File size: 17,970 Bytes
32dec54 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 |
import os, gc, random, re
import gradio as gr
import torch, spaces
from PIL import Image, ImageFilter
import numpy as np
import qrcode
from qrcode.constants import ERROR_CORRECT_H
from diffusers import (
StableDiffusionControlNetPipeline,
StableDiffusionControlNetImg2ImgPipeline, # for Hi-Res Fix
ControlNetModel,
DPMSolverMultistepScheduler,
)
# Quiet matplotlib cache warning on Spaces
os.environ.setdefault("MPLCONFIGDIR", "/tmp/mpl")
# ---- base models for the two tabs ----
BASE_MODELS = {
"stable-diffusion-v1-5": "runwayml/stable-diffusion-v1-5",
"dream": "Lykon/dreamshaper-8",
}
# ControlNet (QR Monster v2 for SD15)
CN_QRMON = "monster-labs/control_v1p_sd15_qrcode_monster"
DTYPE = torch.float16
# ---------- helpers ----------
def snap8(x: int) -> int:
x = max(256, min(1024, int(x)))
return x - (x % 8)
def normalize_color(c):
if c is None: return "white"
if isinstance(c, (tuple, list)):
r, g, b = (int(max(0, min(255, round(float(x))))) for x in c[:3]); return (r, g, b)
if isinstance(c, str):
s = c.strip()
if s.startswith("#"): return s
m = re.match(r"rgba?\(\s*([0-9.]+)\s*,\s*([0-9.]+)\s*,\s*([0-9.]+)", s, re.IGNORECASE)
if m:
r = int(max(0, min(255, round(float(m.group(1))))))
g = int(max(0, min(255, round(float(m.group(2))))))
b = int(max(0, min(255, round(float(m.group(3))))))
return (r, g, b)
return s
return "white"
def make_qr(url="https://example.com", size=768, border=12, back_color="#FFFFFF", blur_radius=0.0):
"""
IMPORTANT for Method 1: give ControlNet a sharp, black-on-WHITE QR (no blur).
"""
qr = qrcode.QRCode(version=None, error_correction=ERROR_CORRECT_H, box_size=10, border=int(border))
qr.add_data(url.strip()); qr.make(fit=True)
img = qr.make_image(fill_color="black", back_color=normalize_color(back_color)).convert("RGB")
img = img.resize((int(size), int(size)), Image.NEAREST)
if blur_radius and blur_radius > 0:
img = img.filter(ImageFilter.GaussianBlur(radius=float(blur_radius)))
return img
def enforce_qr_contrast(stylized: Image.Image, qr_img: Image.Image, strength: float = 0.0, feather: float = 1.0) -> Image.Image:
"""Optional gentle repair. Default OFF for Method 1."""
if strength <= 0: return stylized
q = qr_img.convert("L")
black_mask = q.point(lambda p: 255 if p < 128 else 0).filter(ImageFilter.GaussianBlur(radius=float(feather)))
black = np.asarray(black_mask, dtype=np.float32) / 255.0
white = 1.0 - black
s = np.asarray(stylized.convert("RGB"), dtype=np.float32) / 255.0
s = s * (1.0 - float(strength) * black[..., None])
s = s + (1.0 - s) * (float(strength) * 0.85 * white[..., None])
s = np.clip(s, 0.0, 1.0)
return Image.fromarray((s * 255.0).astype(np.uint8), mode="RGB")
# ---------- lazy pipelines (CPU-offloaded for ZeroGPU) ----------
_CN = None # shared ControlNet QR Monster
_CN_TXT2IMG = {} # per-base-model txt2img pipes
_CN_IMG2IMG = {} # per-base-model img2img pipes
def _base_scheduler_for(pipe):
pipe.scheduler = DPMSolverMultistepScheduler.from_config(
pipe.scheduler.config, use_karras_sigmas=True, algorithm_type="dpmsolver++"
)
pipe.enable_attention_slicing()
pipe.enable_vae_slicing()
pipe.enable_model_cpu_offload()
return pipe
def get_cn():
global _CN
if _CN is None:
_CN = ControlNetModel.from_pretrained(CN_QRMON, torch_dtype=DTYPE, use_safetensors=True)
return _CN
def get_qrmon_txt2img_pipe(model_id: str):
if model_id not in _CN_TXT2IMG:
pipe = StableDiffusionControlNetPipeline.from_pretrained(
model_id,
controlnet=get_cn(),
torch_dtype=DTYPE,
safety_checker=None,
use_safetensors=True,
low_cpu_mem_usage=True,
)
_CN_TXT2IMG[model_id] = _base_scheduler_for(pipe)
return _CN_TXT2IMG[model_id]
def get_qrmon_img2img_pipe(model_id: str):
if model_id not in _CN_IMG2IMG:
pipe = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(
model_id,
controlnet=get_cn(),
torch_dtype=DTYPE,
safety_checker=None,
use_safetensors=True,
low_cpu_mem_usage=True,
)
_CN_IMG2IMG[model_id] = _base_scheduler_for(pipe)
return _CN_IMG2IMG[model_id]
# -------- Method 1: QR control model in text-to-image (+ optional Hi-Res Fix) --------
def _qr_txt2img_core(model_id: str,
url: str, style_prompt: str, negative: str,
steps: int, cfg: float, size: int, border: int,
qr_weight: float, seed: int,
use_hires: bool, hires_upscale: float, hires_strength: float,
repair_strength: float, feather: float):
s = snap8(size)
# Control image: crisp black-on-white QR
qr_img = make_qr(url=url, size=s, border=int(border), back_color="#FFFFFF", blur_radius=0.0)
# Seed / generator
if int(seed) < 0:
seed = random.randint(0, 2**31 - 1)
gen = torch.Generator(device="cuda").manual_seed(int(seed))
# --- Stage A: txt2img with ControlNet
pipe = get_qrmon_txt2img_pipe(model_id)
if torch.cuda.is_available(): torch.cuda.empty_cache()
gc.collect()
with torch.autocast(device_type="cuda", dtype=DTYPE):
out = pipe(
prompt=str(style_prompt),
negative_prompt=str(negative or ""),
image=qr_img, # control image for txt2img
controlnet_conditioning_scale=float(qr_weight), # ~1.0–1.2 works well
control_guidance_start=0.0,
control_guidance_end=1.0,
num_inference_steps=int(steps),
guidance_scale=float(cfg),
width=s, height=s,
generator=gen,
)
lowres = out.images[0]
lowres = enforce_qr_contrast(lowres, qr_img, strength=float(repair_strength), feather=float(feather))
# --- Optional Stage B: Hi-Res Fix (img2img with same QR)
final = lowres
if use_hires:
up = max(1.0, min(2.0, float(hires_upscale)))
W = snap8(int(s * up)); H = W
pipe2 = get_qrmon_img2img_pipe(model_id)
if torch.cuda.is_available(): torch.cuda.empty_cache()
gc.collect()
with torch.autocast(device_type="cuda", dtype=DTYPE):
out2 = pipe2(
prompt=str(style_prompt),
negative_prompt=str(negative or ""),
image=lowres, # init image
control_image=qr_img, # same QR
strength=float(hires_strength), # ~0.7 like "Hires Fix"
controlnet_conditioning_scale=float(qr_weight),
control_guidance_start=0.0,
control_guidance_end=1.0,
num_inference_steps=int(steps),
guidance_scale=float(cfg),
width=W, height=H,
generator=gen,
)
final = out2.images[0]
final = enforce_qr_contrast(final, qr_img, strength=float(repair_strength), feather=float(feather))
return final, lowres, qr_img
# ===================== NEW: helpers for img2img =====================
def center_square(im: Image.Image) -> Image.Image:
"""Center-crop to square (keeps max area)."""
w, h = im.size
if w == h:
return im
if w > h:
off = (w - h) // 2
return im.crop((off, 0, off + h, h))
else:
off = (h - w) // 2
return im.crop((0, off, w, off + w))
def prep_init_image(init_img: Image.Image, target: int) -> Image.Image:
"""Center-crop to square, then resize to target (multiple of 8)."""
s = snap8(target)
im = center_square(init_img.convert("RGB"))
return im.resize((s, s), Image.LANCZOS)
# ================== NEW: img2img + QR Control core ==================
def _qr_img2img_core(model_id: str,
init_image: Image.Image,
url: str,
style_prompt: str,
negative: str,
steps: int,
cfg: float,
size: int,
border: int,
qr_weight: float,
seed: int,
strength: float, # how strongly to transform the init image (0.2–0.8 typical)
repair_strength: float,
feather: float):
"""
Blend an uploaded image with a scannable QR pattern:
- SD (img2img) generates/stylizes from init_image
- ControlNet (QR Monster) enforces QR structure
"""
# target square
s = snap8(size)
init = prep_init_image(init_image, s)
# crisp control QR
qr_img = make_qr(url=url, size=s, border=int(border), back_color="#FFFFFF", blur_radius=0.0)
# seed/generator
if int(seed) < 0:
seed = random.randint(0, 2**31 - 1)
gen = torch.Generator(device="cuda").manual_seed(int(seed))
pipe = get_qrmon_img2img_pipe(model_id)
if torch.cuda.is_available():
torch.cuda.empty_cache()
gc.collect()
# run
with torch.autocast(device_type="cuda", dtype=DTYPE):
out = pipe(
prompt=str(style_prompt),
negative_prompt=str(negative or ""),
image=init, # init image (img2img)
control_image=qr_img, # ControlNet input (QR)
strength=float(strength), # how far to move from init image
controlnet_conditioning_scale=float(qr_weight),
control_guidance_start=0.0,
control_guidance_end=0.85,
num_inference_steps=int(steps),
guidance_scale=float(cfg),
width=s, height=s,
generator=gen,
)
final = out.images[0]
# optional gentle repair to nudge contrast inside QR cells
final = enforce_qr_contrast(final, qr_img, strength=float(repair_strength), feather=float(feather))
# we’ll also return the resized init and QR used for UI
return final, init, qr_img
# ============== NEW: wrapper so Gradio can bind with api_name ==========
@spaces.GPU(duration=120)
def qr_img2img_blend(model_key: str,
init_image: Image.Image,
url: str, style_prompt: str, negative: str,
steps: int, cfg: float, size: int, border: int,
qr_weight: float, seed: int,
strength: float, # img2img denoise strength
repair_strength: float, feather: float):
"""
model_key is one of BASE_MODELS keys; we resolve to model_id here.
"""
model_id = BASE_MODELS.get(model_key, BASE_MODELS["stable-diffusion-v1-5"])
return _qr_img2img_core(model_id,
init_image,
url, style_prompt, negative,
steps, cfg, size, border,
qr_weight, seed,
strength,
repair_strength, feather)
# Wrappers for each tab (so Gradio can bind without passing the model id)
@spaces.GPU(duration=120)
def qr_txt2img_sd15(*args):
return _qr_txt2img_core(BASE_MODELS["stable-diffusion-v1-5"], *args)
@spaces.GPU(duration=120)
def qr_txt2img_dream(*args):
return _qr_txt2img_core(BASE_MODELS["dream"], *args)
# ---------- UI ----------
with gr.Blocks() as demo:
gr.Markdown("# ZeroGPU • Method 1: QR Control (two base models)")
# ---- Tab 1: stable-diffusion-v1-5 (anime/illustration) ----
with gr.Tab("stable-diffusion-v1-5"):
url1 = gr.Textbox(label="URL/Text", value="http://www.mybirdfire.com")
s_prompt1 = gr.Textbox(label="Style prompt", value="japanese painting, elegant shrine and torii, distant mount fuji, autumn maple trees, warm sunlight, 1girl in kimono, highly detailed, intricate patterns, anime key visual, dramatic composition")
s_negative1= gr.Textbox(label="Negative prompt", value="ugly, low quality, blurry, nsfw, watermark, text, low contrast, deformed, extra digits")
size1 = gr.Slider(384, 1024, value=512, step=64, label="Canvas (px)")
steps1 = gr.Slider(10, 50, value=20, step=1, label="Steps")
cfg1 = gr.Slider(1.0, 12.0, value=7.0, step=0.1, label="CFG")
border1 = gr.Slider(2, 16, value=4, step=1, label="QR border (quiet zone)")
qr_w1 = gr.Slider(0.6, 1.6, value=1.5, step=0.05, label="QR control weight")
seed1 = gr.Number(value=-1, precision=0, label="Seed (-1 random)")
use_hires1 = gr.Checkbox(value=True, label="Hi-Res Fix (img2img upscale)")
hires_up1 = gr.Slider(1.0, 2.0, value=2.0, step=0.25, label="Hi-Res upscale (×)")
hires_str1 = gr.Slider(0.3, 0.9, value=0.7, step=0.05, label="Hi-Res denoise strength")
repair1 = gr.Slider(0.0, 1.0, value=0.0, step=0.05, label="Post repair strength (optional)")
feather1 = gr.Slider(0.0, 3.0, value=1.0, step=0.1, label="Repair feather (px)")
final_img1 = gr.Image(label="Final (or Hi-Res) image")
low_img1 = gr.Image(label="Low-res (Stage A) preview")
ctrl_img1 = gr.Image(label="Control QR used")
gr.Button("Generate with stable-diffusion-v1-5").click(
qr_txt2img_sd15,
[url1, s_prompt1, s_negative1, steps1, cfg1, size1, border1, qr_w1, seed1,
use_hires1, hires_up1, hires_str1, repair1, feather1],
[final_img1, low_img1, ctrl_img1],
api_name="qr_txt2img_sd15"
)
# ---- Tab 2: DreamShaper (general art/painterly) ----
with gr.Tab("DreamShaper 8"):
url2 = gr.Textbox(label="URL/Text", value="http://www.mybirdfire.com")
s_prompt2 = gr.Textbox(label="Style prompt", value="ornate baroque palace interior, gilded details, chandeliers, volumetric light, ultra detailed, cinematic")
s_negative2= gr.Textbox(label="Negative prompt", value="lowres, low contrast, blurry, jpeg artifacts, watermark, text, bad anatomy")
size2 = gr.Slider(384, 1024, value=512, step=64, label="Canvas (px)")
steps2 = gr.Slider(10, 50, value=24, step=1, label="Steps")
cfg2 = gr.Slider(1.0, 12.0, value=6.8, step=0.1, label="CFG")
border2 = gr.Slider(2, 16, value=8, step=1, label="QR border (quiet zone)")
qr_w2 = gr.Slider(0.6, 1.6, value=1.5, step=0.05, label="QR control weight")
seed2 = gr.Number(value=-1, precision=0, label="Seed (-1 random)")
use_hires2 = gr.Checkbox(value=True, label="Hi-Res Fix (img2img upscale)")
hires_up2 = gr.Slider(1.0, 2.0, value=2.0, step=0.25, label="Hi-Res upscale (×)")
hires_str2 = gr.Slider(0.3, 0.9, value=0.7, step=0.05, label="Hi-Res denoise strength")
repair2 = gr.Slider(0.0, 1.0, value=0.0, step=0.05, label="Post repair strength (optional)")
feather2 = gr.Slider(0.0, 3.0, value=1.0, step=0.1, label="Repair feather (px)")
final_img2 = gr.Image(label="Final (or Hi-Res) image")
low_img2 = gr.Image(label="Low-res (Stage A) preview")
ctrl_img2 = gr.Image(label="Control QR used")
gr.Button("Generate with DreamShaper 8").click(
qr_txt2img_dream,
[url2, s_prompt2, s_negative2, steps2, cfg2, size2, border2, qr_w2, seed2,
use_hires2, hires_up2, hires_str2, repair2, feather2],
[final_img2, low_img2, ctrl_img2],
api_name="qr_txt2img_dream"
)
# ------------------- NEW TAB: Image Blend (img2img) -------------------
with gr.Tab("Image Blend (img2img + QR)"):
# choose base model via dropdown (no need for separate tabs)
m_key = gr.Dropdown(choices=list(BASE_MODELS.keys()),
value="stable-diffusion-v1-5",
label="Base model")
init_up = gr.Image(label="Upload base image", type="pil")
url_b = gr.Textbox(label="URL/Text", value="http://www.mybirdfire.com")
s_prompt_b = gr.Textbox(label="Style prompt", value="highly detailed, cinematic lighting, rich textures")
s_negative_b = gr.Textbox(label="Negative prompt", value="ugly, low quality, blurry, watermark, text")
size_b = gr.Slider(384, 1024, value=768, step=64, label="Canvas (px, target)")
steps_b = gr.Slider(10, 50, value=28, step=1, label="Steps")
cfg_b = gr.Slider(1.0, 12.0, value=7.0, step=0.1, label="CFG")
border_b = gr.Slider(2, 16, value=8, step=1, label="QR border (quiet zone)")
qr_w_b = gr.Slider(0.6, 1.6, value=1.2, step=0.05, label="QR control weight")
seed_b = gr.Number(value=-1, precision=0, label="Seed (-1 random)")
# img2img denoise strength controls how much we change the uploaded image
strength_b = gr.Slider(0.2, 0.9, value=0.55, step=0.05, label="Img2Img denoise strength (blend amount)")
repair_b = gr.Slider(0.0, 1.0, value=0.0, step=0.05, label="Post repair strength (optional)")
feather_b = gr.Slider(0.0, 3.0, value=1.0, step=0.1, label="Repair feather (px)")
final_b = gr.Image(label="Final blended image")
init_b = gr.Image(label="(Resized) init image used")
ctrl_b = gr.Image(label="Control QR used")
gr.Button("Blend Uploaded Image with QR").click(
qr_img2img_blend,
[m_key, init_up, url_b, s_prompt_b, s_negative_b, steps_b, cfg_b, size_b, border_b,
qr_w_b, seed_b, strength_b, repair_b, feather_b],
[final_b, init_b, ctrl_b],
api_name="qr_img2img_blend"
)
if __name__ == "__main__":
demo.queue(max_size=12).launch()
|