Spaces:
Sleeping
Sleeping
Jon Taylor
commited on
Commit
•
6776a75
1
Parent(s):
70de1d6
added reference image to test diffusion
Browse files- app/pipeline.py +51 -2
- app/pipeline_test.py +4 -1
- requirements.txt +1 -0
app/pipeline.py
CHANGED
@@ -13,13 +13,16 @@ try:
|
|
13 |
except:
|
14 |
pass
|
15 |
|
16 |
-
import psutil
|
17 |
from pydantic import BaseModel, Field
|
18 |
from PIL import Image
|
|
|
19 |
import math
|
20 |
import time
|
21 |
import os
|
22 |
|
|
|
|
|
|
|
23 |
taesd_model = "madebyollin/taesd"
|
24 |
controlnet_model = "thibaud/controlnet-sd21-canny-diffusers"
|
25 |
base_model = "stabilityai/sd-turbo"
|
@@ -168,7 +171,7 @@ class Pipeline:
|
|
168 |
taesd_model, torch_dtype=torch_dtype, use_safetensors=True
|
169 |
).to(device)
|
170 |
|
171 |
-
if os.getenv("TORCH_COMPILE"
|
172 |
self.pipe.unet = torch.compile(
|
173 |
self.pipe.unet, mode="reduce-overhead", fullgraph=True
|
174 |
)
|
@@ -181,3 +184,49 @@ class Pipeline:
|
|
181 |
image=[Image.new("RGB", (768, 768))],
|
182 |
control_image=[Image.new("RGB", (768, 768))],
|
183 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
except:
|
14 |
pass
|
15 |
|
|
|
16 |
from pydantic import BaseModel, Field
|
17 |
from PIL import Image
|
18 |
+
import psutil
|
19 |
import math
|
20 |
import time
|
21 |
import os
|
22 |
|
23 |
+
from dotenv import load_dotenv
|
24 |
+
load_dotenv()
|
25 |
+
|
26 |
taesd_model = "madebyollin/taesd"
|
27 |
controlnet_model = "thibaud/controlnet-sd21-canny-diffusers"
|
28 |
base_model = "stabilityai/sd-turbo"
|
|
|
171 |
taesd_model, torch_dtype=torch_dtype, use_safetensors=True
|
172 |
).to(device)
|
173 |
|
174 |
+
if bool(os.getenv("TORCH_COMPILE")):
|
175 |
self.pipe.unet = torch.compile(
|
176 |
self.pipe.unet, mode="reduce-overhead", fullgraph=True
|
177 |
)
|
|
|
184 |
image=[Image.new("RGB", (768, 768))],
|
185 |
control_image=[Image.new("RGB", (768, 768))],
|
186 |
)
|
187 |
+
|
188 |
+
def predict(self, params: "Pipeline.InputParams", image) -> Image.Image:
|
189 |
+
generator = torch.manual_seed(params.seed)
|
190 |
+
prompt_embeds = self.pipe.compel_proc(params.prompt)
|
191 |
+
control_image = self.canny_torch(
|
192 |
+
image, params.canny_low_threshold, params.canny_high_threshold
|
193 |
+
)
|
194 |
+
steps = params.steps
|
195 |
+
strength = params.strength
|
196 |
+
if int(steps * strength) < 1:
|
197 |
+
steps = math.ceil(1 / max(0.10, strength))
|
198 |
+
last_time = time.time()
|
199 |
+
results = self.pipe(
|
200 |
+
image=image,
|
201 |
+
control_image=control_image,
|
202 |
+
prompt_embeds=prompt_embeds,
|
203 |
+
generator=generator,
|
204 |
+
strength=strength,
|
205 |
+
num_inference_steps=steps,
|
206 |
+
guidance_scale=params.guidance_scale,
|
207 |
+
width=params.width,
|
208 |
+
height=params.height,
|
209 |
+
output_type="pil",
|
210 |
+
controlnet_conditioning_scale=params.controlnet_scale,
|
211 |
+
control_guidance_start=params.controlnet_start,
|
212 |
+
control_guidance_end=params.controlnet_end,
|
213 |
+
)
|
214 |
+
print(f"Time taken: {time.time() - last_time}")
|
215 |
+
|
216 |
+
nsfw_content_detected = (
|
217 |
+
results.nsfw_content_detected[0]
|
218 |
+
if "nsfw_content_detected" in results
|
219 |
+
else False
|
220 |
+
)
|
221 |
+
if nsfw_content_detected:
|
222 |
+
return None
|
223 |
+
result_image = results.images[0]
|
224 |
+
|
225 |
+
if os.getenv("CONTROL_NET_OVERLAY"):
|
226 |
+
# paste control_image on top of result_image
|
227 |
+
w0, h0 = (200, 200)
|
228 |
+
control_image = control_image.resize((w0, h0))
|
229 |
+
w1, h1 = result_image.size
|
230 |
+
result_image.paste(control_image, (w1 - w0, h1 - h0))
|
231 |
+
|
232 |
+
return result_image
|
app/pipeline_test.py
CHANGED
@@ -1,9 +1,12 @@
|
|
1 |
from pipeline import Pipeline
|
2 |
from device import device, torch_dtype
|
|
|
3 |
|
4 |
def main():
|
5 |
p = Pipeline(device, torch_dtype)
|
6 |
-
|
|
|
|
|
7 |
|
8 |
if __name__ == "__main__":
|
9 |
main()
|
|
|
1 |
from pipeline import Pipeline
|
2 |
from device import device, torch_dtype
|
3 |
+
from diffusers.utils import load_image
|
4 |
|
5 |
def main():
|
6 |
p = Pipeline(device, torch_dtype)
|
7 |
+
params = Pipeline.InputParams()
|
8 |
+
image = load_image("https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png")
|
9 |
+
p.predict(params, image).show()
|
10 |
|
11 |
if __name__ == "__main__":
|
12 |
main()
|
requirements.txt
CHANGED
@@ -10,6 +10,7 @@ pillow
|
|
10 |
pydantic
|
11 |
utils
|
12 |
psutil
|
|
|
13 |
|
14 |
transformers==4.35.2
|
15 |
torch==2.1.1
|
|
|
10 |
pydantic
|
11 |
utils
|
12 |
psutil
|
13 |
+
dotenv
|
14 |
|
15 |
transformers==4.35.2
|
16 |
torch==2.1.1
|