Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
|
@@ -2,15 +2,71 @@ import os
|
|
| 2 |
import random
|
| 3 |
import uuid
|
| 4 |
import json
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
import gradio as gr
|
| 6 |
-
import numpy as np
|
| 7 |
-
from PIL import Image
|
| 8 |
import spaces
|
| 9 |
import torch
|
| 10 |
-
|
| 11 |
-
from
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
|
| 13 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
bad_words = json.loads(os.getenv('BAD_WORDS', "[]"))
|
| 15 |
bad_words_negative = json.loads(os.getenv('BAD_WORDS_NEGATIVE', "[]"))
|
| 16 |
default_negative = os.getenv("default_negative", "")
|
|
@@ -24,218 +80,11 @@ def check_text(prompt, negative=""):
|
|
| 24 |
return True
|
| 25 |
return False
|
| 26 |
|
| 27 |
-
#
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
"negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly",
|
| 33 |
-
},
|
| 34 |
-
{
|
| 35 |
-
"name": "2560 x 1440",
|
| 36 |
-
"prompt": "hyper-realistic 4K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic",
|
| 37 |
-
"negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly",
|
| 38 |
-
},
|
| 39 |
-
{
|
| 40 |
-
"name": "HD+",
|
| 41 |
-
"prompt": "hyper-realistic 2K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic",
|
| 42 |
-
"negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly",
|
| 43 |
-
},
|
| 44 |
-
{
|
| 45 |
-
"name": "Style Zero",
|
| 46 |
-
"prompt": "{prompt}",
|
| 47 |
-
"negative_prompt": "",
|
| 48 |
-
},
|
| 49 |
-
]
|
| 50 |
-
|
| 51 |
-
# Collage styles--------------------------------------------------------------------
|
| 52 |
-
collage_style_list = [
|
| 53 |
-
{
|
| 54 |
-
"name": "Hi-Res",
|
| 55 |
-
"prompt": "hyper-realistic 8K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic",
|
| 56 |
-
"negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly",
|
| 57 |
-
},
|
| 58 |
-
{
|
| 59 |
-
"name": "B & W",
|
| 60 |
-
"prompt": "black and white collage of {prompt}. monochromatic, timeless, classic, dramatic contrast",
|
| 61 |
-
"negative_prompt": "colorful, vibrant, bright, flashy",
|
| 62 |
-
},
|
| 63 |
-
{
|
| 64 |
-
"name": "Polaroid",
|
| 65 |
-
"prompt": "collage of polaroid photos featuring {prompt}. vintage style, high contrast, nostalgic, instant film aesthetic",
|
| 66 |
-
"negative_prompt": "digital, modern, low quality, blurry",
|
| 67 |
-
},
|
| 68 |
-
{
|
| 69 |
-
"name": "Watercolor",
|
| 70 |
-
"prompt": "watercolor collage of {prompt}. soft edges, translucent colors, painterly effects",
|
| 71 |
-
"negative_prompt": "digital, sharp lines, solid colors",
|
| 72 |
-
},
|
| 73 |
-
{
|
| 74 |
-
"name": "Cinematic",
|
| 75 |
-
"prompt": "cinematic collage of {prompt}. film stills, movie posters, dramatic lighting",
|
| 76 |
-
"negative_prompt": "static, lifeless, mundane",
|
| 77 |
-
},
|
| 78 |
-
{
|
| 79 |
-
"name": "Nostalgic",
|
| 80 |
-
"prompt": "nostalgic collage of {prompt}. retro imagery, vintage objects, sentimental journey",
|
| 81 |
-
"negative_prompt": "contemporary, futuristic, forward-looking",
|
| 82 |
-
},
|
| 83 |
-
{
|
| 84 |
-
"name": "Vintage",
|
| 85 |
-
"prompt": "vintage collage of {prompt}. aged paper, sepia tones, retro imagery, antique vibes",
|
| 86 |
-
"negative_prompt": "modern, contemporary, futuristic, high-tech",
|
| 87 |
-
},
|
| 88 |
-
{
|
| 89 |
-
"name": "Scrapbook",
|
| 90 |
-
"prompt": "scrapbook style collage of {prompt}. mixed media, hand-cut elements, textures, paper, stickers, doodles",
|
| 91 |
-
"negative_prompt": "clean, digital, modern, low quality",
|
| 92 |
-
},
|
| 93 |
-
{
|
| 94 |
-
"name": "NeoNGlow",
|
| 95 |
-
"prompt": "neon glow collage of {prompt}. vibrant colors, glowing effects, futuristic vibes",
|
| 96 |
-
"negative_prompt": "dull, muted colors, vintage, retro",
|
| 97 |
-
},
|
| 98 |
-
{
|
| 99 |
-
"name": "Geometric",
|
| 100 |
-
"prompt": "geometric collage of {prompt}. abstract shapes, colorful, sharp edges, modern design, high quality",
|
| 101 |
-
"negative_prompt": "blurry, low quality, traditional, dull",
|
| 102 |
-
},
|
| 103 |
-
{
|
| 104 |
-
"name": "Thematic",
|
| 105 |
-
"prompt": "thematic collage of {prompt}. cohesive theme, well-organized, matching colors, creative layout",
|
| 106 |
-
"negative_prompt": "random, messy, unorganized, clashing colors",
|
| 107 |
-
},
|
| 108 |
-
{
|
| 109 |
-
"name": "Cherry",
|
| 110 |
-
"prompt": "Duotone style Cherry tone applied to {prompt}",
|
| 111 |
-
"negative_prompt": "",
|
| 112 |
-
},
|
| 113 |
-
{
|
| 114 |
-
"name": "Fuchsia",
|
| 115 |
-
"prompt": "Duotone style Fuchsia tone applied to {prompt}",
|
| 116 |
-
"negative_prompt": "",
|
| 117 |
-
},
|
| 118 |
-
{
|
| 119 |
-
"name": "Pop",
|
| 120 |
-
"prompt": "Duotone style Pop tone applied to {prompt}",
|
| 121 |
-
"negative_prompt": "",
|
| 122 |
-
},
|
| 123 |
-
{
|
| 124 |
-
"name": "Violet",
|
| 125 |
-
"prompt": "Duotone style Violet applied to {prompt}",
|
| 126 |
-
"negative_prompt": "",
|
| 127 |
-
},
|
| 128 |
-
{
|
| 129 |
-
"name": "Sea Blue",
|
| 130 |
-
"prompt": "Duotone style Sea Blue applied to {prompt}",
|
| 131 |
-
"negative_prompt": "",
|
| 132 |
-
},
|
| 133 |
-
{
|
| 134 |
-
"name": "Sea Green",
|
| 135 |
-
"prompt": "Duotone style Sea Green applied to {prompt}",
|
| 136 |
-
"negative_prompt": "",
|
| 137 |
-
},
|
| 138 |
-
{
|
| 139 |
-
"name": "Mustard",
|
| 140 |
-
"prompt": "Duotone style Mustard applied to {prompt}",
|
| 141 |
-
"negative_prompt": "",
|
| 142 |
-
},
|
| 143 |
-
{
|
| 144 |
-
"name": "Amber",
|
| 145 |
-
"prompt": "Duotone style Amber applied to {prompt}",
|
| 146 |
-
"negative_prompt": "",
|
| 147 |
-
},
|
| 148 |
-
{
|
| 149 |
-
"name": "Pomelo",
|
| 150 |
-
"prompt": "Duotone style Pomelo applied to {prompt}",
|
| 151 |
-
"negative_prompt": "",
|
| 152 |
-
},
|
| 153 |
-
{
|
| 154 |
-
"name": "Peppermint",
|
| 155 |
-
"prompt": "Duotone style Peppermint applied to {prompt}",
|
| 156 |
-
"negative_prompt": "",
|
| 157 |
-
},
|
| 158 |
-
{
|
| 159 |
-
"name": "Mystic",
|
| 160 |
-
"prompt": "Duotone style Mystic tone applied to {prompt}",
|
| 161 |
-
"negative_prompt": "",
|
| 162 |
-
},
|
| 163 |
-
{
|
| 164 |
-
"name": "Pastel",
|
| 165 |
-
"prompt": "Duotone style Pastel applied to {prompt}",
|
| 166 |
-
"negative_prompt": "",
|
| 167 |
-
},
|
| 168 |
-
{
|
| 169 |
-
"name": "Coral",
|
| 170 |
-
"prompt": "Duotone style Coral applied to {prompt}",
|
| 171 |
-
"negative_prompt": "",
|
| 172 |
-
},
|
| 173 |
-
{
|
| 174 |
-
"name": "No Style",
|
| 175 |
-
"prompt": "{prompt}",
|
| 176 |
-
"negative_prompt": "",
|
| 177 |
-
},
|
| 178 |
-
]
|
| 179 |
-
|
| 180 |
-
# Filters--------------------------------------------------------------------
|
| 181 |
-
filters = {
|
| 182 |
-
"Vivid": {
|
| 183 |
-
"prompt": "extra vivid {prompt}",
|
| 184 |
-
"negative_prompt": "washed out, dull"
|
| 185 |
-
},
|
| 186 |
-
"Playa": {
|
| 187 |
-
"prompt": "{prompt} set in a vast playa",
|
| 188 |
-
"negative_prompt": "forest, mountains"
|
| 189 |
-
},
|
| 190 |
-
"Desert": {
|
| 191 |
-
"prompt": "{prompt} set in a desert landscape",
|
| 192 |
-
"negative_prompt": "ocean, city"
|
| 193 |
-
},
|
| 194 |
-
"West": {
|
| 195 |
-
"prompt": "{prompt} with a western theme",
|
| 196 |
-
"negative_prompt": "eastern, modern"
|
| 197 |
-
},
|
| 198 |
-
"Blush": {
|
| 199 |
-
"prompt": "{prompt} with a soft blush color palette",
|
| 200 |
-
"negative_prompt": "harsh colors, neon"
|
| 201 |
-
},
|
| 202 |
-
"Minimalist": {
|
| 203 |
-
"prompt": "{prompt} with a minimalist design",
|
| 204 |
-
"negative_prompt": "cluttered, ornate"
|
| 205 |
-
},
|
| 206 |
-
"Zero filter": {
|
| 207 |
-
"prompt": "{prompt}",
|
| 208 |
-
"negative_prompt": ""
|
| 209 |
-
},
|
| 210 |
-
}
|
| 211 |
-
|
| 212 |
-
styles = {k["name"]: (k["prompt"], k["negative_prompt"]) for k in style_list}
|
| 213 |
-
collage_styles = {k["name"]: (k["prompt"], k["negative_prompt"]) for k in collage_style_list}
|
| 214 |
-
filter_styles = {k: (v["prompt"], v["negative_prompt"]) for k, v in filters.items()}
|
| 215 |
-
|
| 216 |
-
STYLE_NAMES = list(styles.keys())
|
| 217 |
-
COLLAGE_STYLE_NAMES = list(collage_styles.keys())
|
| 218 |
-
FILTER_NAMES = list(filters.keys())
|
| 219 |
-
DEFAULT_STYLE_NAME = "3840 x 2160"
|
| 220 |
-
DEFAULT_COLLAGE_STYLE_NAME = "Hi-Res"
|
| 221 |
-
DEFAULT_FILTER_NAME = "Zero filter"
|
| 222 |
-
|
| 223 |
-
def apply_style(style_name: str, positive: str, negative: str = "") -> Tuple[str, str]:
|
| 224 |
-
if style_name in styles:
|
| 225 |
-
p, n = styles.get(style_name, styles[DEFAULT_STYLE_NAME])
|
| 226 |
-
elif style_name in collage_styles:
|
| 227 |
-
p, n = collage_styles.get(style_name, collage_styles[DEFAULT_COLLAGE_STYLE_NAME])
|
| 228 |
-
elif style_name in filter_styles:
|
| 229 |
-
p, n = filter_styles.get(style_name, filter_styles[DEFAULT_FILTER_NAME])
|
| 230 |
-
else:
|
| 231 |
-
p, n = styles[DEFAULT_STYLE_NAME]
|
| 232 |
-
|
| 233 |
-
if not negative:
|
| 234 |
-
negative = ""
|
| 235 |
-
return p.replace("{prompt}", positive), n + negative
|
| 236 |
-
|
| 237 |
-
if not torch.cuda.is_available():
|
| 238 |
-
DESCRIPTION = "\n<p>⚠️Running on CPU, This may not work on CPU.</p>"
|
| 239 |
|
| 240 |
MAX_SEED = np.iinfo(np.int32).max
|
| 241 |
CACHE_EXAMPLES = torch.cuda.is_available() and os.getenv("CACHE_EXAMPLES", "0") == "1"
|
|
@@ -243,53 +92,46 @@ MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "2048"))
|
|
| 243 |
USE_TORCH_COMPILE = os.getenv("USE_TORCH_COMPILE", "0") == "1"
|
| 244 |
ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD", "0") == "1"
|
| 245 |
|
| 246 |
-
|
| 247 |
-
# Set dtype based on device: half for CUDA, float32 for CPU
|
| 248 |
dtype = torch.float16 if device.type == "cuda" else torch.float32
|
| 249 |
|
| 250 |
-
# Load
|
| 251 |
if torch.cuda.is_available():
|
|
|
|
| 252 |
pipe = StableDiffusionXLPipeline.from_pretrained(
|
| 253 |
-
#"SG161222/RealVisXL_V5.0",
|
| 254 |
"SG161222/RealVisXL_V5.0_Lightning",
|
| 255 |
torch_dtype=dtype,
|
| 256 |
use_safetensors=True,
|
| 257 |
add_watermarker=False
|
| 258 |
).to(device)
|
| 259 |
-
# Ensure text encoder uses half precision on GPU
|
| 260 |
pipe.text_encoder = pipe.text_encoder.half()
|
| 261 |
-
|
| 262 |
if ENABLE_CPU_OFFLOAD:
|
| 263 |
pipe.enable_model_cpu_offload()
|
| 264 |
else:
|
| 265 |
pipe.to(device)
|
| 266 |
print("Loaded RealVisXL_V5.0_Lightning on Device!")
|
| 267 |
-
|
| 268 |
if USE_TORCH_COMPILE:
|
| 269 |
pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
|
| 270 |
print("Model RealVisXL_V5.0_Lightning Compiled!")
|
| 271 |
|
| 272 |
-
#
|
| 273 |
pipe2 = StableDiffusionXLPipeline.from_pretrained(
|
| 274 |
-
#"SG161222/RealVisXL_V4.0",
|
| 275 |
"SG161222/RealVisXL_V4.0_Lightning",
|
| 276 |
torch_dtype=dtype,
|
| 277 |
use_safetensors=True,
|
| 278 |
add_watermarker=False,
|
| 279 |
).to(device)
|
| 280 |
pipe2.text_encoder = pipe2.text_encoder.half()
|
| 281 |
-
|
| 282 |
if ENABLE_CPU_OFFLOAD:
|
| 283 |
pipe2.enable_model_cpu_offload()
|
| 284 |
else:
|
| 285 |
pipe2.to(device)
|
| 286 |
print("Loaded RealVisXL_V4.0 on Device!")
|
| 287 |
-
|
| 288 |
if USE_TORCH_COMPILE:
|
| 289 |
pipe2.unet = torch.compile(pipe2.unet, mode="reduce-overhead", fullgraph=True)
|
| 290 |
print("Model RealVisXL_V4.0 Compiled!")
|
| 291 |
|
| 292 |
-
#
|
| 293 |
pipe3 = StableDiffusionXLPipeline.from_pretrained(
|
| 294 |
"SG161222/RealVisXL_V3.0_Turbo",
|
| 295 |
torch_dtype=dtype,
|
|
@@ -297,18 +139,15 @@ if torch.cuda.is_available():
|
|
| 297 |
add_watermarker=False,
|
| 298 |
).to(device)
|
| 299 |
pipe3.text_encoder = pipe3.text_encoder.half()
|
| 300 |
-
|
| 301 |
if ENABLE_CPU_OFFLOAD:
|
| 302 |
pipe3.enable_model_cpu_offload()
|
| 303 |
else:
|
| 304 |
pipe3.to(device)
|
| 305 |
print("Loaded Animagine XL 4.0 on Device!")
|
| 306 |
-
|
| 307 |
if USE_TORCH_COMPILE:
|
| 308 |
pipe3.unet = torch.compile(pipe3.unet, mode="reduce-overhead", fullgraph=True)
|
| 309 |
print("Model Animagine XL 4.0 Compiled!")
|
| 310 |
else:
|
| 311 |
-
# On CPU, load all models in float32
|
| 312 |
pipe = StableDiffusionXLPipeline.from_pretrained(
|
| 313 |
"SG161222/RealVisXL_V5.0_Lightning",
|
| 314 |
torch_dtype=dtype,
|
|
@@ -329,7 +168,7 @@ else:
|
|
| 329 |
).to(device)
|
| 330 |
print("Running on CPU; models loaded in float32.")
|
| 331 |
|
| 332 |
-
#
|
| 333 |
DEFAULT_MODEL = "Lightning 5"
|
| 334 |
MODEL_CHOICES = [DEFAULT_MODEL, "Lightning 4", "Turbo v3"]
|
| 335 |
models = {
|
|
@@ -338,266 +177,232 @@ models = {
|
|
| 338 |
"Turbo v3": pipe3
|
| 339 |
}
|
| 340 |
|
| 341 |
-
def
|
| 342 |
-
|
| 343 |
-
|
| 344 |
-
def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
|
| 345 |
-
if randomize_seed:
|
| 346 |
-
seed = random.randint(0, MAX_SEED)
|
| 347 |
-
return seed
|
| 348 |
-
|
| 349 |
-
@spaces.GPU
|
| 350 |
-
def generate(
|
| 351 |
-
prompt: str,
|
| 352 |
-
negative_prompt: str = "",
|
| 353 |
-
use_negative_prompt: bool = False,
|
| 354 |
-
style: str = DEFAULT_STYLE_NAME,
|
| 355 |
-
collage_style: str = DEFAULT_COLLAGE_STYLE_NAME,
|
| 356 |
-
filter_name: str = DEFAULT_FILTER_NAME,
|
| 357 |
-
grid_size: str = "2x2",
|
| 358 |
-
seed: int = 0,
|
| 359 |
-
width: int = 1024,
|
| 360 |
-
height: int = 1024,
|
| 361 |
-
guidance_scale: float = 3,
|
| 362 |
-
randomize_seed: bool = False,
|
| 363 |
-
model_choice: str = DEFAULT_MODEL,
|
| 364 |
-
use_resolution_binning: bool = True,
|
| 365 |
-
progress=gr.Progress(track_tqdm=True),
|
| 366 |
-
):
|
| 367 |
-
if check_text(prompt, negative_prompt):
|
| 368 |
raise ValueError("Prompt contains restricted words.")
|
| 369 |
|
| 370 |
-
if collage_style != "No Style":
|
| 371 |
-
prompt, negative_prompt = apply_style(collage_style, prompt, negative_prompt)
|
| 372 |
-
elif filter_name != "No Filter":
|
| 373 |
-
prompt, negative_prompt = apply_style(filter_name, prompt, negative_prompt)
|
| 374 |
-
else:
|
| 375 |
-
prompt, negative_prompt = apply_style(style, prompt, negative_prompt)
|
| 376 |
-
|
| 377 |
seed = int(randomize_seed_fn(seed, randomize_seed))
|
| 378 |
generator = torch.Generator(device=device).manual_seed(seed)
|
| 379 |
|
| 380 |
-
|
| 381 |
-
negative_prompt = ""
|
| 382 |
-
negative_prompt += default_negative
|
| 383 |
-
|
| 384 |
grid_sizes = {
|
| 385 |
"2x1": (2, 1),
|
| 386 |
"1x2": (1, 2),
|
| 387 |
"2x2": (2, 2),
|
| 388 |
-
"2x3": (2, 3),
|
| 389 |
-
"3x2": (3, 2),
|
| 390 |
"1x1": (1, 1)
|
| 391 |
}
|
| 392 |
-
|
| 393 |
-
|
| 394 |
-
num_images = grid_size_x * grid_size_y
|
| 395 |
|
| 396 |
options = {
|
| 397 |
"prompt": prompt,
|
| 398 |
-
"negative_prompt":
|
| 399 |
"width": width,
|
| 400 |
"height": height,
|
| 401 |
"guidance_scale": guidance_scale,
|
| 402 |
"num_inference_steps": 30,
|
| 403 |
"generator": generator,
|
| 404 |
"num_images_per_prompt": num_images,
|
| 405 |
-
"use_resolution_binning":
|
| 406 |
"output_type": "pil",
|
| 407 |
}
|
| 408 |
|
| 409 |
if device.type == "cuda":
|
| 410 |
torch.cuda.empty_cache()
|
| 411 |
|
| 412 |
-
# Choose pipeline based on user selection
|
| 413 |
selected_pipe = models.get(model_choice, pipe)
|
| 414 |
images = selected_pipe(**options).images
|
| 415 |
|
| 416 |
-
|
|
|
|
| 417 |
for i, img in enumerate(images[:num_images]):
|
| 418 |
-
grid_img.paste(img, (i %
|
| 419 |
|
| 420 |
unique_name = str(uuid.uuid4()) + ".png"
|
| 421 |
-
|
| 422 |
return [unique_name], seed
|
| 423 |
|
| 424 |
-
|
| 425 |
-
|
| 426 |
-
|
| 427 |
-
|
| 428 |
-
|
| 429 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 430 |
|
| 431 |
css = '''
|
| 432 |
-
.gradio-container {
|
| 433 |
-
max-width: 888px !important;
|
| 434 |
-
margin: 0 auto !important;
|
| 435 |
-
display: flex;
|
| 436 |
-
flex-direction: column;
|
| 437 |
-
align-items: center;
|
| 438 |
-
}
|
| 439 |
h1 {
|
| 440 |
-
|
|
|
|
| 441 |
}
|
| 442 |
-
'''
|
| 443 |
|
| 444 |
-
|
| 445 |
-
|
| 446 |
-
|
| 447 |
-
|
| 448 |
-
|
| 449 |
-
|
| 450 |
-
|
| 451 |
|
| 452 |
-
|
| 453 |
-
|
| 454 |
-
|
| 455 |
-
|
| 456 |
-
|
| 457 |
-
|
| 458 |
-
|
| 459 |
-
|
| 460 |
-
|
| 461 |
-
|
| 462 |
-
|
| 463 |
-
|
| 464 |
-
|
| 465 |
-
|
| 466 |
-
|
| 467 |
-
|
| 468 |
-
|
| 469 |
-
|
| 470 |
-
|
| 471 |
-
|
| 472 |
-
|
| 473 |
-
|
| 474 |
-
|
| 475 |
-
|
| 476 |
-
|
| 477 |
-
|
| 478 |
-
filter_selection = gr.Dropdown(
|
| 479 |
-
show_label=True,
|
| 480 |
-
container=True,
|
| 481 |
-
interactive=True,
|
| 482 |
-
choices=FILTER_NAMES,
|
| 483 |
-
value=DEFAULT_FILTER_NAME,
|
| 484 |
-
label="Filter Type",
|
| 485 |
-
)
|
| 486 |
-
with gr.Row(visible=True):
|
| 487 |
-
collage_style_selection = gr.Dropdown(
|
| 488 |
-
show_label=True,
|
| 489 |
-
container=True,
|
| 490 |
-
interactive=True,
|
| 491 |
-
choices=COLLAGE_STYLE_NAMES,
|
| 492 |
-
value=DEFAULT_COLLAGE_STYLE_NAME,
|
| 493 |
-
label="Collage Template + Duotone Canvas",
|
| 494 |
-
)
|
| 495 |
-
with gr.Row(visible=True):
|
| 496 |
-
style_selection = gr.Dropdown(
|
| 497 |
-
show_label=True,
|
| 498 |
-
container=True,
|
| 499 |
-
interactive=True,
|
| 500 |
-
choices=STYLE_NAMES,
|
| 501 |
-
value=DEFAULT_STYLE_NAME,
|
| 502 |
-
label="Quality Style",
|
| 503 |
-
)
|
| 504 |
-
with gr.Accordion("Advanced options", open=False):
|
| 505 |
-
use_negative_prompt = gr.Checkbox(label="Use negative prompt", value=True, visible=True)
|
| 506 |
-
negative_prompt = gr.Text(
|
| 507 |
-
label="Negative prompt",
|
| 508 |
-
max_lines=1,
|
| 509 |
-
placeholder="Enter a negative prompt",
|
| 510 |
-
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",
|
| 511 |
-
visible=True,
|
| 512 |
-
)
|
| 513 |
-
with gr.Row():
|
| 514 |
-
num_inference_steps = gr.Slider(
|
| 515 |
-
label="Steps",
|
| 516 |
-
minimum=10,
|
| 517 |
-
maximum=60,
|
| 518 |
-
step=1,
|
| 519 |
-
value=30,
|
| 520 |
-
)
|
| 521 |
-
with gr.Row():
|
| 522 |
-
num_images_per_prompt = gr.Slider(
|
| 523 |
-
label="Images",
|
| 524 |
-
minimum=1,
|
| 525 |
-
maximum=5,
|
| 526 |
-
step=1,
|
| 527 |
-
value=2,
|
| 528 |
-
)
|
| 529 |
-
seed = gr.Slider(
|
| 530 |
-
label="Seed",
|
| 531 |
-
minimum=0,
|
| 532 |
-
maximum=MAX_SEED,
|
| 533 |
-
step=1,
|
| 534 |
-
value=0,
|
| 535 |
-
visible=True
|
| 536 |
-
)
|
| 537 |
-
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
| 538 |
-
with gr.Row(visible=True):
|
| 539 |
-
width = gr.Slider(
|
| 540 |
-
label="Width",
|
| 541 |
-
minimum=512,
|
| 542 |
-
maximum=2048,
|
| 543 |
-
step=64,
|
| 544 |
-
value=1024,
|
| 545 |
-
)
|
| 546 |
-
height = gr.Slider(
|
| 547 |
-
label="Height",
|
| 548 |
-
minimum=512,
|
| 549 |
-
maximum=2048,
|
| 550 |
-
step=64,
|
| 551 |
-
value=1024,
|
| 552 |
-
)
|
| 553 |
-
with gr.Row():
|
| 554 |
-
guidance_scale = gr.Slider(
|
| 555 |
-
label="Guidance Scale",
|
| 556 |
-
minimum=0.1,
|
| 557 |
-
maximum=20.0,
|
| 558 |
-
step=0.1,
|
| 559 |
-
value=6,
|
| 560 |
-
)
|
| 561 |
-
with gr.Column(scale=2):
|
| 562 |
-
result = gr.Gallery(label="Result", columns=1, show_label=False)
|
| 563 |
-
gr.Examples(
|
| 564 |
-
examples=examples,
|
| 565 |
-
inputs=prompt,
|
| 566 |
-
outputs=[result, seed],
|
| 567 |
-
fn=generate,
|
| 568 |
-
cache_examples=CACHE_EXAMPLES,
|
| 569 |
-
)
|
| 570 |
-
use_negative_prompt.change(
|
| 571 |
-
fn=lambda x: gr.update(visible=x),
|
| 572 |
-
inputs=use_negative_prompt,
|
| 573 |
-
outputs=negative_prompt,
|
| 574 |
-
api_name=False,
|
| 575 |
-
)
|
| 576 |
-
gr.on(
|
| 577 |
-
triggers=[
|
| 578 |
-
prompt.submit,
|
| 579 |
-
negative_prompt.submit,
|
| 580 |
-
run_button.click,
|
| 581 |
-
],
|
| 582 |
-
fn=generate,
|
| 583 |
-
inputs=[
|
| 584 |
-
prompt,
|
| 585 |
-
negative_prompt,
|
| 586 |
-
use_negative_prompt,
|
| 587 |
-
style_selection,
|
| 588 |
-
collage_style_selection,
|
| 589 |
-
filter_selection,
|
| 590 |
-
grid_size_selection,
|
| 591 |
-
seed,
|
| 592 |
-
width,
|
| 593 |
-
height,
|
| 594 |
-
guidance_scale,
|
| 595 |
-
randomize_seed,
|
| 596 |
-
model_selection,
|
| 597 |
-
],
|
| 598 |
-
outputs=[result, seed],
|
| 599 |
-
api_name="run",
|
| 600 |
-
)
|
| 601 |
|
| 602 |
if __name__ == "__main__":
|
| 603 |
-
demo.queue(max_size=
|
|
|
|
| 2 |
import random
|
| 3 |
import uuid
|
| 4 |
import json
|
| 5 |
+
import time
|
| 6 |
+
import asyncio
|
| 7 |
+
import re
|
| 8 |
+
from threading import Thread
|
| 9 |
+
|
| 10 |
import gradio as gr
|
|
|
|
|
|
|
| 11 |
import spaces
|
| 12 |
import torch
|
| 13 |
+
import numpy as np
|
| 14 |
+
from PIL import Image
|
| 15 |
+
import edge_tts
|
| 16 |
+
|
| 17 |
+
from transformers import (
|
| 18 |
+
AutoModelForCausalLM,
|
| 19 |
+
AutoTokenizer,
|
| 20 |
+
TextIteratorStreamer,
|
| 21 |
+
Qwen2VLForConditionalGeneration,
|
| 22 |
+
AutoProcessor,
|
| 23 |
+
)
|
| 24 |
+
from transformers.image_utils import load_image
|
| 25 |
+
from diffusers import StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler
|
| 26 |
+
|
| 27 |
+
MAX_MAX_NEW_TOKENS = 2048
|
| 28 |
+
DEFAULT_MAX_NEW_TOKENS = 1024
|
| 29 |
+
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
|
| 30 |
+
|
| 31 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
| 32 |
+
|
| 33 |
+
# Load text-only model and tokenizer for chat generation
|
| 34 |
+
model_id = "prithivMLmods/FastThink-0.5B-Tiny"
|
| 35 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 36 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 37 |
+
model_id,
|
| 38 |
+
device_map="auto",
|
| 39 |
+
torch_dtype=torch.bfloat16,
|
| 40 |
+
)
|
| 41 |
+
model.eval()
|
| 42 |
|
| 43 |
+
# TTS Voices and processor for multimodal chat
|
| 44 |
+
TTS_VOICES = [
|
| 45 |
+
"en-US-JennyNeural", # @tts1
|
| 46 |
+
"en-US-GuyNeural", # @tts2
|
| 47 |
+
]
|
| 48 |
+
MODEL_ID_VL = "prithivMLmods/Qwen2-VL-OCR-2B-Instruct"
|
| 49 |
+
processor = AutoProcessor.from_pretrained(MODEL_ID_VL, trust_remote_code=True)
|
| 50 |
+
model_m = Qwen2VLForConditionalGeneration.from_pretrained(
|
| 51 |
+
MODEL_ID_VL,
|
| 52 |
+
trust_remote_code=True,
|
| 53 |
+
torch_dtype=torch.float16
|
| 54 |
+
).to("cuda").eval()
|
| 55 |
+
|
| 56 |
+
# A helper function to convert text to speech via Edge TTS
|
| 57 |
+
async def text_to_speech(text: str, voice: str, output_file="output.mp3"):
|
| 58 |
+
communicate = edge_tts.Communicate(text, voice)
|
| 59 |
+
await communicate.save(output_file)
|
| 60 |
+
return output_file
|
| 61 |
+
|
| 62 |
+
def clean_chat_history(chat_history):
|
| 63 |
+
cleaned = []
|
| 64 |
+
for msg in chat_history:
|
| 65 |
+
if isinstance(msg, dict) and isinstance(msg.get("content"), str):
|
| 66 |
+
cleaned.append(msg)
|
| 67 |
+
return cleaned
|
| 68 |
+
|
| 69 |
+
# Restricted words check (if any)
|
| 70 |
bad_words = json.loads(os.getenv('BAD_WORDS', "[]"))
|
| 71 |
bad_words_negative = json.loads(os.getenv('BAD_WORDS_NEGATIVE', "[]"))
|
| 72 |
default_negative = os.getenv("default_negative", "")
|
|
|
|
| 80 |
return True
|
| 81 |
return False
|
| 82 |
|
| 83 |
+
# Use the same random seed function for both text and image generation
|
| 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 |
MAX_SEED = np.iinfo(np.int32).max
|
| 90 |
CACHE_EXAMPLES = torch.cuda.is_available() and os.getenv("CACHE_EXAMPLES", "0") == "1"
|
|
|
|
| 92 |
USE_TORCH_COMPILE = os.getenv("USE_TORCH_COMPILE", "0") == "1"
|
| 93 |
ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD", "0") == "1"
|
| 94 |
|
| 95 |
+
# Set dtype based on device: use half for CUDA, float32 otherwise.
|
|
|
|
| 96 |
dtype = torch.float16 if device.type == "cuda" else torch.float32
|
| 97 |
|
| 98 |
+
# Load image generation pipelines for the three model choices.
|
| 99 |
if torch.cuda.is_available():
|
| 100 |
+
# Lightning 5 model
|
| 101 |
pipe = StableDiffusionXLPipeline.from_pretrained(
|
|
|
|
| 102 |
"SG161222/RealVisXL_V5.0_Lightning",
|
| 103 |
torch_dtype=dtype,
|
| 104 |
use_safetensors=True,
|
| 105 |
add_watermarker=False
|
| 106 |
).to(device)
|
|
|
|
| 107 |
pipe.text_encoder = pipe.text_encoder.half()
|
|
|
|
| 108 |
if ENABLE_CPU_OFFLOAD:
|
| 109 |
pipe.enable_model_cpu_offload()
|
| 110 |
else:
|
| 111 |
pipe.to(device)
|
| 112 |
print("Loaded RealVisXL_V5.0_Lightning on Device!")
|
|
|
|
| 113 |
if USE_TORCH_COMPILE:
|
| 114 |
pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
|
| 115 |
print("Model RealVisXL_V5.0_Lightning Compiled!")
|
| 116 |
|
| 117 |
+
# Lightning 4 model
|
| 118 |
pipe2 = StableDiffusionXLPipeline.from_pretrained(
|
|
|
|
| 119 |
"SG161222/RealVisXL_V4.0_Lightning",
|
| 120 |
torch_dtype=dtype,
|
| 121 |
use_safetensors=True,
|
| 122 |
add_watermarker=False,
|
| 123 |
).to(device)
|
| 124 |
pipe2.text_encoder = pipe2.text_encoder.half()
|
|
|
|
| 125 |
if ENABLE_CPU_OFFLOAD:
|
| 126 |
pipe2.enable_model_cpu_offload()
|
| 127 |
else:
|
| 128 |
pipe2.to(device)
|
| 129 |
print("Loaded RealVisXL_V4.0 on Device!")
|
|
|
|
| 130 |
if USE_TORCH_COMPILE:
|
| 131 |
pipe2.unet = torch.compile(pipe2.unet, mode="reduce-overhead", fullgraph=True)
|
| 132 |
print("Model RealVisXL_V4.0 Compiled!")
|
| 133 |
|
| 134 |
+
# Turbo v3 model
|
| 135 |
pipe3 = StableDiffusionXLPipeline.from_pretrained(
|
| 136 |
"SG161222/RealVisXL_V3.0_Turbo",
|
| 137 |
torch_dtype=dtype,
|
|
|
|
| 139 |
add_watermarker=False,
|
| 140 |
).to(device)
|
| 141 |
pipe3.text_encoder = pipe3.text_encoder.half()
|
|
|
|
| 142 |
if ENABLE_CPU_OFFLOAD:
|
| 143 |
pipe3.enable_model_cpu_offload()
|
| 144 |
else:
|
| 145 |
pipe3.to(device)
|
| 146 |
print("Loaded Animagine XL 4.0 on Device!")
|
|
|
|
| 147 |
if USE_TORCH_COMPILE:
|
| 148 |
pipe3.unet = torch.compile(pipe3.unet, mode="reduce-overhead", fullgraph=True)
|
| 149 |
print("Model Animagine XL 4.0 Compiled!")
|
| 150 |
else:
|
|
|
|
| 151 |
pipe = StableDiffusionXLPipeline.from_pretrained(
|
| 152 |
"SG161222/RealVisXL_V5.0_Lightning",
|
| 153 |
torch_dtype=dtype,
|
|
|
|
| 168 |
).to(device)
|
| 169 |
print("Running on CPU; models loaded in float32.")
|
| 170 |
|
| 171 |
+
# Define available model choices and their mapping.
|
| 172 |
DEFAULT_MODEL = "Lightning 5"
|
| 173 |
MODEL_CHOICES = [DEFAULT_MODEL, "Lightning 4", "Turbo v3"]
|
| 174 |
models = {
|
|
|
|
| 177 |
"Turbo v3": pipe3
|
| 178 |
}
|
| 179 |
|
| 180 |
+
def generate_image_grid(prompt: str, seed: int, grid_size: str, width: int, height: int,
|
| 181 |
+
guidance_scale: float, randomize_seed: bool, model_choice: str):
|
| 182 |
+
if check_text(prompt, ""):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 183 |
raise ValueError("Prompt contains restricted words.")
|
| 184 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 185 |
seed = int(randomize_seed_fn(seed, randomize_seed))
|
| 186 |
generator = torch.Generator(device=device).manual_seed(seed)
|
| 187 |
|
| 188 |
+
# Define supported grid sizes.
|
|
|
|
|
|
|
|
|
|
| 189 |
grid_sizes = {
|
| 190 |
"2x1": (2, 1),
|
| 191 |
"1x2": (1, 2),
|
| 192 |
"2x2": (2, 2),
|
|
|
|
|
|
|
| 193 |
"1x1": (1, 1)
|
| 194 |
}
|
| 195 |
+
grid_size_tuple = grid_sizes.get(grid_size, (1, 1))
|
| 196 |
+
num_images = grid_size_tuple[0] * grid_size_tuple[1]
|
|
|
|
| 197 |
|
| 198 |
options = {
|
| 199 |
"prompt": prompt,
|
| 200 |
+
"negative_prompt": default_negative,
|
| 201 |
"width": width,
|
| 202 |
"height": height,
|
| 203 |
"guidance_scale": guidance_scale,
|
| 204 |
"num_inference_steps": 30,
|
| 205 |
"generator": generator,
|
| 206 |
"num_images_per_prompt": num_images,
|
| 207 |
+
"use_resolution_binning": True,
|
| 208 |
"output_type": "pil",
|
| 209 |
}
|
| 210 |
|
| 211 |
if device.type == "cuda":
|
| 212 |
torch.cuda.empty_cache()
|
| 213 |
|
|
|
|
| 214 |
selected_pipe = models.get(model_choice, pipe)
|
| 215 |
images = selected_pipe(**options).images
|
| 216 |
|
| 217 |
+
# Create a grid image.
|
| 218 |
+
grid_img = Image.new('RGB', (width * grid_size_tuple[0], height * grid_size_tuple[1]))
|
| 219 |
for i, img in enumerate(images[:num_images]):
|
| 220 |
+
grid_img.paste(img, ((i % grid_size_tuple[0]) * width, (i // grid_size_tuple[0]) * height))
|
| 221 |
|
| 222 |
unique_name = str(uuid.uuid4()) + ".png"
|
| 223 |
+
grid_img.save(unique_name)
|
| 224 |
return [unique_name], seed
|
| 225 |
|
| 226 |
+
# -----------------------------
|
| 227 |
+
# Main generate() Function
|
| 228 |
+
# -----------------------------
|
| 229 |
+
@spaces.GPU
|
| 230 |
+
def generate(
|
| 231 |
+
input_dict: dict,
|
| 232 |
+
chat_history: list[dict],
|
| 233 |
+
max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
|
| 234 |
+
temperature: float = 0.6,
|
| 235 |
+
top_p: float = 0.9,
|
| 236 |
+
top_k: int = 50,
|
| 237 |
+
repetition_penalty: float = 1.2,
|
| 238 |
+
):
|
| 239 |
+
text = input_dict["text"]
|
| 240 |
+
files = input_dict.get("files", [])
|
| 241 |
+
|
| 242 |
+
lower_text = text.lower().strip()
|
| 243 |
+
# Check if the prompt is an image generation command using model flags.
|
| 244 |
+
if (lower_text.startswith("@lightningv5") or
|
| 245 |
+
lower_text.startswith("@lightningv4") or
|
| 246 |
+
lower_text.startswith("@turbov3")):
|
| 247 |
+
|
| 248 |
+
# Determine model choice based on flag.
|
| 249 |
+
model_choice = None
|
| 250 |
+
if "@lightningv5" in lower_text:
|
| 251 |
+
model_choice = "Lightning 5"
|
| 252 |
+
elif "@lightningv4" in lower_text:
|
| 253 |
+
model_choice = "Lightning 4"
|
| 254 |
+
elif "@turbov3" in lower_text:
|
| 255 |
+
model_choice = "Turbo v3"
|
| 256 |
+
|
| 257 |
+
# Parse grid size flag e.g. "@2x2"
|
| 258 |
+
grid_match = re.search(r"@(\d+x\d+)", lower_text)
|
| 259 |
+
grid_size = grid_match.group(1) if grid_match else "1x1"
|
| 260 |
+
|
| 261 |
+
# Remove the model and grid flags from the prompt.
|
| 262 |
+
prompt_clean = re.sub(r"@lightningv5", "", text, flags=re.IGNORECASE)
|
| 263 |
+
prompt_clean = re.sub(r"@lightningv4", "", prompt_clean, flags=re.IGNORECASE)
|
| 264 |
+
prompt_clean = re.sub(r"@turbov3", "", prompt_clean, flags=re.IGNORECASE)
|
| 265 |
+
prompt_clean = re.sub(r"@\d+x\d+", "", prompt_clean, flags=re.IGNORECASE)
|
| 266 |
+
prompt_clean = prompt_clean.strip().strip('"')
|
| 267 |
+
|
| 268 |
+
# Default parameters for image generation.
|
| 269 |
+
width = 1024
|
| 270 |
+
height = 1024
|
| 271 |
+
guidance_scale = 6.0
|
| 272 |
+
seed_val = 0
|
| 273 |
+
randomize_seed = True
|
| 274 |
+
use_resolution_binning = True
|
| 275 |
+
|
| 276 |
+
yield "Generating image grid..."
|
| 277 |
+
image_paths, used_seed = generate_image_grid(
|
| 278 |
+
prompt_clean,
|
| 279 |
+
seed_val,
|
| 280 |
+
grid_size,
|
| 281 |
+
width,
|
| 282 |
+
height,
|
| 283 |
+
guidance_scale,
|
| 284 |
+
randomize_seed,
|
| 285 |
+
model_choice,
|
| 286 |
+
)
|
| 287 |
+
yield gr.Image(image_paths[0])
|
| 288 |
+
return
|
| 289 |
+
|
| 290 |
+
# Otherwise, handle text/chat (and TTS) generation.
|
| 291 |
+
tts_prefix = "@tts"
|
| 292 |
+
is_tts = any(text.strip().lower().startswith(f"{tts_prefix}{i}") for i in range(1, 3))
|
| 293 |
+
voice_index = next((i for i in range(1, 3) if text.strip().lower().startswith(f"{tts_prefix}{i}")), None)
|
| 294 |
+
|
| 295 |
+
if is_tts and voice_index:
|
| 296 |
+
voice = TTS_VOICES[voice_index - 1]
|
| 297 |
+
text = text.replace(f"{tts_prefix}{voice_index}", "").strip()
|
| 298 |
+
conversation = [{"role": "user", "content": text}]
|
| 299 |
+
else:
|
| 300 |
+
voice = None
|
| 301 |
+
text = text.replace(tts_prefix, "").strip()
|
| 302 |
+
conversation = clean_chat_history(chat_history)
|
| 303 |
+
conversation.append({"role": "user", "content": text})
|
| 304 |
+
|
| 305 |
+
if files:
|
| 306 |
+
images = [load_image(image) for image in files] if len(files) > 1 else [load_image(files[0])]
|
| 307 |
+
messages = [{
|
| 308 |
+
"role": "user",
|
| 309 |
+
"content": [
|
| 310 |
+
*[{"type": "image", "image": image} for image in images],
|
| 311 |
+
{"type": "text", "text": text},
|
| 312 |
+
]
|
| 313 |
+
}]
|
| 314 |
+
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 315 |
+
inputs = processor(text=[prompt], images=images, return_tensors="pt", padding=True).to("cuda")
|
| 316 |
+
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
| 317 |
+
generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
|
| 318 |
+
thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
|
| 319 |
+
thread.start()
|
| 320 |
+
|
| 321 |
+
buffer = ""
|
| 322 |
+
yield "Thinking..."
|
| 323 |
+
for new_text in streamer:
|
| 324 |
+
buffer += new_text
|
| 325 |
+
buffer = buffer.replace("<|im_end|>", "")
|
| 326 |
+
time.sleep(0.01)
|
| 327 |
+
yield buffer
|
| 328 |
+
else:
|
| 329 |
+
input_ids = tokenizer.apply_chat_template(conversation, add_generation_prompt=True, return_tensors="pt")
|
| 330 |
+
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
| 331 |
+
input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
|
| 332 |
+
gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
|
| 333 |
+
input_ids = input_ids.to(model.device)
|
| 334 |
+
streamer = TextIteratorStreamer(tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
|
| 335 |
+
generation_kwargs = {
|
| 336 |
+
"input_ids": input_ids,
|
| 337 |
+
"streamer": streamer,
|
| 338 |
+
"max_new_tokens": max_new_tokens,
|
| 339 |
+
"do_sample": True,
|
| 340 |
+
"top_p": top_p,
|
| 341 |
+
"top_k": top_k,
|
| 342 |
+
"temperature": temperature,
|
| 343 |
+
"num_beams": 1,
|
| 344 |
+
"repetition_penalty": repetition_penalty,
|
| 345 |
+
}
|
| 346 |
+
t = Thread(target=model.generate, kwargs=generation_kwargs)
|
| 347 |
+
t.start()
|
| 348 |
+
|
| 349 |
+
outputs = []
|
| 350 |
+
for new_text in streamer:
|
| 351 |
+
outputs.append(new_text)
|
| 352 |
+
yield "".join(outputs)
|
| 353 |
+
|
| 354 |
+
final_response = "".join(outputs)
|
| 355 |
+
yield final_response
|
| 356 |
+
|
| 357 |
+
if is_tts and voice:
|
| 358 |
+
output_file = asyncio.run(text_to_speech(final_response, voice))
|
| 359 |
+
yield gr.Audio(output_file, autoplay=True)
|
| 360 |
+
|
| 361 |
+
|
| 362 |
+
DESCRIPTION = """
|
| 363 |
+
# IMAGINEO 4K ⚡
|
| 364 |
+
"""
|
| 365 |
|
| 366 |
css = '''
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 367 |
h1 {
|
| 368 |
+
text-align: center;
|
| 369 |
+
display: block;
|
| 370 |
}
|
|
|
|
| 371 |
|
| 372 |
+
#duplicate-button {
|
| 373 |
+
margin: auto;
|
| 374 |
+
color: #fff;
|
| 375 |
+
background: #1565c0;
|
| 376 |
+
border-radius: 100vh;
|
| 377 |
+
}
|
| 378 |
+
'''
|
| 379 |
|
| 380 |
+
demo = gr.ChatInterface(
|
| 381 |
+
fn=generate,
|
| 382 |
+
additional_inputs=[
|
| 383 |
+
gr.Slider(label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS),
|
| 384 |
+
gr.Slider(label="Temperature", minimum=0.1, maximum=4.0, step=0.1, value=0.6),
|
| 385 |
+
gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9),
|
| 386 |
+
gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50),
|
| 387 |
+
gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2),
|
| 388 |
+
],
|
| 389 |
+
examples=[
|
| 390 |
+
["@tts1 Who is Nikola Tesla, and why did he die?"],
|
| 391 |
+
['@lightningv5 @2x2 "Chocolate dripping from a donut against a yellow background, in the style of brocore, hyper-realistic"'],
|
| 392 |
+
['@lightningv4 @1x1 "A serene landscape with mountains"'],
|
| 393 |
+
['@turbov3 @2x1 "Abstract art, colorful and vibrant"'],
|
| 394 |
+
["Write a Python function to check if a number is prime."],
|
| 395 |
+
["@tts2 What causes rainbows to form?"],
|
| 396 |
+
],
|
| 397 |
+
cache_examples=False,
|
| 398 |
+
type="messages",
|
| 399 |
+
description=DESCRIPTION,
|
| 400 |
+
css=css,
|
| 401 |
+
fill_height=True,
|
| 402 |
+
textbox=gr.MultimodalTextbox(label="Query Input", file_types=["image"], file_count="multiple"),
|
| 403 |
+
stop_btn="Stop Generation",
|
| 404 |
+
multimodal=True,
|
| 405 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 406 |
|
| 407 |
if __name__ == "__main__":
|
| 408 |
+
demo.queue(max_size=20).launch(share=True)
|