox/
42
+ .nox/
43
+ .coverage
44
+ .coverage.*
45
+ .cache
46
+ nosetests.xml
47
+ coverage.xml
48
+ *.cover
49
+ *.py,cover
50
+ .hypothesis/
51
+ .pytest_cache/
52
+ cover/
53
+
54
+ # Translations
55
+ *.mo
56
+ *.pot
57
+
58
+ # Django stuff:
59
+ *.log
60
+ local_settings.py
61
+ db.sqlite3
62
+ db.sqlite3-journal
63
+
64
+ # Flask stuff:
65
+ instance/
66
+ .webassets-cache
67
+
68
+ # Scrapy stuff:
69
+ .scrapy
70
+
71
+ # Sphinx documentation
72
+ docs/_build/
73
+
74
+ # PyBuilder
75
+ .pybuilder/
76
+ target/
77
+
78
+ # Jupyter Notebook
79
+ .ipynb_checkpoints
80
+
81
+ # IPython
82
+ profile_default/
83
+ ipython_config.py
84
+
85
+ # pyenv
86
+ # For a library or package, you might want to ignore these files since the code is
87
+ # intended to run in multiple environments; otherwise, check them in:
88
+ # .python-version
89
+
90
+ # pipenv
91
+ # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
92
+ # However, in case of collaboration, if having platform-specific dependencies or dependencies
93
+ # having no cross-platform support, pipenv may install dependencies that don't work, or not
94
+ # install all needed dependencies.
95
+ #Pipfile.lock
96
+
97
+ # poetry
98
+ # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
99
+ # This is especially recommended for binary packages to ensure reproducibility, and is more
100
+ # commonly ignored for libraries.
101
+ # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
102
+ #poetry.lock
103
+
104
+ # pdm
105
+ # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
106
+ #pdm.lock
107
+ # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
108
+ # in version control.
109
+ # https://pdm.fming.dev/#use-with-ide
110
+ .pdm.toml
111
+
112
+ # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
113
+ __pypackages__/
114
+
115
+ # Celery stuff
116
+ celerybeat-schedule
117
+ celerybeat.pid
118
+
119
+ # SageMath parsed files
120
+ *.sage.py
121
+
122
+ # Environments
123
+ .env
124
+ .venv
125
+ env/
126
+ venv/
127
+ ENV/
128
+ env.bak/
129
+ venv.bak/
130
+
131
+ # Spyder project settings
132
+ .spyderproject
133
+ .spyproject
134
+
135
+ # Rope project settings
136
+ .ropeproject
137
+
138
+ # mkdocs documentation
139
+ /site
140
+
141
+ # mypy
142
+ .mypy_cache/
143
+ .dmypy.json
144
+ dmypy.json
145
+
146
+ # Pyre type checker
147
+ .pyre/
148
+
149
+ # pytype static type analyzer
150
+ .pytype/
151
+
152
+ # Cython debug symbols
153
+ cython_debug/
154
+
155
+ # PyCharm
156
+ # JetBrains specific template is maintained in a separate JetBrains.gitignore that can
157
+ # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
158
+ # and can be added to the global gitignore or merged into this file. For a more nuclear
159
+ # option (not recommended) you can uncomment the following to ignore the entire idea folder.
160
+ #.idea
161
+ *.pt
162
+ *.pth
163
+ *.ckpt
164
+ *.bin
165
+ *.safetensors
166
+
167
+ # Editor setting metadata
168
+ .idea/
169
+ .vscode/
170
+ detected_maps/
171
+ annotator/downloads/
172
+
173
+ # test results and expectations
174
+ web_tests/results/
175
+ web_tests/expectations/
176
+ tests/web_api/full_coverage/results/
177
+ tests/web_api/full_coverage/expectations/
178
+
179
+ *_diff.png
180
+
181
+ # Presets
182
+ presets/
183
+
184
+ # Ignore existing dir of hand refiner if exists.
185
+ annotator/hand_refiner_portable
extensions-builtin/forge_legacy_preprocessors/LICENSE ADDED
@@ -0,0 +1,674 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ GNU GENERAL PUBLIC LICENSE
2
+ Version 3, 29 June 2007
3
+
4
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
5
+ Everyone is permitted to copy and distribute verbatim copies
6
+ of this license document, but changing it is not allowed.
7
+
8
+ Preamble
9
+
10
+ The GNU General Public License is a free, copyleft license for
11
+ software and other kinds of works.
12
+
13
+ The licenses for most software and other practical works are designed
14
+ to take away your freedom to share and change the works. By contrast,
15
+ the GNU General Public License is intended to guarantee your freedom to
16
+ share and change all versions of a program--to make sure it remains free
17
+ software for all its users. We, the Free Software Foundation, use the
18
+ GNU General Public License for most of our software; it applies also to
19
+ any other work released this way by its authors. You can apply it to
20
+ your programs, too.
21
+
22
+ When we speak of free software, we are referring to freedom, not
23
+ price. Our General Public Licenses are designed to make sure that you
24
+ have the freedom to distribute copies of free software (and charge for
25
+ them if you wish), that you receive source code or can get it if you
26
+ want it, that you can change the software or use pieces of it in new
27
+ free programs, and that you know you can do these things.
28
+
29
+ To protect your rights, we need to prevent others from denying you
30
+ these rights or asking you to surrender the rights. Therefore, you have
31
+ certain responsibilities if you distribute copies of the software, or if
32
+ you modify it: responsibilities to respect the freedom of others.
33
+
34
+ For example, if you distribute copies of such a program, whether
35
+ gratis or for a fee, you must pass on to the recipients the same
36
+ freedoms that you received. You must make sure that they, too, receive
37
+ or can get the source code. And you must show them these terms so they
38
+ know their rights.
39
+
40
+ Developers that use the GNU GPL protect your rights with two steps:
41
+ (1) assert copyright on the software, and (2) offer you this License
42
+ giving you legal permission to copy, distribute and/or modify it.
43
+
44
+ For the developers' and authors' protection, the GPL clearly explains
45
+ that there is no warranty for this free software. For both users' and
46
+ authors' sake, the GPL requires that modified versions be marked as
47
+ changed, so that their problems will not be attributed erroneously to
48
+ authors of previous versions.
49
+
50
+ Some devices are designed to deny users access to install or run
51
+ modified versions of the software inside them, although the manufacturer
52
+ can do so. This is fundamentally incompatible with the aim of
53
+ protecting users' freedom to change the software. The systematic
54
+ pattern of such abuse occurs in the area of products for individuals to
55
+ use, which is precisely where it is most unacceptable. Therefore, we
56
+ have designed this version of the GPL to prohibit the practice for those
57
+ products. If such problems arise substantially in other domains, we
58
+ stand ready to extend this provision to those domains in future versions
59
+ of the GPL, as needed to protect the freedom of users.
60
+
61
+ Finally, every program is threatened constantly by software patents.
62
+ States should not allow patents to restrict development and use of
63
+ software on general-purpose computers, but in those that do, we wish to
64
+ avoid the special danger that patents applied to a free program could
65
+ make it effectively proprietary. To prevent this, the GPL assures that
66
+ patents cannot be used to render the program non-free.
67
+
68
+ The precise terms and conditions for copying, distribution and
69
+ modification follow.
70
+
71
+ TERMS AND CONDITIONS
72
+
73
+ 0. Definitions.
74
+
75
+ "This License" refers to version 3 of the GNU General Public License.
76
+
77
+ "Copyright" also means copyright-like laws that apply to other kinds of
78
+ works, such as semiconductor masks.
79
+
80
+ "The Program" refers to any copyrightable work licensed under this
81
+ License. Each licensee is addressed as "you". "Licensees" and
82
+ "recipients" may be individuals or organizations.
83
+
84
+ To "modify" a work means to copy from or adapt all or part of the work
85
+ in a fashion requiring copyright permission, other than the making of an
86
+ exact copy. The resulting work is called a "modified version" of the
87
+ earlier work or a work "based on" the earlier work.
88
+
89
+ A "covered work" means either the unmodified Program or a work based
90
+ on the Program.
91
+
92
+ To "propagate" a work means to do anything with it that, without
93
+ permission, would make you directly or secondarily liable for
94
+ infringement under applicable copyright law, except executing it on a
95
+ computer or modifying a private copy. Propagation includes copying,
96
+ distribution (with or without modification), making available to the
97
+ public, and in some countries other activities as well.
98
+
99
+ To "convey" a work means any kind of propagation that enables other
100
+ parties to make or receive copies. Mere interaction with a user through
101
+ a computer network, with no transfer of a copy, is not conveying.
102
+
103
+ An interactive user interface displays "Appropriate Legal Notices"
104
+ to the extent that it includes a convenient and prominently visible
105
+ feature that (1) displays an appropriate copyright notice, and (2)
106
+ tells the user that there is no warranty for the work (except to the
107
+ extent that warranties are provided), that licensees may convey the
108
+ work under this License, and how to view a copy of this License. If
109
+ the interface presents a list of user commands or options, such as a
110
+ menu, a prominent item in the list meets this criterion.
111
+
112
+ 1. Source Code.
113
+
114
+ The "source code" for a work means the preferred form of the work
115
+ for making modifications to it. "Object code" means any non-source
116
+ form of a work.
117
+
118
+ A "Standard Interface" means an interface that either is an official
119
+ standard defined by a recognized standards body, or, in the case of
120
+ interfaces specified for a particular programming language, one that
121
+ is widely used among developers working in that language.
122
+
123
+ The "System Libraries" of an executable work include anything, other
124
+ than the work as a whole, that (a) is included in the normal form of
125
+ packaging a Major Component, but which is not part of that Major
126
+ Component, and (b) serves only to enable use of the work with that
127
+ Major Component, or to implement a Standard Interface for which an
128
+ implementation is available to the public in source code form. A
129
+ "Major Component", in this context, means a major essential component
130
+ (kernel, window system, and so on) of the specific operating system
131
+ (if any) on which the executable work runs, or a compiler used to
132
+ produce the work, or an object code interpreter used to run it.
133
+
134
+ The "Corresponding Source" for a work in object code form means all
135
+ the source code needed to generate, install, and (for an executable
136
+ work) run the object code and to modify the work, including scripts to
137
+ control those activities. However, it does not include the work's
138
+ System Libraries, or general-purpose tools or generally available free
139
+ programs which are used unmodified in performing those activities but
140
+ which are not part of the work. For example, Corresponding Source
141
+ includes interface definition files associated with source files for
142
+ the work, and the source code for shared libraries and dynamically
143
+ linked subprograms that the work is specifically designed to require,
144
+ such as by intimate data communication or control flow between those
145
+ subprograms and other parts of the work.
146
+
147
+ The Corresponding Source need not include anything that users
148
+ can regenerate automatically from other parts of the Corresponding
149
+ Source.
150
+
151
+ The Corresponding Source for a work in source code form is that
152
+ same work.
153
+
154
+ 2. Basic Permissions.
155
+
156
+ All rights granted under this License are granted for the term of
157
+ copyright on the Program, and are irrevocable provided the stated
158
+ conditions are met. This License explicitly affirms your unlimited
159
+ permission to run the unmodified Program. The output from running a
160
+ covered work is covered by this License only if the output, given its
161
+ content, constitutes a covered work. This License acknowledges your
162
+ rights of fair use or other equivalent, as provided by copyright law.
163
+
164
+ You may make, run and propagate covered works that you do not
165
+ convey, without conditions so long as your license otherwise remains
166
+ in force. You may convey covered works to others for the sole purpose
167
+ of having them make modifications exclusively for you, or provide you
168
+ with facilities for running those works, provided that you comply with
169
+ the terms of this License in conveying all material for which you do
170
+ not control copyright. Those thus making or running the covered works
171
+ for you must do so exclusively on your behalf, under your direction
172
+ and control, on terms that prohibit them from making any copies of
173
+ your copyrighted material outside their relationship with you.
174
+
175
+ Conveying under any other circumstances is permitted solely under
176
+ the conditions stated below. Sublicensing is not allowed; section 10
177
+ makes it unnecessary.
178
+
179
+ 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
180
+
181
+ No covered work shall be deemed part of an effective technological
182
+ measure under any applicable law fulfilling obligations under article
183
+ 11 of the WIPO copyright treaty adopted on 20 December 1996, or
184
+ similar laws prohibiting or restricting circumvention of such
185
+ measures.
186
+
187
+ When you convey a covered work, you waive any legal power to forbid
188
+ circumvention of technological measures to the extent such circumvention
189
+ is effected by exercising rights under this License with respect to
190
+ the covered work, and you disclaim any intention to limit operation or
191
+ modification of the work as a means of enforcing, against the work's
192
+ users, your or third parties' legal rights to forbid circumvention of
193
+ technological measures.
194
+
195
+ 4. Conveying Verbatim Copies.
196
+
197
+ You may convey verbatim copies of the Program's source code as you
198
+ receive it, in any medium, provided that you conspicuously and
199
+ appropriately publish on each copy an appropriate copyright notice;
200
+ keep intact all notices stating that this License and any
201
+ non-permissive terms added in accord with section 7 apply to the code;
202
+ keep intact all notices of the absence of any warranty; and give all
203
+ recipients a copy of this License along with the Program.
204
+
205
+ You may charge any price or no price for each copy that you convey,
206
+ and you may offer support or warranty protection for a fee.
207
+
208
+ 5. Conveying Modified Source Versions.
209
+
210
+ You may convey a work based on the Program, or the modifications to
211
+ produce it from the Program, in the form of source code under the
212
+ terms of section 4, provided that you also meet all of these conditions:
213
+
214
+ a) The work must carry prominent notices stating that you modified
215
+ it, and giving a relevant date.
216
+
217
+ b) The work must carry prominent notices stating that it is
218
+ released under this License and any conditions added under section
219
+ 7. This requirement modifies the requirement in section 4 to
220
+ "keep intact all notices".
221
+
222
+ c) You must license the entire work, as a whole, under this
223
+ License to anyone who comes into possession of a copy. This
224
+ License will therefore apply, along with any applicable section 7
225
+ additional terms, to the whole of the work, and all its parts,
226
+ regardless of how they are packaged. This License gives no
227
+ permission to license the work in any other way, but it does not
228
+ invalidate such permission if you have separately received it.
229
+
230
+ d) If the work has interactive user interfaces, each must display
231
+ Appropriate Legal Notices; however, if the Program has interactive
232
+ interfaces that do not display Appropriate Legal Notices, your
233
+ work need not make them do so.
234
+
235
+ A compilation of a covered work with other separate and independent
236
+ works, which are not by their nature extensions of the covered work,
237
+ and which are not combined with it such as to form a larger program,
238
+ in or on a volume of a storage or distribution medium, is called an
239
+ "aggregate" if the compilation and its resulting copyright are not
240
+ used to limit the access or legal rights of the compilation's users
241
+ beyond what the individual works permit. Inclusion of a covered work
242
+ in an aggregate does not cause this License to apply to the other
243
+ parts of the aggregate.
244
+
245
+ 6. Conveying Non-Source Forms.
246
+
247
+ You may convey a covered work in object code form under the terms
248
+ of sections 4 and 5, provided that you also convey the
249
+ machine-readable Corresponding Source under the terms of this License,
250
+ in one of these ways:
251
+
252
+ a) Convey the object code in, or embodied in, a physical product
253
+ (including a physical distribution medium), accompanied by the
254
+ Corresponding Source fixed on a durable physical medium
255
+ customarily used for software interchange.
256
+
257
+ b) Convey the object code in, or embodied in, a physical product
258
+ (including a physical distribution medium), accompanied by a
259
+ written offer, valid for at least three years and valid for as
260
+ long as you offer spare parts or customer support for that product
261
+ model, to give anyone who possesses the object code either (1) a
262
+ copy of the Corresponding Source for all the software in the
263
+ product that is covered by this License, on a durable physical
264
+ medium customarily used for software interchange, for a price no
265
+ more than your reasonable cost of physically performing this
266
+ conveying of source, or (2) access to copy the
267
+ Corresponding Source from a network server at no charge.
268
+
269
+ c) Convey individual copies of the object code with a copy of the
270
+ written offer to provide the Corresponding Source. This
271
+ alternative is allowed only occasionally and noncommercially, and
272
+ only if you received the object code with such an offer, in accord
273
+ with subsection 6b.
274
+
275
+ d) Convey the object code by offering access from a designated
276
+ place (gratis or for a charge), and offer equivalent access to the
277
+ Corresponding Source in the same way through the same place at no
278
+ further charge. You need not require recipients to copy the
279
+ Corresponding Source along with the object code. If the place to
280
+ copy the object code is a network server, the Corresponding Source
281
+ may be on a different server (operated by you or a third party)
282
+ that supports equivalent copying facilities, provided you maintain
283
+ clear directions next to the object code saying where to find the
284
+ Corresponding Source. Regardless of what server hosts the
285
+ Corresponding Source, you remain obligated to ensure that it is
286
+ available for as long as needed to satisfy these requirements.
287
+
288
+ e) Convey the object code using peer-to-peer transmission, provided
289
+ you inform other peers where the object code and Corresponding
290
+ Source of the work are being offered to the general public at no
291
+ charge under subsection 6d.
292
+
293
+ A separable portion of the object code, whose source code is excluded
294
+ from the Corresponding Source as a System Library, need not be
295
+ included in conveying the object code work.
296
+
297
+ A "User Product" is either (1) a "consumer product", which means any
298
+ tangible personal property which is normally used for personal, family,
299
+ or household purposes, or (2) anything designed or sold for incorporation
300
+ into a dwelling. In determining whether a product is a consumer product,
301
+ doubtful cases shall be resolved in favor of coverage. For a particular
302
+ product received by a particular user, "normally used" refers to a
303
+ typical or common use of that class of product, regardless of the status
304
+ of the particular user or of the way in which the particular user
305
+ actually uses, or expects or is expected to use, the product. A product
306
+ is a consumer product regardless of whether the product has substantial
307
+ commercial, industrial or non-consumer uses, unless such uses represent
308
+ the only significant mode of use of the product.
309
+
310
+ "Installation Information" for a User Product means any methods,
311
+ procedures, authorization keys, or other information required to install
312
+ and execute modified versions of a covered work in that User Product from
313
+ a modified version of its Corresponding Source. The information must
314
+ suffice to ensure that the continued functioning of the modified object
315
+ code is in no case prevented or interfered with solely because
316
+ modification has been made.
317
+
318
+ If you convey an object code work under this section in, or with, or
319
+ specifically for use in, a User Product, and the conveying occurs as
320
+ part of a transaction in which the right of possession and use of the
321
+ User Product is transferred to the recipient in perpetuity or for a
322
+ fixed term (regardless of how the transaction is characterized), the
323
+ Corresponding Source conveyed under this section must be accompanied
324
+ by the Installation Information. But this requirement does not apply
325
+ if neither you nor any third party retains the ability to install
326
+ modified object code on the User Product (for example, the work has
327
+ been installed in ROM).
328
+
329
+ The requirement to provide Installation Information does not include a
330
+ requirement to continue to provide support service, warranty, or updates
331
+ for a work that has been modified or installed by the recipient, or for
332
+ the User Product in which it has been modified or installed. Access to a
333
+ network may be denied when the modification itself materially and
334
+ adversely affects the operation of the network or violates the rules and
335
+ protocols for communication across the network.
336
+
337
+ Corresponding Source conveyed, and Installation Information provided,
338
+ in accord with this section must be in a format that is publicly
339
+ documented (and with an implementation available to the public in
340
+ source code form), and must require no special password or key for
341
+ unpacking, reading or copying.
342
+
343
+ 7. Additional Terms.
344
+
345
+ "Additional permissions" are terms that supplement the terms of this
346
+ License by making exceptions from one or more of its conditions.
347
+ Additional permissions that are applicable to the entire Program shall
348
+ be treated as though they were included in this License, to the extent
349
+ that they are valid under applicable law. If additional permissions
350
+ apply only to part of the Program, that part may be used separately
351
+ under those permissions, but the entire Program remains governed by
352
+ this License without regard to the additional permissions.
353
+
354
+ When you convey a copy of a covered work, you may at your option
355
+ remove any additional permissions from that copy, or from any part of
356
+ it. (Additional permissions may be written to require their own
357
+ removal in certain cases when you modify the work.) You may place
358
+ additional permissions on material, added by you to a covered work,
359
+ for which you have or can give appropriate copyright permission.
360
+
361
+ Notwithstanding any other provision of this License, for material you
362
+ add to a covered work, you may (if authorized by the copyright holders of
363
+ that material) supplement the terms of this License with terms:
364
+
365
+ a) Disclaiming warranty or limiting liability differently from the
366
+ terms of sections 15 and 16 of this License; or
367
+
368
+ b) Requiring preservation of specified reasonable legal notices or
369
+ author attributions in that material or in the Appropriate Legal
370
+ Notices displayed by works containing it; or
371
+
372
+ c) Prohibiting misrepresentation of the origin of that material, or
373
+ requiring that modified versions of such material be marked in
374
+ reasonable ways as different from the original version; or
375
+
376
+ d) Limiting the use for publicity purposes of names of licensors or
377
+ authors of the material; or
378
+
379
+ e) Declining to grant rights under trademark law for use of some
380
+ trade names, trademarks, or service marks; or
381
+
382
+ f) Requiring indemnification of licensors and authors of that
383
+ material by anyone who conveys the material (or modified versions of
384
+ it) with contractual assumptions of liability to the recipient, for
385
+ any liability that these contractual assumptions directly impose on
386
+ those licensors and authors.
387
+
388
+ All other non-permissive additional terms are considered "further
389
+ restrictions" within the meaning of section 10. If the Program as you
390
+ received it, or any part of it, contains a notice stating that it is
391
+ governed by this License along with a term that is a further
392
+ restriction, you may remove that term. If a license document contains
393
+ a further restriction but permits relicensing or conveying under this
394
+ License, you may add to a covered work material governed by the terms
395
+ of that license document, provided that the further restriction does
396
+ not survive such relicensing or conveying.
397
+
398
+ If you add terms to a covered work in accord with this section, you
399
+ must place, in the relevant source files, a statement of the
400
+ additional terms that apply to those files, or a notice indicating
401
+ where to find the applicable terms.
402
+
403
+ Additional terms, permissive or non-permissive, may be stated in the
404
+ form of a separately written license, or stated as exceptions;
405
+ the above requirements apply either way.
406
+
407
+ 8. Termination.
408
+
409
+ You may not propagate or modify a covered work except as expressly
410
+ provided under this License. Any attempt otherwise to propagate or
411
+ modify it is void, and will automatically terminate your rights under
412
+ this License (including any patent licenses granted under the third
413
+ paragraph of section 11).
414
+
415
+ However, if you cease all violation of this License, then your
416
+ license from a particular copyright holder is reinstated (a)
417
+ provisionally, unless and until the copyright holder explicitly and
418
+ finally terminates your license, and (b) permanently, if the copyright
419
+ holder fails to notify you of the violation by some reasonable means
420
+ prior to 60 days after the cessation.
421
+
422
+ Moreover, your license from a particular copyright holder is
423
+ reinstated permanently if the copyright holder notifies you of the
424
+ violation by some reasonable means, this is the first time you have
425
+ received notice of violation of this License (for any work) from that
426
+ copyright holder, and you cure the violation prior to 30 days after
427
+ your receipt of the notice.
428
+
429
+ Termination of your rights under this section does not terminate the
430
+ licenses of parties who have received copies or rights from you under
431
+ this License. If your rights have been terminated and not permanently
432
+ reinstated, you do not qualify to receive new licenses for the same
433
+ material under section 10.
434
+
435
+ 9. Acceptance Not Required for Having Copies.
436
+
437
+ You are not required to accept this License in order to receive or
438
+ run a copy of the Program. Ancillary propagation of a covered work
439
+ occurring solely as a consequence of using peer-to-peer transmission
440
+ to receive a copy likewise does not require acceptance. However,
441
+ nothing other than this License grants you permission to propagate or
442
+ modify any covered work. These actions infringe copyright if you do
443
+ not accept this License. Therefore, by modifying or propagating a
444
+ covered work, you indicate your acceptance of this License to do so.
445
+
446
+ 10. Automatic Licensing of Downstream Recipients.
447
+
448
+ Each time you convey a covered work, the recipient automatically
449
+ receives a license from the original licensors, to run, modify and
450
+ propagate that work, subject to this License. You are not responsible
451
+ for enforcing compliance by third parties with this License.
452
+
453
+ An "entity transaction" is a transaction transferring control of an
454
+ organization, or substantially all assets of one, or subdividing an
455
+ organization, or merging organizations. If propagation of a covered
456
+ work results from an entity transaction, each party to that
457
+ transaction who receives a copy of the work also receives whatever
458
+ licenses to the work the party's predecessor in interest had or could
459
+ give under the previous paragraph, plus a right to possession of the
460
+ Corresponding Source of the work from the predecessor in interest, if
461
+ the predecessor has it or can get it with reasonable efforts.
462
+
463
+ You may not impose any further restrictions on the exercise of the
464
+ rights granted or affirmed under this License. For example, you may
465
+ not impose a license fee, royalty, or other charge for exercise of
466
+ rights granted under this License, and you may not initiate litigation
467
+ (including a cross-claim or counterclaim in a lawsuit) alleging that
468
+ any patent claim is infringed by making, using, selling, offering for
469
+ sale, or importing the Program or any portion of it.
470
+
471
+ 11. Patents.
472
+
473
+ A "contributor" is a copyright holder who authorizes use under this
474
+ License of the Program or a work on which the Program is based. The
475
+ work thus licensed is called the contributor's "contributor version".
476
+
477
+ A contributor's "essential patent claims" are all patent claims
478
+ owned or controlled by the contributor, whether already acquired or
479
+ hereafter acquired, that would be infringed by some manner, permitted
480
+ by this License, of making, using, or selling its contributor version,
481
+ but do not include claims that would be infringed only as a
482
+ consequence of further modification of the contributor version. For
483
+ purposes of this definition, "control" includes the right to grant
484
+ patent sublicenses in a manner consistent with the requirements of
485
+ this License.
486
+
487
+ Each contributor grants you a non-exclusive, worldwide, royalty-free
488
+ patent license under the contributor's essential patent claims, to
489
+ make, use, sell, offer for sale, import and otherwise run, modify and
490
+ propagate the contents of its contributor version.
491
+
492
+ In the following three paragraphs, a "patent license" is any express
493
+ agreement or commitment, however denominated, not to enforce a patent
494
+ (such as an express permission to practice a patent or covenant not to
495
+ sue for patent infringement). To "grant" such a patent license to a
496
+ party means to make such an agreement or commitment not to enforce a
497
+ patent against the party.
498
+
499
+ If you convey a covered work, knowingly relying on a patent license,
500
+ and the Corresponding Source of the work is not available for anyone
501
+ to copy, free of charge and under the terms of this License, through a
502
+ publicly available network server or other readily accessible means,
503
+ then you must either (1) cause the Corresponding Source to be so
504
+ available, or (2) arrange to deprive yourself of the benefit of the
505
+ patent license for this particular work, or (3) arrange, in a manner
506
+ consistent with the requirements of this License, to extend the patent
507
+ license to downstream recipients. "Knowingly relying" means you have
508
+ actual knowledge that, but for the patent license, your conveying the
509
+ covered work in a country, or your recipient's use of the covered work
510
+ in a country, would infringe one or more identifiable patents in that
511
+ country that you have reason to believe are valid.
512
+
513
+ If, pursuant to or in connection with a single transaction or
514
+ arrangement, you convey, or propagate by procuring conveyance of, a
515
+ covered work, and grant a patent license to some of the parties
516
+ receiving the covered work authorizing them to use, propagate, modify
517
+ or convey a specific copy of the covered work, then the patent license
518
+ you grant is automatically extended to all recipients of the covered
519
+ work and works based on it.
520
+
521
+ A patent license is "discriminatory" if it does not include within
522
+ the scope of its coverage, prohibits the exercise of, or is
523
+ conditioned on the non-exercise of one or more of the rights that are
524
+ specifically granted under this License. You may not convey a covered
525
+ work if you are a party to an arrangement with a third party that is
526
+ in the business of distributing software, under which you make payment
527
+ to the third party based on the extent of your activity of conveying
528
+ the work, and under which the third party grants, to any of the
529
+ parties who would receive the covered work from you, a discriminatory
530
+ patent license (a) in connection with copies of the covered work
531
+ conveyed by you (or copies made from those copies), or (b) primarily
532
+ for and in connection with specific products or compilations that
533
+ contain the covered work, unless you entered into that arrangement,
534
+ or that patent license was granted, prior to 28 March 2007.
535
+
536
+ Nothing in this License shall be construed as excluding or limiting
537
+ any implied license or other defenses to infringement that may
538
+ otherwise be available to you under applicable patent law.
539
+
540
+ 12. No Surrender of Others' Freedom.
541
+
542
+ If conditions are imposed on you (whether by court order, agreement or
543
+ otherwise) that contradict the conditions of this License, they do not
544
+ excuse you from the conditions of this License. If you cannot convey a
545
+ covered work so as to satisfy simultaneously your obligations under this
546
+ License and any other pertinent obligations, then as a consequence you may
547
+ not convey it at all. For example, if you agree to terms that obligate you
548
+ to collect a royalty for further conveying from those to whom you convey
549
+ the Program, the only way you could satisfy both those terms and this
550
+ License would be to refrain entirely from conveying the Program.
551
+
552
+ 13. Use with the GNU Affero General Public License.
553
+
554
+ Notwithstanding any other provision of this License, you have
555
+ permission to link or combine any covered work with a work licensed
556
+ under version 3 of the GNU Affero General Public License into a single
557
+ combined work, and to convey the resulting work. The terms of this
558
+ License will continue to apply to the part which is the covered work,
559
+ but the special requirements of the GNU Affero General Public License,
560
+ section 13, concerning interaction through a network will apply to the
561
+ combination as such.
562
+
563
+ 14. Revised Versions of this License.
564
+
565
+ The Free Software Foundation may publish revised and/or new versions of
566
+ the GNU General Public License from time to time. Such new versions will
567
+ be similar in spirit to the present version, but may differ in detail to
568
+ address new problems or concerns.
569
+
570
+ Each version is given a distinguishing version number. If the
571
+ Program specifies that a certain numbered version of the GNU General
572
+ Public License "or any later version" applies to it, you have the
573
+ option of following the terms and conditions either of that numbered
574
+ version or of any later version published by the Free Software
575
+ Foundation. If the Program does not specify a version number of the
576
+ GNU General Public License, you may choose any version ever published
577
+ by the Free Software Foundation.
578
+
579
+ If the Program specifies that a proxy can decide which future
580
+ versions of the GNU General Public License can be used, that proxy's
581
+ public statement of acceptance of a version permanently authorizes you
582
+ to choose that version for the Program.
583
+
584
+ Later license versions may give you additional or different
585
+ permissions. However, no additional obligations are imposed on any
586
+ author or copyright holder as a result of your choosing to follow a
587
+ later version.
588
+
589
+ 15. Disclaimer of Warranty.
590
+
591
+ THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
592
+ APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
593
+ HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
594
+ OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
595
+ THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
596
+ PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
597
+ IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
598
+ ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
599
+
600
+ 16. Limitation of Liability.
601
+
602
+ IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
603
+ WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
604
+ THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
605
+ GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
606
+ USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
607
+ DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
608
+ PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
609
+ EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
610
+ SUCH DAMAGES.
611
+
612
+ 17. Interpretation of Sections 15 and 16.
613
+
614
+ If the disclaimer of warranty and limitation of liability provided
615
+ above cannot be given local legal effect according to their terms,
616
+ reviewing courts shall apply local law that most closely approximates
617
+ an absolute waiver of all civil liability in connection with the
618
+ Program, unless a warranty or assumption of liability accompanies a
619
+ copy of the Program in return for a fee.
620
+
621
+ END OF TERMS AND CONDITIONS
622
+
623
+ How to Apply These Terms to Your New Programs
624
+
625
+ If you develop a new program, and you want it to be of the greatest
626
+ possible use to the public, the best way to achieve this is to make it
627
+ free software which everyone can redistribute and change under these terms.
628
+
629
+ To do so, attach the following notices to the program. It is safest
630
+ to attach them to the start of each source file to most effectively
631
+ state the exclusion of warranty; and each file should have at least
632
+ the "copyright" line and a pointer to where the full notice is found.
633
+
634
+ <one line to give the program's name and a brief idea of what it does.>
635
+ Copyright (C) <year> <name of author>
636
+
637
+ This program is free software: you can redistribute it and/or modify
638
+ it under the terms of the GNU General Public License as published by
639
+ the Free Software Foundation, either version 3 of the License, or
640
+ (at your option) any later version.
641
+
642
+ This program is distributed in the hope that it will be useful,
643
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
644
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
645
+ GNU General Public License for more details.
646
+
647
+ You should have received a copy of the GNU General Public License
648
+ along with this program. If not, see <https://www.gnu.org/licenses/>.
649
+
650
+ Also add information on how to contact you by electronic and paper mail.
651
+
652
+ If the program does terminal interaction, make it output a short
653
+ notice like this when it starts in an interactive mode:
654
+
655
+ <program> Copyright (C) <year> <name of author>
656
+ This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
657
+ This is free software, and you are welcome to redistribute it
658
+ under certain conditions; type `show c' for details.
659
+
660
+ The hypothetical commands `show w' and `show c' should show the appropriate
661
+ parts of the General Public License. Of course, your program's commands
662
+ might be different; for a GUI interface, you would use an "about box".
663
+
664
+ You should also get your employer (if you work as a programmer) or school,
665
+ if any, to sign a "copyright disclaimer" for the program, if necessary.
666
+ For more information on this, and how to apply and follow the GNU GPL, see
667
+ <https://www.gnu.org/licenses/>.
668
+
669
+ The GNU General Public License does not permit incorporating your program
670
+ into proprietary programs. If your program is a subroutine library, you
671
+ may consider it more useful to permit linking proprietary applications with
672
+ the library. If this is what you want to do, use the GNU Lesser General
673
+ Public License instead of this License. But first, please read
674
+ <https://www.gnu.org/licenses/why-not-lgpl.html>.