Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,7 +1,7 @@
|
|
| 1 |
import os
|
| 2 |
import random
|
| 3 |
import uuid
|
| 4 |
-
from typing import Tuple
|
| 5 |
import gradio as gr
|
| 6 |
import numpy as np
|
| 7 |
from PIL import Image
|
|
@@ -9,31 +9,44 @@ import spaces
|
|
| 9 |
import torch
|
| 10 |
from diffusers import StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler
|
| 11 |
|
| 12 |
-
|
| 13 |
-
|
| 14 |
"""
|
| 15 |
|
| 16 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
MAX_SEED = np.iinfo(np.int32).max
|
| 18 |
-
DEFAULT_STYLE_NAME = "3840 x 2160"
|
| 19 |
-
USE_TORCH_COMPILE = False # Set to True if you want to try torch compile (might be faster but requires compatible hardware/drivers)
|
| 20 |
-
ENABLE_CPU_OFFLOAD = False # Set to True to offload parts of the model to CPU (saves VRAM but slower)
|
| 21 |
|
| 22 |
-
#
|
| 23 |
-
|
| 24 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
"RealVisXL V4.0 Lightning": "SG161222/RealVisXL_V4.0_Lightning",
|
| 26 |
"RealVisXL V5.0 Lightning": "SG161222/RealVisXL_V5.0_Lightning",
|
| 27 |
-
# Add more SDXL base models here if desired
|
| 28 |
-
# "Another SDXL Model": "stabilityai/stable-diffusion-xl-base-1.0", # Example
|
| 29 |
}
|
| 30 |
|
| 31 |
-
#
|
| 32 |
-
|
|
|
|
| 33 |
|
| 34 |
-
#
|
| 35 |
LORA_OPTIONS = {
|
| 36 |
-
# Name: (HuggingFace Repo ID, Weight Filename, Adapter Name)
|
| 37 |
"Realism (face/character)👦🏻": ("prithivMLmods/Canopus-Realism-LoRA", "Canopus-Realism-LoRA.safetensors", "rlms"),
|
| 38 |
"Pixar (art/toons)🙀": ("prithivMLmods/Canopus-Pixar-Art", "Canopus-Pixar-Art.safetensors", "pixar"),
|
| 39 |
"Photoshoot (camera/film)📸": ("prithivMLmods/Canopus-Photo-Shoot-Mini-LoRA", "Canopus-Photo-Shoot-Mini-LoRA.safetensors", "photo"),
|
|
@@ -49,88 +62,47 @@ LORA_OPTIONS = {
|
|
| 49 |
"Art Minimalistic (paint/semireal)🎨": ("prithivMLmods/Canopus-Art-Medium-LoRA", "Canopus-Art-Medium-LoRA.safetensors", "mdm"),
|
| 50 |
}
|
| 51 |
|
| 52 |
-
#
|
| 53 |
style_list = [
|
| 54 |
{
|
| 55 |
"name": "3840 x 2160",
|
| 56 |
"prompt": "hyper-realistic 8K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic",
|
| 57 |
-
"negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly
|
| 58 |
},
|
| 59 |
{
|
| 60 |
"name": "2560 x 1440",
|
| 61 |
"prompt": "hyper-realistic 4K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic",
|
| 62 |
-
"negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly
|
| 63 |
},
|
| 64 |
{
|
| 65 |
"name": "HD+",
|
| 66 |
"prompt": "hyper-realistic 2K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic",
|
| 67 |
-
"negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly
|
| 68 |
},
|
| 69 |
{
|
| 70 |
"name": "Style Zero",
|
| 71 |
"prompt": "{prompt}",
|
| 72 |
-
"negative_prompt": "
|
| 73 |
},
|
| 74 |
]
|
|
|
|
| 75 |
styles = {k["name"]: (k["prompt"], k["negative_prompt"]) for k in style_list}
|
|
|
|
| 76 |
STYLE_NAMES = list(styles.keys())
|
| 77 |
|
| 78 |
-
#
|
| 79 |
-
def save_image(img):
|
| 80 |
-
unique_name = str(uuid.uuid4()) + ".png"
|
| 81 |
-
img.save(unique_name)
|
| 82 |
-
return unique_name
|
| 83 |
-
|
| 84 |
-
def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
|
| 85 |
-
if randomize_seed:
|
| 86 |
-
seed = random.randint(0, MAX_SEED)
|
| 87 |
-
return seed
|
| 88 |
-
|
| 89 |
def apply_style(style_name: str, positive: str, negative: str = "") -> Tuple[str, str]:
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
# Combine the base negative prompt with the user's negative prompt
|
| 94 |
-
# Ensure user's negative prompt is appended correctly
|
| 95 |
-
if negative and base_n:
|
| 96 |
-
combined_n = f"{base_n}, {negative}"
|
| 97 |
-
elif negative:
|
| 98 |
-
combined_n = negative
|
| 99 |
-
else:
|
| 100 |
-
combined_n = base_n
|
| 101 |
-
|
| 102 |
-
# Apply the positive prompt template
|
| 103 |
-
final_p = base_p.replace("{prompt}", positive)
|
| 104 |
-
|
| 105 |
-
return final_p, combined_n
|
| 106 |
-
|
| 107 |
-
def load_predefined_images():
|
| 108 |
-
# Ensure the assets directory and images exist
|
| 109 |
-
asset_dir = "assets"
|
| 110 |
-
image_files = [
|
| 111 |
-
"1.png", "2.png", "3.png",
|
| 112 |
-
"4.png", "5.png", "6.png",
|
| 113 |
-
"7.png", "8.png", "9.png",
|
| 114 |
-
]
|
| 115 |
-
predefined_images = []
|
| 116 |
-
if os.path.exists(asset_dir):
|
| 117 |
-
for img_file in image_files:
|
| 118 |
-
img_path = os.path.join(asset_dir, img_file)
|
| 119 |
-
if os.path.exists(img_path):
|
| 120 |
-
predefined_images.append(img_path)
|
| 121 |
-
else:
|
| 122 |
-
print(f"Warning: Predefined image not found: {img_path}")
|
| 123 |
else:
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
return
|
| 128 |
-
|
| 129 |
|
| 130 |
-
#
|
| 131 |
@spaces.GPU(duration=180, enable_queue=True)
|
| 132 |
def generate(
|
| 133 |
-
selected_base_model_name: str, # New input for base model selection
|
| 134 |
prompt: str,
|
| 135 |
negative_prompt: str = "",
|
| 136 |
use_negative_prompt: bool = False,
|
|
@@ -138,212 +110,110 @@ def generate(
|
|
| 138 |
width: int = 1024,
|
| 139 |
height: int = 1024,
|
| 140 |
guidance_scale: float = 3,
|
| 141 |
-
num_inference_steps: int = 4, # Lightning models use fewer steps
|
| 142 |
randomize_seed: bool = False,
|
| 143 |
style_name: str = DEFAULT_STYLE_NAME,
|
| 144 |
-
|
|
|
|
| 145 |
progress=gr.Progress(track_tqdm=True),
|
| 146 |
):
|
| 147 |
-
|
| 148 |
-
raise gr.Error("GPU not available. This Space requires a GPU to run.")
|
| 149 |
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
if selected_base_model_name in loaded_pipelines:
|
| 156 |
-
print(f"Using cached pipeline: {selected_base_model_name}")
|
| 157 |
-
pipe = loaded_pipelines[selected_base_model_name]
|
| 158 |
-
else:
|
| 159 |
-
print(f"Loading pipeline: {selected_base_model_name}")
|
| 160 |
-
model_id = pipelines_info[selected_base_model_name]
|
| 161 |
-
pipe = StableDiffusionXLPipeline.from_pretrained(
|
| 162 |
-
model_id,
|
| 163 |
-
torch_dtype=torch.float16,
|
| 164 |
-
use_safetensors=True,
|
| 165 |
-
variant="fp16" if torch.cuda.is_available() else None # Use fp16 variant if available on GPU
|
| 166 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 167 |
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
pipe.to("cuda") # Default: move entire pipeline to GPU
|
| 174 |
-
|
| 175 |
-
# Configure scheduler (important for Lightning models)
|
| 176 |
-
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
|
| 177 |
-
|
| 178 |
-
# Load ALL LoRAs onto this newly loaded pipeline instance
|
| 179 |
-
print(f"Loading LoRAs for {selected_base_model_name}...")
|
| 180 |
-
for lora_name, (model_repo, weight_file, adapter_tag) in LORA_OPTIONS.items():
|
| 181 |
-
try:
|
| 182 |
-
print(f" Loading LoRA: {lora_name} ({adapter_tag})")
|
| 183 |
-
pipe.load_lora_weights(model_repo, weight_name=weight_file, adapter_name=adapter_tag)
|
| 184 |
-
except Exception as e:
|
| 185 |
-
print(f" Failed to load LoRA {lora_name}: {e}")
|
| 186 |
-
# Optionally raise an error or continue without this LoRA
|
| 187 |
-
# raise gr.Error(f"Failed to load LoRA {lora_name}. Check repo/file names.")
|
| 188 |
-
|
| 189 |
-
if USE_TORCH_COMPILE:
|
| 190 |
-
print("Attempting to compile UNet (may take time)...")
|
| 191 |
-
try:
|
| 192 |
-
pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
|
| 193 |
-
print("UNet compiled successfully.")
|
| 194 |
-
except Exception as e:
|
| 195 |
-
print(f"Torch compile failed: {e}. Running without compilation.")
|
| 196 |
-
|
| 197 |
-
# Cache the fully loaded and configured pipeline
|
| 198 |
-
loaded_pipelines[selected_base_model_name] = pipe
|
| 199 |
-
print(f"Pipeline {selected_base_model_name} loaded and cached.")
|
| 200 |
-
|
| 201 |
-
# --- Prompt Styling ---
|
| 202 |
-
positive_prompt, effective_negative_prompt = apply_style(style_name, prompt, negative_prompt if use_negative_prompt else "")
|
| 203 |
-
|
| 204 |
-
# --- LoRA Selection ---
|
| 205 |
-
if lora_choice not in LORA_OPTIONS:
|
| 206 |
-
raise gr.Error(f"Selected LoRA '{lora_choice}' not found in options.")
|
| 207 |
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
# Note: LoRA weight/scale is often handled within the pipeline or during loading.
|
| 212 |
-
# If you need adjustable LoRA scale, you might need `add_weighted_adapter` or similar.
|
| 213 |
-
# For simplicity here, we assume the default scale is used.
|
| 214 |
-
# cross_attention_kwargs={"scale": 0.8} # Example if you need to set scale explicitly
|
| 215 |
|
| 216 |
-
#
|
| 217 |
-
|
| 218 |
-
generator = torch.Generator("cuda").manual_seed(seed)
|
| 219 |
-
images = pipe(
|
| 220 |
prompt=positive_prompt,
|
| 221 |
negative_prompt=effective_negative_prompt,
|
| 222 |
width=width,
|
| 223 |
height=height,
|
| 224 |
guidance_scale=guidance_scale,
|
| 225 |
-
num_inference_steps=
|
| 226 |
-
generator=generator,
|
| 227 |
num_images_per_prompt=1,
|
| 228 |
-
|
| 229 |
output_type="pil",
|
| 230 |
).images
|
| 231 |
-
|
| 232 |
image_paths = [save_image(img) for img in images]
|
| 233 |
-
print("Image generation complete.")
|
| 234 |
return image_paths, seed
|
| 235 |
|
| 236 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 237 |
css = '''
|
| 238 |
-
.gradio-container{max-width:
|
| 239 |
h1{text-align:center}
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
/* Make gallery taller */
|
| 243 |
-
#result_gallery .h-\[400px\] {
|
| 244 |
-
height: 600px !important; /* Adjust height as needed */
|
| 245 |
}
|
| 246 |
-
#predefined_gallery .h-\[400px\] {
|
| 247 |
-
height: 300px !important; /* Adjust height as needed */
|
| 248 |
-
}
|
| 249 |
-
footer { visibility: hidden }
|
| 250 |
'''
|
| 251 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 252 |
with gr.Blocks(css=css) as demo:
|
| 253 |
gr.Markdown(DESCRIPTIONz)
|
| 254 |
-
|
| 255 |
-
with gr.Row(elem_id="model-select-row"):
|
| 256 |
-
model_selector = gr.Dropdown(
|
| 257 |
-
label="Select Base Model",
|
| 258 |
-
choices=list(pipelines_info.keys()),
|
| 259 |
-
value=list(pipelines_info.keys())[0], # Default to the first model
|
| 260 |
-
scale=1
|
| 261 |
-
)
|
| 262 |
-
model_choice = gr.Dropdown(
|
| 263 |
-
label="Select LoRA Style",
|
| 264 |
-
choices=list(LORA_OPTIONS.keys()),
|
| 265 |
-
value="Realism (face/character)👦🏻", # Default LoRA
|
| 266 |
-
scale=1
|
| 267 |
-
)
|
| 268 |
-
|
| 269 |
with gr.Group():
|
| 270 |
with gr.Row():
|
| 271 |
prompt = gr.Text(
|
| 272 |
label="Prompt",
|
| 273 |
show_label=False,
|
| 274 |
-
max_lines=
|
| 275 |
-
placeholder="Enter your prompt
|
| 276 |
container=False,
|
| 277 |
-
scale=5, # Make prompt input wider
|
| 278 |
)
|
| 279 |
-
run_button = gr.Button("
|
| 280 |
-
|
| 281 |
-
# Use Tabs for Main Result and Examples/Gallery
|
| 282 |
-
with gr.Tabs():
|
| 283 |
-
with gr.TabItem("Result", id="result_tab"):
|
| 284 |
-
result = gr.Gallery(
|
| 285 |
-
label="Generated Image", elem_id="result_gallery",
|
| 286 |
-
columns=1, preview=True, show_label=False, height=600 # Make gallery taller
|
| 287 |
-
)
|
| 288 |
-
# Display the seed used for the generated image
|
| 289 |
-
used_seed = gr.Number(label="Seed Used", interactive=False)
|
| 290 |
-
|
| 291 |
-
with gr.TabItem("Examples & Predefined Gallery", id="examples_tab"):
|
| 292 |
-
gr.Markdown("### Prompt Examples")
|
| 293 |
-
gr.Examples(
|
| 294 |
-
examples=[
|
| 295 |
-
"cinematic photo, a man sitting on a chair in a dark room, realistic", # Realism example
|
| 296 |
-
"pixar style 3d render of a cute cat astronaut exploring mars", # Pixar example
|
| 297 |
-
"studio photography, high fashion model wearing a futuristic silver hoodie, dramatic lighting", # Photoshoot/Clothing example
|
| 298 |
-
"minimalist vector art illustration of a mountain range at sunset, liquid style", # Minimalist/Liquid example
|
| 299 |
-
"pencil sketch drawing of an old wise wizard with a long beard", # Pencil Art example
|
| 300 |
-
],
|
| 301 |
-
inputs=[prompt], # Only update the prompt field from examples
|
| 302 |
-
outputs=[result, used_seed], # Define outputs for example generation
|
| 303 |
-
fn=lambda p: generate( # Need a lambda to pass default values for other args
|
| 304 |
-
selected_base_model_name=list(pipelines_info.keys())[0], # Use default model for examples
|
| 305 |
-
prompt=p,
|
| 306 |
-
lora_choice="Realism (face/character)👦🏻", # Use default LoRA for examples
|
| 307 |
-
# Add other default args from 'generate' signature if needed
|
| 308 |
-
negative_prompt="(deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers:1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation",
|
| 309 |
-
use_negative_prompt=True,
|
| 310 |
-
seed=0, # Or make examples use random seed?
|
| 311 |
-
width=1024,
|
| 312 |
-
height=1024,
|
| 313 |
-
guidance_scale=3.0,
|
| 314 |
-
num_inference_steps=4,
|
| 315 |
-
randomize_seed=True, # Randomize seed for examples
|
| 316 |
-
style_name=DEFAULT_STYLE_NAME,
|
| 317 |
-
),
|
| 318 |
-
cache_examples=False, # Recalculate examples if needed
|
| 319 |
-
label="Click an example to generate"
|
| 320 |
-
)
|
| 321 |
-
gr.Markdown("### Predefined Image Gallery")
|
| 322 |
-
predefined_gallery = gr.Gallery(
|
| 323 |
-
label="Image Gallery", elem_id="predefined_gallery",
|
| 324 |
-
columns=3, show_label=False, value=load_predefined_images(), height=300
|
| 325 |
-
)
|
| 326 |
-
|
| 327 |
-
|
| 328 |
-
with gr.Accordion("⚙️ Advanced Settings", open=False):
|
| 329 |
-
style_selection = gr.Radio(
|
| 330 |
-
show_label=True,
|
| 331 |
-
container=True,
|
| 332 |
-
interactive=True,
|
| 333 |
-
choices=STYLE_NAMES,
|
| 334 |
-
value=DEFAULT_STYLE_NAME,
|
| 335 |
-
label="Image Quality Style",
|
| 336 |
-
)
|
| 337 |
-
with gr.Row():
|
| 338 |
-
use_negative_prompt = gr.Checkbox(label="Use Negative Prompt", value=True)
|
| 339 |
-
randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
|
| 340 |
|
|
|
|
|
|
|
| 341 |
negative_prompt = gr.Text(
|
| 342 |
-
label="Negative
|
| 343 |
-
|
| 344 |
-
|
| 345 |
-
|
| 346 |
-
|
|
|
|
| 347 |
)
|
| 348 |
seed = gr.Slider(
|
| 349 |
label="Seed",
|
|
@@ -351,45 +221,62 @@ with gr.Blocks(css=css) as demo:
|
|
| 351 |
maximum=MAX_SEED,
|
| 352 |
step=1,
|
| 353 |
value=0,
|
| 354 |
-
visible=True
|
| 355 |
-
interactive=True
|
| 356 |
)
|
| 357 |
-
|
| 358 |
-
with gr.Row():
|
| 359 |
width = gr.Slider(
|
| 360 |
label="Width",
|
| 361 |
minimum=512,
|
| 362 |
-
maximum=
|
| 363 |
-
step=
|
| 364 |
value=1024,
|
| 365 |
)
|
| 366 |
height = gr.Slider(
|
| 367 |
label="Height",
|
| 368 |
minimum=512,
|
| 369 |
-
maximum=
|
| 370 |
-
step=
|
| 371 |
value=1024,
|
| 372 |
)
|
| 373 |
-
|
| 374 |
with gr.Row():
|
| 375 |
guidance_scale = gr.Slider(
|
| 376 |
-
label="Guidance Scale
|
| 377 |
-
minimum=0.
|
| 378 |
-
maximum=
|
| 379 |
step=0.1,
|
| 380 |
-
value=
|
| 381 |
-
)
|
| 382 |
-
num_inference_steps = gr.Slider(
|
| 383 |
-
label="Inference Steps",
|
| 384 |
-
minimum=1,
|
| 385 |
-
maximum=20, # Lightning models need very few steps
|
| 386 |
-
step=1,
|
| 387 |
-
value=4, # Default steps for Lightning
|
| 388 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 389 |
|
| 390 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 391 |
|
| 392 |
-
# Show/hide negative prompt input based on checkbox
|
| 393 |
use_negative_prompt.change(
|
| 394 |
fn=lambda x: gr.update(visible=x),
|
| 395 |
inputs=use_negative_prompt,
|
|
@@ -397,54 +284,38 @@ with gr.Blocks(css=css) as demo:
|
|
| 397 |
api_name=False,
|
| 398 |
)
|
| 399 |
|
| 400 |
-
|
| 401 |
-
|
| 402 |
-
|
| 403 |
-
|
| 404 |
-
|
| 405 |
-
|
| 406 |
-
)
|
| 407 |
-
|
| 408 |
-
# Main generation trigger
|
| 409 |
-
inputs_list = [
|
| 410 |
-
model_selector, # Add model selector
|
| 411 |
-
prompt,
|
| 412 |
-
negative_prompt,
|
| 413 |
-
use_negative_prompt,
|
| 414 |
-
seed,
|
| 415 |
-
width,
|
| 416 |
-
height,
|
| 417 |
-
guidance_scale,
|
| 418 |
-
num_inference_steps, # Add steps slider
|
| 419 |
-
randomize_seed,
|
| 420 |
-
style_selection,
|
| 421 |
-
model_choice, # This is the LoRA choice dropdown
|
| 422 |
-
]
|
| 423 |
-
outputs_list = [result, used_seed] # Output gallery and the seed number
|
| 424 |
-
|
| 425 |
-
prompt.submit(
|
| 426 |
fn=generate,
|
| 427 |
-
inputs=
|
| 428 |
-
|
| 429 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 430 |
)
|
| 431 |
-
run_button.click(
|
| 432 |
-
fn=generate,
|
| 433 |
-
inputs=inputs_list,
|
| 434 |
-
outputs=outputs_list,
|
| 435 |
-
api_name="run_button_click" # Optional: Define API name
|
| 436 |
-
)
|
| 437 |
-
|
| 438 |
-
# --- Launch ---
|
| 439 |
-
if __name__ == "__main__":
|
| 440 |
-
if not torch.cuda.is_available():
|
| 441 |
-
print("Warning: No CUDA GPU detected. Running on CPU will be extremely slow or may fail.")
|
| 442 |
-
DESCRIPTIONz += "\n<p>⚠️<b>WARNING: No GPU detected. Running on CPU is very slow and may not work reliably.</b> Consider using a GPU instance.</p>"
|
| 443 |
-
# Optionally disable parts of the UI or exit if CPU is unacceptable
|
| 444 |
-
# exit()
|
| 445 |
|
| 446 |
-
|
| 447 |
-
|
| 448 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 449 |
|
| 450 |
-
|
|
|
|
|
|
| 1 |
import os
|
| 2 |
import random
|
| 3 |
import uuid
|
| 4 |
+
from typing import Tuple
|
| 5 |
import gradio as gr
|
| 6 |
import numpy as np
|
| 7 |
from PIL import Image
|
|
|
|
| 9 |
import torch
|
| 10 |
from diffusers import StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler
|
| 11 |
|
| 12 |
+
# Description for the Gradio interface
|
| 13 |
+
DESCRIPTIONz = """## SDXL-LoRA-DLC ⚡
|
| 14 |
"""
|
| 15 |
|
| 16 |
+
# Function to save generated images
|
| 17 |
+
def save_image(img):
|
| 18 |
+
unique_name = str(uuid.uuid4()) + ".png"
|
| 19 |
+
img.save(unique_name)
|
| 20 |
+
return unique_name
|
| 21 |
+
|
| 22 |
+
# Function to handle seed randomization
|
| 23 |
+
def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
|
| 24 |
+
if randomize_seed:
|
| 25 |
+
seed = random.randint(0, MAX_SEED)
|
| 26 |
+
return seed
|
| 27 |
+
|
| 28 |
MAX_SEED = np.iinfo(np.int32).max
|
|
|
|
|
|
|
|
|
|
| 29 |
|
| 30 |
+
# Warning if running on CPU
|
| 31 |
+
if not torch.cuda.is_available():
|
| 32 |
+
DESCRIPTIONz += "\n<p>⚠️Running on CPU, This may not work on CPU. If it runs for an extended time or if you encounter errors, try running it on a GPU by duplicating the space using @spaces.GPU(). +import spaces.📍</p>"
|
| 33 |
+
|
| 34 |
+
# Configuration flags (unchanged)
|
| 35 |
+
USE_TORCH_COMPILE = 0
|
| 36 |
+
ENABLE_CPU_OFFLOAD = 0
|
| 37 |
+
|
| 38 |
+
# Define available base models
|
| 39 |
+
base_models = {
|
| 40 |
"RealVisXL V4.0 Lightning": "SG161222/RealVisXL_V4.0_Lightning",
|
| 41 |
"RealVisXL V5.0 Lightning": "SG161222/RealVisXL_V5.0_Lightning",
|
|
|
|
|
|
|
| 42 |
}
|
| 43 |
|
| 44 |
+
# Global variables to manage the current pipeline
|
| 45 |
+
current_base_model = None
|
| 46 |
+
current_pipeline = None
|
| 47 |
|
| 48 |
+
# Define LoRA options
|
| 49 |
LORA_OPTIONS = {
|
|
|
|
| 50 |
"Realism (face/character)👦🏻": ("prithivMLmods/Canopus-Realism-LoRA", "Canopus-Realism-LoRA.safetensors", "rlms"),
|
| 51 |
"Pixar (art/toons)🙀": ("prithivMLmods/Canopus-Pixar-Art", "Canopus-Pixar-Art.safetensors", "pixar"),
|
| 52 |
"Photoshoot (camera/film)📸": ("prithivMLmods/Canopus-Photo-Shoot-Mini-LoRA", "Canopus-Photo-Shoot-Mini-LoRA.safetensors", "photo"),
|
|
|
|
| 62 |
"Art Minimalistic (paint/semireal)🎨": ("prithivMLmods/Canopus-Art-Medium-LoRA", "Canopus-Art-Medium-LoRA.safetensors", "mdm"),
|
| 63 |
}
|
| 64 |
|
| 65 |
+
# Define style options
|
| 66 |
style_list = [
|
| 67 |
{
|
| 68 |
"name": "3840 x 2160",
|
| 69 |
"prompt": "hyper-realistic 8K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic",
|
| 70 |
+
"negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly",
|
| 71 |
},
|
| 72 |
{
|
| 73 |
"name": "2560 x 1440",
|
| 74 |
"prompt": "hyper-realistic 4K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic",
|
| 75 |
+
"negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly",
|
| 76 |
},
|
| 77 |
{
|
| 78 |
"name": "HD+",
|
| 79 |
"prompt": "hyper-realistic 2K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic",
|
| 80 |
+
"negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly",
|
| 81 |
},
|
| 82 |
{
|
| 83 |
"name": "Style Zero",
|
| 84 |
"prompt": "{prompt}",
|
| 85 |
+
"negative_prompt": "",
|
| 86 |
},
|
| 87 |
]
|
| 88 |
+
|
| 89 |
styles = {k["name"]: (k["prompt"], k["negative_prompt"]) for k in style_list}
|
| 90 |
+
DEFAULT_STYLE_NAME = "3840 x 2160"
|
| 91 |
STYLE_NAMES = list(styles.keys())
|
| 92 |
|
| 93 |
+
# Function to apply selected style
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
def apply_style(style_name: str, positive: str, negative: str = "") -> Tuple[str, str]:
|
| 95 |
+
if style_name in styles:
|
| 96 |
+
p, n = styles.get(style_name, styles[DEFAULT_STYLE_NAME])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 97 |
else:
|
| 98 |
+
p, n = styles[DEFAULT_STYLE_NAME]
|
| 99 |
+
if not negative:
|
| 100 |
+
negative = ""
|
| 101 |
+
return p.replace("{prompt}", positive), n + negative
|
|
|
|
| 102 |
|
| 103 |
+
# Generation function with model selection
|
| 104 |
@spaces.GPU(duration=180, enable_queue=True)
|
| 105 |
def generate(
|
|
|
|
| 106 |
prompt: str,
|
| 107 |
negative_prompt: str = "",
|
| 108 |
use_negative_prompt: bool = False,
|
|
|
|
| 110 |
width: int = 1024,
|
| 111 |
height: int = 1024,
|
| 112 |
guidance_scale: float = 3,
|
|
|
|
| 113 |
randomize_seed: bool = False,
|
| 114 |
style_name: str = DEFAULT_STYLE_NAME,
|
| 115 |
+
lora_model: str = "Realism (face/character)👦🏻",
|
| 116 |
+
base_model: str = "RealVisXL V5.0 Lightning",
|
| 117 |
progress=gr.Progress(track_tqdm=True),
|
| 118 |
):
|
| 119 |
+
global current_base_model, current_pipeline
|
|
|
|
| 120 |
|
| 121 |
+
# Load the pipeline if the base model has changed
|
| 122 |
+
if base_model != current_base_model:
|
| 123 |
+
model_id = base_models[base_model]
|
| 124 |
+
current_pipeline = StableDiffusionXLPipeline.from_pretrained(
|
| 125 |
+
model_id, torch_dtype=torch.float16, use_safetensors=True
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 126 |
)
|
| 127 |
+
current_pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(
|
| 128 |
+
current_pipeline.scheduler.config
|
| 129 |
+
)
|
| 130 |
+
for lora_display_name, (lora_model, lora_weight, adapter_name) in LORA_OPTIONS.items():
|
| 131 |
+
current_pipeline.load_lora_weights(
|
| 132 |
+
lora_model, weight_name=lora_weight, adapter_name=adapter_name
|
| 133 |
+
)
|
| 134 |
+
current_pipeline.to("cuda")
|
| 135 |
+
current_base_model = base_model
|
| 136 |
|
| 137 |
+
# Handle seed and prompts
|
| 138 |
+
seed = int(randomize_seed_fn(seed, randomize_seed))
|
| 139 |
+
positive_prompt, effective_negative_prompt = apply_style(style_name, prompt, negative_prompt)
|
| 140 |
+
if not use_negative_prompt:
|
| 141 |
+
effective_negative_prompt = ""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 142 |
|
| 143 |
+
# Set the LoRA adapter
|
| 144 |
+
_, _, adapter_name = LORA_OPTIONS[lora_model]
|
| 145 |
+
current_pipeline.set_adapters(adapter_name)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 146 |
|
| 147 |
+
# Generate the image
|
| 148 |
+
images = current_pipeline(
|
|
|
|
|
|
|
| 149 |
prompt=positive_prompt,
|
| 150 |
negative_prompt=effective_negative_prompt,
|
| 151 |
width=width,
|
| 152 |
height=height,
|
| 153 |
guidance_scale=guidance_scale,
|
| 154 |
+
num_inference_steps=20,
|
|
|
|
| 155 |
num_images_per_prompt=1,
|
| 156 |
+
cross_attention_kwargs={"scale": 0.65},
|
| 157 |
output_type="pil",
|
| 158 |
).images
|
|
|
|
| 159 |
image_paths = [save_image(img) for img in images]
|
|
|
|
| 160 |
return image_paths, seed
|
| 161 |
|
| 162 |
+
# Example prompts
|
| 163 |
+
examples = [
|
| 164 |
+
"Realism: Man in the style of dark beige and brown, uhd image, youthful protagonists, nonrepresentational ",
|
| 165 |
+
"Pixar: A young man with light brown wavy hair and light brown eyes sitting in an armchair and looking directly at the camera, pixar style, disney pixar, office background, ultra detailed, 1 man",
|
| 166 |
+
"Hoodie: Front view, capture a urban style, Superman Hoodie, technical materials, fabric small point label on text Blue theory, the design is minimal, with a raised collar, fabric is a Light yellow, low angle to capture the Hoodies form and detailing, f/5.6 to focus on the hoodies craftsmanship, solid grey background, studio light setting, with batman logo in the chest region of the t-shirt",
|
| 167 |
+
]
|
| 168 |
+
|
| 169 |
+
# CSS styling
|
| 170 |
css = '''
|
| 171 |
+
.gradio-container{max-width: 545px !important}
|
| 172 |
h1{text-align:center}
|
| 173 |
+
footer {
|
| 174 |
+
visibility: hidden
|
|
|
|
|
|
|
|
|
|
| 175 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 176 |
'''
|
| 177 |
|
| 178 |
+
# Function to load predefined images
|
| 179 |
+
def load_predefined_images():
|
| 180 |
+
predefined_images = [
|
| 181 |
+
"assets/1.png",
|
| 182 |
+
"assets/2.png",
|
| 183 |
+
"assets/3.png",
|
| 184 |
+
"assets/4.png",
|
| 185 |
+
"assets/5.png",
|
| 186 |
+
"assets/6.png",
|
| 187 |
+
"assets/7.png",
|
| 188 |
+
"assets/8.png",
|
| 189 |
+
"assets/9.png",
|
| 190 |
+
]
|
| 191 |
+
return predefined_images
|
| 192 |
+
|
| 193 |
+
# Gradio interface
|
| 194 |
with gr.Blocks(css=css) as demo:
|
| 195 |
gr.Markdown(DESCRIPTIONz)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 196 |
with gr.Group():
|
| 197 |
with gr.Row():
|
| 198 |
prompt = gr.Text(
|
| 199 |
label="Prompt",
|
| 200 |
show_label=False,
|
| 201 |
+
max_lines=1,
|
| 202 |
+
placeholder="Enter your prompt with resp. tag!",
|
| 203 |
container=False,
|
|
|
|
| 204 |
)
|
| 205 |
+
run_button = gr.Button("Run", scale=0)
|
| 206 |
+
result = gr.Gallery(label="Result", columns=1, preview=True, show_label=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 207 |
|
| 208 |
+
with gr.Accordion("Advanced options", open=False, visible=False):
|
| 209 |
+
use_negative_prompt = gr.Checkbox(label="Use negative prompt", value=True)
|
| 210 |
negative_prompt = gr.Text(
|
| 211 |
+
label="Negative prompt",
|
| 212 |
+
lines=4,
|
| 213 |
+
max_lines=6,
|
| 214 |
+
value="(deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers:1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation",
|
| 215 |
+
placeholder="Enter a negative prompt",
|
| 216 |
+
visible=True,
|
| 217 |
)
|
| 218 |
seed = gr.Slider(
|
| 219 |
label="Seed",
|
|
|
|
| 221 |
maximum=MAX_SEED,
|
| 222 |
step=1,
|
| 223 |
value=0,
|
| 224 |
+
visible=True
|
|
|
|
| 225 |
)
|
| 226 |
+
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
| 227 |
+
with gr.Row(visible=True):
|
| 228 |
width = gr.Slider(
|
| 229 |
label="Width",
|
| 230 |
minimum=512,
|
| 231 |
+
maximum=2048,
|
| 232 |
+
step=8,
|
| 233 |
value=1024,
|
| 234 |
)
|
| 235 |
height = gr.Slider(
|
| 236 |
label="Height",
|
| 237 |
minimum=512,
|
| 238 |
+
maximum=2048,
|
| 239 |
+
step=8,
|
| 240 |
value=1024,
|
| 241 |
)
|
|
|
|
| 242 |
with gr.Row():
|
| 243 |
guidance_scale = gr.Slider(
|
| 244 |
+
label="Guidance Scale",
|
| 245 |
+
minimum=0.1,
|
| 246 |
+
maximum=20.0,
|
| 247 |
step=0.1,
|
| 248 |
+
value=3.0,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 249 |
)
|
| 250 |
+
style_selection = gr.Radio(
|
| 251 |
+
show_label=True,
|
| 252 |
+
container=True,
|
| 253 |
+
interactive=True,
|
| 254 |
+
choices=STYLE_NAMES,
|
| 255 |
+
value=DEFAULT_STYLE_NAME,
|
| 256 |
+
label="Quality Style",
|
| 257 |
+
)
|
| 258 |
+
|
| 259 |
+
# Add base model and LoRA selection dropdowns
|
| 260 |
+
with gr.Row():
|
| 261 |
+
base_model_choice = gr.Dropdown(
|
| 262 |
+
label="Base Model",
|
| 263 |
+
choices=list(base_models.keys()),
|
| 264 |
+
value="RealVisXL V5.0 Lightning"
|
| 265 |
+
)
|
| 266 |
+
model_choice = gr.Dropdown(
|
| 267 |
+
label="LoRA Selection",
|
| 268 |
+
choices=list(LORA_OPTIONS.keys()),
|
| 269 |
+
value="Realism (face/character)👦🏻"
|
| 270 |
+
)
|
| 271 |
|
| 272 |
+
gr.Examples(
|
| 273 |
+
examples=examples,
|
| 274 |
+
inputs=prompt,
|
| 275 |
+
outputs=[result, seed],
|
| 276 |
+
fn=generate,
|
| 277 |
+
cache_examples=False,
|
| 278 |
+
)
|
| 279 |
|
|
|
|
| 280 |
use_negative_prompt.change(
|
| 281 |
fn=lambda x: gr.update(visible=x),
|
| 282 |
inputs=use_negative_prompt,
|
|
|
|
| 284 |
api_name=False,
|
| 285 |
)
|
| 286 |
|
| 287 |
+
gr.on(
|
| 288 |
+
triggers=[
|
| 289 |
+
prompt.submit,
|
| 290 |
+
negative_prompt.submit,
|
| 291 |
+
run_button.click,
|
| 292 |
+
],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 293 |
fn=generate,
|
| 294 |
+
inputs=[
|
| 295 |
+
prompt,
|
| 296 |
+
negative_prompt,
|
| 297 |
+
use_negative_prompt,
|
| 298 |
+
seed,
|
| 299 |
+
width,
|
| 300 |
+
height,
|
| 301 |
+
guidance_scale,
|
| 302 |
+
randomize_seed,
|
| 303 |
+
style_selection,
|
| 304 |
+
model_choice,
|
| 305 |
+
base_model_choice,
|
| 306 |
+
],
|
| 307 |
+
outputs=[result, seed],
|
| 308 |
+
api_name="run",
|
| 309 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 310 |
|
| 311 |
+
with gr.Column(scale=3):
|
| 312 |
+
gr.Markdown("### Image Gallery")
|
| 313 |
+
predefined_gallery = gr.Gallery(
|
| 314 |
+
label="Image Gallery",
|
| 315 |
+
columns=3,
|
| 316 |
+
show_label=False,
|
| 317 |
+
value=load_predefined_images()
|
| 318 |
+
)
|
| 319 |
|
| 320 |
+
if __name__ == "__main__":
|
| 321 |
+
demo.queue(max_size=30).launch()
|