Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -0,0 +1,311 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import torch
|
3 |
+
from diffusers import StableDiffusionPipeline
|
4 |
+
from PIL import Image
|
5 |
+
import numpy as np
|
6 |
+
import os
|
7 |
+
from huggingface_hub import hf_hub_download
|
8 |
+
import warnings
|
9 |
+
from transformers import CLIPProcessor, CLIPModel
|
10 |
+
warnings.filterwarnings("ignore")
|
11 |
+
|
12 |
+
# Check if CUDA is available
|
13 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
14 |
+
print(f"Using device: {device}")
|
15 |
+
|
16 |
+
# Load CLIP model for semantic guidance
|
17 |
+
print("Loading CLIP model for semantic guidance...")
|
18 |
+
clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").to(device)
|
19 |
+
clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
20 |
+
|
21 |
+
# Dictionary of available concepts
|
22 |
+
CONCEPTS = {
|
23 |
+
"canna-lily-flowers102": {
|
24 |
+
"repo_id": "sd-concepts-library/canna-lily-flowers102",
|
25 |
+
"type": "object",
|
26 |
+
"description": "Canna lily flower style"
|
27 |
+
},
|
28 |
+
"samurai-jack": {
|
29 |
+
"repo_id": "sd-concepts-library/samurai-jack",
|
30 |
+
"type": "style",
|
31 |
+
"description": "Samurai Jack animation style"
|
32 |
+
},
|
33 |
+
"babies-poster": {
|
34 |
+
"repo_id": "sd-concepts-library/babies-poster",
|
35 |
+
"type": "style",
|
36 |
+
"description": "Babies poster art style"
|
37 |
+
},
|
38 |
+
"animal-toy": {
|
39 |
+
"repo_id": "sd-concepts-library/animal-toy",
|
40 |
+
"type": "object",
|
41 |
+
"description": "Animal toy style"
|
42 |
+
},
|
43 |
+
"sword-lily-flowers102": {
|
44 |
+
"repo_id": "sd-concepts-library/sword-lily-flowers102",
|
45 |
+
"type": "object",
|
46 |
+
"description": "Sword lily flower style"
|
47 |
+
}
|
48 |
+
}
|
49 |
+
|
50 |
+
def car_loss(image):
|
51 |
+
"""Custom loss function that encourages the presence of cars in the image"""
|
52 |
+
# Convert PIL image to tensor if needed
|
53 |
+
if isinstance(image, Image.Image):
|
54 |
+
image = np.array(image)
|
55 |
+
image = torch.tensor(image, device=device)
|
56 |
+
|
57 |
+
# Process image for CLIP
|
58 |
+
with torch.no_grad():
|
59 |
+
# Convert to PIL for CLIP processing
|
60 |
+
pil_image = Image.fromarray(image.cpu().numpy().astype(np.uint8))
|
61 |
+
|
62 |
+
# Get CLIP features for the image
|
63 |
+
inputs = clip_processor(
|
64 |
+
text=["a photo of a car", "a photo without cars"],
|
65 |
+
images=pil_image,
|
66 |
+
return_tensors="pt",
|
67 |
+
padding=True
|
68 |
+
).to(device)
|
69 |
+
|
70 |
+
# Get similarity scores
|
71 |
+
outputs = clip_model(**inputs)
|
72 |
+
logits_per_image = outputs.logits_per_image
|
73 |
+
|
74 |
+
# Higher score for the first text (with cars) is better
|
75 |
+
car_score = logits_per_image[0][0]
|
76 |
+
no_car_score = logits_per_image[0][1]
|
77 |
+
|
78 |
+
# We want to maximize car_score and minimize no_car_score
|
79 |
+
loss = -(car_score - no_car_score)
|
80 |
+
|
81 |
+
return loss
|
82 |
+
|
83 |
+
def generate_image(pipe, prompt, seed, guidance_scale=7.5, num_inference_steps=30, use_car_guidance=False):
|
84 |
+
"""Generate an image with optional car guidance"""
|
85 |
+
generator = torch.Generator(device).manual_seed(seed)
|
86 |
+
custom_loss = car_loss if use_car_guidance else None
|
87 |
+
|
88 |
+
if custom_loss:
|
89 |
+
try:
|
90 |
+
# Start with a standard generation
|
91 |
+
init_images = pipe(
|
92 |
+
prompt,
|
93 |
+
guidance_scale=guidance_scale,
|
94 |
+
num_inference_steps=num_inference_steps // 2,
|
95 |
+
generator=generator
|
96 |
+
).images
|
97 |
+
|
98 |
+
init_image = init_images[0]
|
99 |
+
|
100 |
+
# Refine using car guidance
|
101 |
+
from diffusers import StableDiffusionImg2ImgPipeline
|
102 |
+
|
103 |
+
img2img_pipe = StableDiffusionImg2ImgPipeline(
|
104 |
+
vae=pipe.vae,
|
105 |
+
text_encoder=pipe.text_encoder,
|
106 |
+
tokenizer=pipe.tokenizer,
|
107 |
+
unet=pipe.unet,
|
108 |
+
scheduler=pipe.scheduler,
|
109 |
+
safety_checker=None,
|
110 |
+
feature_extractor=None,
|
111 |
+
).to(device)
|
112 |
+
|
113 |
+
strength = 0.75
|
114 |
+
current_image = init_image
|
115 |
+
|
116 |
+
for i in range(5):
|
117 |
+
current_loss = custom_loss(current_image)
|
118 |
+
|
119 |
+
refined_images = img2img_pipe(
|
120 |
+
prompt=prompt + ", with beautiful cars",
|
121 |
+
image=current_image,
|
122 |
+
strength=strength,
|
123 |
+
guidance_scale=guidance_scale,
|
124 |
+
generator=generator,
|
125 |
+
).images
|
126 |
+
|
127 |
+
current_image = refined_images[0]
|
128 |
+
strength *= 0.8
|
129 |
+
|
130 |
+
return current_image
|
131 |
+
|
132 |
+
except Exception as e:
|
133 |
+
print(f"Error in car-guided generation: {e}")
|
134 |
+
return pipe(
|
135 |
+
prompt,
|
136 |
+
guidance_scale=guidance_scale,
|
137 |
+
num_inference_steps=num_inference_steps,
|
138 |
+
generator=generator
|
139 |
+
).images[0]
|
140 |
+
else:
|
141 |
+
return pipe(
|
142 |
+
prompt,
|
143 |
+
guidance_scale=guidance_scale,
|
144 |
+
num_inference_steps=num_inference_steps,
|
145 |
+
generator=generator
|
146 |
+
).images[0]
|
147 |
+
|
148 |
+
# Cache for loaded models and concepts
|
149 |
+
loaded_models = {}
|
150 |
+
|
151 |
+
def get_model_with_concept(concept_name):
|
152 |
+
"""Get or load a model with the specified concept"""
|
153 |
+
if concept_name not in loaded_models:
|
154 |
+
concept_info = CONCEPTS[concept_name]
|
155 |
+
|
156 |
+
# Download concept embedding
|
157 |
+
concept_path = f"concepts/{concept_name}.bin"
|
158 |
+
os.makedirs("concepts", exist_ok=True)
|
159 |
+
|
160 |
+
if not os.path.exists(concept_path):
|
161 |
+
file = hf_hub_download(
|
162 |
+
repo_id=concept_info["repo_id"],
|
163 |
+
filename="learned_embeds.bin",
|
164 |
+
repo_type="model"
|
165 |
+
)
|
166 |
+
import shutil
|
167 |
+
shutil.copy(file, concept_path)
|
168 |
+
|
169 |
+
# Load model and concept
|
170 |
+
pipe = StableDiffusionPipeline.from_pretrained(
|
171 |
+
"stabilityai/stable-diffusion-2",
|
172 |
+
torch_dtype=torch.float32 if device == "cpu" else torch.float16,
|
173 |
+
safety_checker=None
|
174 |
+
).to(device)
|
175 |
+
|
176 |
+
pipe.load_textual_inversion(concept_path)
|
177 |
+
loaded_models[concept_name] = pipe
|
178 |
+
|
179 |
+
return loaded_models[concept_name]
|
180 |
+
|
181 |
+
def generate_images(concept_name, base_prompt, seed, use_car_guidance):
|
182 |
+
"""Generate images using the selected concept"""
|
183 |
+
try:
|
184 |
+
# Get model with concept
|
185 |
+
pipe = get_model_with_concept(concept_name)
|
186 |
+
|
187 |
+
# Construct prompt based on concept type
|
188 |
+
if CONCEPTS[concept_name]["type"] == "object":
|
189 |
+
prompt = f"A {base_prompt} with a <{concept_name}>"
|
190 |
+
else:
|
191 |
+
prompt = f"<{concept_name}> {base_prompt}"
|
192 |
+
|
193 |
+
# Generate image
|
194 |
+
image = generate_image(
|
195 |
+
pipe=pipe,
|
196 |
+
prompt=prompt,
|
197 |
+
seed=int(seed),
|
198 |
+
use_car_guidance=use_car_guidance
|
199 |
+
)
|
200 |
+
|
201 |
+
return image
|
202 |
+
except Exception as e:
|
203 |
+
raise gr.Error(f"Error generating image: {str(e)}")
|
204 |
+
|
205 |
+
# Create Gradio interface
|
206 |
+
with gr.Blocks(title="Stable Diffusion Style Explorer") as demo:
|
207 |
+
gr.Markdown("""
|
208 |
+
# Stable Diffusion Style Explorer
|
209 |
+
|
210 |
+
Generate images using various concepts from the SD Concepts Library, with optional car guidance.
|
211 |
+
|
212 |
+
## How to use:
|
213 |
+
1. Select a concept from the dropdown
|
214 |
+
2. Enter a base prompt (or use the default)
|
215 |
+
3. Set a seed for reproducibility
|
216 |
+
4. Choose whether to use car guidance
|
217 |
+
5. Click Generate!
|
218 |
+
|
219 |
+
Check out the examples below to see different combinations of concepts and prompts!
|
220 |
+
""")
|
221 |
+
|
222 |
+
with gr.Row():
|
223 |
+
with gr.Column():
|
224 |
+
concept = gr.Dropdown(
|
225 |
+
choices=list(CONCEPTS.keys()),
|
226 |
+
value="samurai-jack",
|
227 |
+
label="Select Concept"
|
228 |
+
)
|
229 |
+
|
230 |
+
prompt = gr.Textbox(
|
231 |
+
value="A serene landscape with mountains and a lake at sunset",
|
232 |
+
label="Base Prompt"
|
233 |
+
)
|
234 |
+
|
235 |
+
seed = gr.Number(
|
236 |
+
value=42,
|
237 |
+
label="Seed",
|
238 |
+
precision=0
|
239 |
+
)
|
240 |
+
|
241 |
+
car_guidance = gr.Checkbox(
|
242 |
+
value=False,
|
243 |
+
label="Use Car Guidance"
|
244 |
+
)
|
245 |
+
|
246 |
+
generate_btn = gr.Button("Generate Image")
|
247 |
+
|
248 |
+
with gr.Column():
|
249 |
+
output_image = gr.Image(label="Generated Image")
|
250 |
+
|
251 |
+
concept.change(
|
252 |
+
fn=lambda x: gr.Markdown(f"Selected concept: {CONCEPTS[x]['description']} ({CONCEPTS[x]['type']})"),
|
253 |
+
inputs=[concept],
|
254 |
+
outputs=[gr.Markdown()]
|
255 |
+
)
|
256 |
+
|
257 |
+
generate_btn.click(
|
258 |
+
fn=generate_images,
|
259 |
+
inputs=[concept, prompt, seed, car_guidance],
|
260 |
+
outputs=[output_image]
|
261 |
+
)
|
262 |
+
|
263 |
+
# Gallery of pre-generated examples
|
264 |
+
gr.Markdown("### 🖼️ Pre-generated Examples")
|
265 |
+
|
266 |
+
with gr.Row():
|
267 |
+
# Samurai Jack examples
|
268 |
+
with gr.Column():
|
269 |
+
gr.Markdown("**Samurai Jack Style**")
|
270 |
+
gr.Image("Assignment17/Assignment17/outputs/samurai-jack_normal.png",
|
271 |
+
label="Without Car Guidance")
|
272 |
+
gr.Image("Assignment17/Assignment17/outputs/samurai-jack_car.png",
|
273 |
+
label="With Car Guidance")
|
274 |
+
|
275 |
+
with gr.Row():
|
276 |
+
# Canna Lily examples
|
277 |
+
with gr.Column():
|
278 |
+
gr.Markdown("**Canna Lily Object**")
|
279 |
+
gr.Image("Assignment17/Assignment17/outputs/canna-lily-flowers102_normal.png",
|
280 |
+
label="Without Car Guidance")
|
281 |
+
gr.Image("Assignment17/Assignment17/outputs/canna-lily-flowers102_car.png",
|
282 |
+
label="With Car Guidance")
|
283 |
+
|
284 |
+
with gr.Row():
|
285 |
+
# Babies Poster examples
|
286 |
+
with gr.Column():
|
287 |
+
gr.Markdown("**Babies Poster Style**")
|
288 |
+
gr.Image("Assignment17/Assignment17/outputs/babies-poster_normal.png",
|
289 |
+
label="Without Car Guidance")
|
290 |
+
gr.Image("Assignment17/Assignment17/outputs/babies-poster_car.png",
|
291 |
+
label="With Car Guidance")
|
292 |
+
|
293 |
+
with gr.Row():
|
294 |
+
# Animal Toy examples
|
295 |
+
with gr.Column():
|
296 |
+
gr.Markdown("**Animal Toy Object**")
|
297 |
+
gr.Image("Assignment17/Assignment17/outputs/animal-toy_normal.png",
|
298 |
+
label="Without Car Guidance")
|
299 |
+
gr.Image("Assignment17/Assignment17/outputs/animal-toy_car.png",
|
300 |
+
label="With Car Guidance")
|
301 |
+
|
302 |
+
with gr.Row():
|
303 |
+
# Sword Lily examples
|
304 |
+
with gr.Column():
|
305 |
+
gr.Markdown("**Sword Lily Object**")
|
306 |
+
gr.Image("Assignment17/Assignment17/outputs/sword-lily-flowers102_normal.png",
|
307 |
+
label="Without Car Guidance")
|
308 |
+
gr.Image("Assignment17/Assignment17/outputs/sword-lily-flowers102_car.png",
|
309 |
+
label="With Car Guidance")
|
310 |
+
|
311 |
+
demo.launch()
|