neopian-diffusion / README.md
crumb's picture
Update README.md
88e3f09
|
raw
history blame
3.36 kB
---
language:
- en
tags:
- stable-diffusion
- text-to-image
license: creativeml-openrail-m
---
# Neopian-Diffusion (wip, not done training the style isnt there yet)
Stable Diffusion models, starting with [runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5), trained on images extracted from gifs from https://www.neopets.com/funimages.phtml. CLIP ViT-B/32 (OpenAI) was used to filter the best matching frame of the GIF for every given caption/GIF pair. The frame with the minimum spherical distance was chosen and saved for training. In total this amounts to 1950 images around 100x100px. The DreamBooth models were finetuned on a Colab T4 with the term "low-resolution" concatenated onto prompts at varying weights, to hopefully combat artifacting in the final results (see this link for a hypothesis from someone on Discord about using negative terms while training Textual Inversions https://cdn.discordapp.com/attachments/1008246088148463648/1041538692432527470/image.png).
Example chosen frame of GIF from CLIP
| Caption | Unprocessed GIF | Chosen Frame |
| --- | --- | --- |
| "yurble_baby_clap" | ![](https://images.neopets.com/template_images/yurble_baby_clap.gif) | ![](https://cdn.discordapp.com/attachments/1010693530181718146/1043310485413576794/yurble_baby_clap.jpg) |
## Training Details
Stage 1 (0-8k steps) The text encoder was trained along with the UNet at half precision for 15% of the total 8,000 steps (1,200 steps), and then the UNet was trained alone for the rest. I used a polynomial learning rate decay starting at 2e-6 (the default in fast-DreamBooth). "low quality" concatenated onto 1/3 of the prompts. Trained at 448x448
Stage 2 (8k-16k) Text encoder trained 50% of steps, random choice "low quality" "lowres" "jpeg" concatenated onto 10% of prompts, starting at 1e-6 lr, trained at 384x384
## How to use with `diffusers` library (section from [openjourney](https://huggingface.co/openjourney/openjourney))
### Installing necessary libraries
_NOTE: This model currently works on a computer which has at least one NVIDIA GPU with CUDA support_.
```
pip install diffusers transformers ftfy scipy accelerate
```
### Logging in
For logging in, you have to use `huggingface-cli login` command.
### Importing necessary libraries
```python
import torch
from torch import autocast
from diffusers.models import AutoencoderKL
from diffusers import StableDiffusionPipeline
```
### Creating the pipeline
```python
pipe = StableDiffusionPipeline.from_pretrained("doohickey/neopian-diffusion", use_auth_token=True)
pipe = pipe.to("cuda")
```
### (Optional) Disabling NSFW Filter
_NOTE: Remember disabling this is not recommended, but since people had problems with some very basic prompts, we offer this. Remember AI art has a vast majority of users, so keep underage and sensitive users safe._
```python
def dummy(images, **kwargs):
return images, False
pipe.safety_checker = dummy
```
### Image Generation
```python
prompt = "my prompt"
with autocast("cuda"):
image = pipe(prompt=prompt, num_inference_steps=100, width=512, height=512, guidance_scale=15).images[0]
image.save("image.png")
```
## Neopets Copyright Notice
"Don't forget, if you use these images on a non-Neopets page, you need to include our Copyright Notice." https://www.neopets.com/terms.phtml