Spaces:
Running on Zero
Running on Zero
File size: 31,566 Bytes
0b10f88 09f3ae8 0b10f88 12c9226 0b10f88 12c9226 0b10f88 12c9226 0b10f88 12c9226 0b10f88 12c9226 0b10f88 12c9226 0b10f88 8b17a6c 0b10f88 12c9226 0b10f88 12c9226 0b10f88 09f3ae8 0b10f88 09f3ae8 0b10f88 12c9226 0b10f88 12c9226 0b10f88 12c9226 0b10f88 12c9226 0b10f88 09f3ae8 0b10f88 09f3ae8 0b10f88 09f3ae8 0b10f88 09f3ae8 0b10f88 8b17a6c 0b10f88 09f3ae8 0b10f88 09f3ae8 0b10f88 09f3ae8 0b10f88 09f3ae8 0b10f88 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 | """GenRF: Generative Registration-free Image Fusion - Hugging Face Space entry point."""
from pathlib import Path
from functools import lru_cache
import html
import os
import tempfile
import time
import gradio as gr
from PIL import Image, ImageOps
ZERO_GPU = os.getenv("GENRF_ZERO_GPU", "1") == "1"
if ZERO_GPU:
from spaces import GPU
else:
def GPU(**kwargs):
return lambda function: function
try:
import torch
_HAS_TORCH = True
except Exception: # CPU-only / local UI preview
torch = None
_HAS_TORCH = False
ROOT = Path(__file__).resolve().parent
BASE = "black-forest-labs/flux-3-image-klein-base"
DEFAULT = "Fuse the infrared and visible images. Restore degraded visibility, preserve the visible scene geometry and natural colors, and make salient infrared targets clear without ghosts or halos."
PROMPTS = {
"IVIF": DEFAULT,
"MEF": "Fuse the two differently exposed images of the same scene into a naturally exposed RGB image. Preserve scene geometry, realistic colors, and fine details in both shadows and highlights. Avoid ghosting, halos, clipped highlights, and invented details.",
"MFF": "Fuse the two differently focused images of the same scene into a sharp all-in-focus RGB image. Preserve scene geometry, realistic colors, and fine details in both near and far focus regions. Avoid ghosting, halos, defocus blur, and invented details.",
}
MODES = {
"IVIF": "IVIF · Infrared / Visible",
"MEF": "MEF · Multi-exposure",
"MFF": "MFF · Multi-focus",
}
INPUTS = {
"IVIF": (("ir", "Infrared"), ("vi", "Visible")),
"MEF": (("over", "Overexposed"), ("under", "Underexposed")),
"MFF": (("near", "Near focus"), ("far", "Far focus")),
}
GROUPS = [
("IVIF", "Infrared / Visible Image Fusion"),
("MEF", "Multi-exposure Image Fusion"),
("MFF", "Multi-focus Image Fusion"),
]
DEFAULTS = {"steps": 30, "cfg": 3.5, "seed": 2026}
PAGE_TITLE = "GenRF: Generative Registration-free Image Fusion"
def task_name(mode):
return mode.split(" · ")[0]
def scenes(mode):
return sorted(
path.name
for path in (ROOT / "examples" / task_name(mode).lower()).iterdir()
if path.is_dir()
)
def example(mode, name):
folder = ROOT / "examples" / task_name(mode).lower() / name
if name not in scenes(mode):
raise gr.Error("Select an example from the list.")
files = [
next(path for path in folder.glob(stem + ".*") if path.is_file())
for stem, _ in INPUTS[task_name(mode)]
]
with Image.open(files[0]) as first, Image.open(files[1]) as second, \
Image.open(folder / "genrf.png") as result:
return first.copy(), second.copy(), result.copy(), None
def align_pair(first, second):
"""Bring two differently sized sources to a common size without distortion.
The first (infrared / reference) image is padded symmetrically with zeros until
its aspect ratio equals that of the second (visible) image. The visible image is
then resampled to the padded size, so neither source is stretched. Returns the
aligned pair together with a flag indicating whether alignment was applied.
"""
if first.size == second.size:
return first, second, False
wi, hi = first.size
wv, hv = second.size
ratio = wv / hv
if wi / hi > ratio:
target_h = max(hi, round(wi / ratio))
top = (target_h - hi) // 2
bottom = target_h - hi - top
first = ImageOps.expand(first, border=(0, top, 0, bottom), fill=0)
else:
target_w = max(wi, round(hi * ratio))
left = (target_w - wi) // 2
right = target_w - wi - left
first = ImageOps.expand(first, border=(left, 0, right, 0), fill=0)
second = second.resize(first.size, Image.Resampling.LANCZOS)
return first, second, True
@lru_cache(maxsize=1)
def pipeline():
import torch
from diffusers import ModularPipeline
token = os.environ.get("HF_TOKEN")
pipe = ModularPipeline.from_pretrained(BASE, token=token)
pipe._blocks.sub_blocks.pop("prompt_upsampler", None)
pipe._blocks.sub_blocks.pop("text_encoder", None)
pipe.load_components(names=["vae", "transformer"], dtype=torch.bfloat16)
pipe.to("cuda")
pipe.update_components(guider=pipe.guider.new(guidance_scale=DEFAULTS["cfg"]))
pipe.load_lora_weights(str(ROOT / "weights"))
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()
return pipe
@lru_cache(maxsize=4)
def embeddings(prompt):
import torch
from transformers import AutoTokenizer, Qwen3VLForConditionalGeneration
cache = torch.load(ROOT / "eval_prompt_flux3.pt", map_location="cpu", weights_only=True)
if prompt == cache["prompt"]:
return cache["pos"], cache["neg"]
token = os.environ.get("HF_TOKEN")
tokenizer = AutoTokenizer.from_pretrained(BASE, subfolder="text_encoder", token=token)
encoder = Qwen3VLForConditionalGeneration.from_pretrained(
BASE, subfolder="text_encoder", token=token, torch_dtype=torch.bfloat16
).eval().to("cuda")
try:
with torch.inference_mode():
raw = tokenizer.apply_chat_template(
[{"role": "user", "content": prompt}], tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(
[raw], return_tensors="pt", padding="max_length", padding_side="right",
truncation=True, max_length=512,
).to("cuda")
attn = inputs["attention_mask"]
positions = torch.arange(512, device="cuda")[None].masked_fill(attn == 0, 1)
hidden = encoder(
input_ids=inputs["input_ids"], attention_mask=attn, position_ids=positions,
output_hidden_states=True, use_cache=False,
).hidden_states
pos = torch.stack(hidden[4:33:4], dim=2).flatten(2).cpu()
finally:
del encoder
torch.cuda.empty_cache()
return pos, cache["neg"]
@GPU(duration=120)
def fuse(first, second, prompt, steps, cfg, seed):
import numpy as np
import torch
if first is None or second is None:
raise gr.Error("Upload both source images or pick an example below.")
if not torch.cuda.is_available():
raise gr.Error("Live fusion requires a CUDA GPU. You can still browse the archived examples.")
first, second = [ImageOps.exif_transpose(im).convert("RGB") for im in (first, second)]
first, second, _ = align_pair(first, second)
width, height = first.size
if min(width, height) < 16:
raise gr.Error("Each side must be at least 16 pixels.")
limit = 1280
if max(width, height) > limit:
size = tuple(max(16, round(side * limit / max(width, height))) for side in first.size)
first, second = [im.resize(size, Image.Resampling.LANCZOS) for im in (first, second)]
width, height = size
right, bottom = (-width) % 16, (-height) % 16
inputs = [
Image.fromarray(np.pad(np.asarray(im), ((0, bottom), (0, right), (0, 0)), mode="reflect"))
for im in (first, second)
]
pos, neg = embeddings(prompt)
pipe = pipeline()
pipe.update_components(guider=pipe.guider.new(guidance_scale=float(cfg)))
with torch.inference_mode():
image = pipe(
image=inputs, height=height + bottom, width=width + right,
num_inference_steps=int(steps),
generator=torch.Generator("cpu").manual_seed(int(seed)),
prompt_embeds=pos.to("cuda"), negative_prompt_embeds=neg.to("cuda"),
output="images",
)[0]
return image.crop((0, 0, width, height))
# --------------------------------------------------------------------------
# Presentation layer
# --------------------------------------------------------------------------
THEME = gr.themes.Soft(
primary_hue=gr.themes.colors.indigo,
secondary_hue=gr.themes.colors.slate,
neutral_hue=gr.themes.colors.slate,
font=["ui-sans-serif", "system-ui", "-apple-system", "Segoe UI", "sans-serif"],
)
CSS = r"""
:root {
--genrf-bg:#f5f7fb; --genrf-surface:#ffffff; --genrf-surface-2:#f8fafc;
--genrf-border:#e5eaf2; --genrf-border-strong:#d6deea;
--genrf-ink:#0f172a; --genrf-ink-2:#526077; --genrf-muted:#94a3b8;
--genrf-accent:#4f46e5; --genrf-accent-2:#7c6cf0; --genrf-accent-soft:#eef2ff;
--genrf-shadow:0 1px 2px rgba(15,23,42,.04), 0 10px 28px rgba(15,23,42,.045);
}
.dark, body.dark, html.dark, .gradio-container.dark {
--genrf-bg:#0b1120; --genrf-surface:#131c30; --genrf-surface-2:#101827;
--genrf-border:#26334c; --genrf-border-strong:#33425f;
--genrf-ink:#eef2fb; --genrf-ink-2:#aab7cd; --genrf-muted:#7b89a2;
--genrf-accent:#8f96ff; --genrf-accent-2:#6f78ef; --genrf-accent-soft:#1c2540;
--genrf-shadow:0 1px 2px rgba(0,0,0,.24), 0 10px 28px rgba(0,0,0,.22);
}
.gradio-container { max-width:1180px !important; margin-inline:auto !important;
background:var(--genrf-bg) !important; }
.dark .gradio-container { color:var(--genrf-ink) !important; }
.genrf-hero { display:flex; flex-wrap:wrap; align-items:center; gap:18px 28px; position:relative; overflow:hidden; padding:24px 28px; margin-bottom:16px;
border:1px solid var(--genrf-border); border-radius:20px; box-shadow:var(--genrf-shadow);
background:radial-gradient(700px 220px at 8% -30%, rgba(99,102,241,.20), transparent 70%),
linear-gradient(115deg, var(--genrf-accent-soft) 0%, var(--genrf-surface) 62%,
var(--genrf-surface-2) 100%); }
.genrf-hero-main { flex:1 1 520px; min-width:0; }
.genrf-links { display:flex; gap:8px; flex-wrap:wrap; margin-left:auto; }
.genrf-link { display:inline-flex; align-items:center; gap:7px; padding:8px 13px; border-radius:10px;
border:1px solid var(--genrf-border-strong); background:var(--genrf-surface); color:var(--genrf-ink);
font-size:12px; font-weight:600; text-decoration:none !important; box-shadow:0 1px 2px rgba(15,23,42,.04);
transition:border-color .15s ease, color .15s ease, transform .15s ease; }
.genrf-link:hover { border-color:var(--genrf-accent); color:var(--genrf-accent); transform:translateY(-1px); }
.genrf-link svg { width:14px; height:14px; display:block; }
.genrf-hero h1 { margin:14px 0 8px; font-size:clamp(21px,2.4vw,30px); font-weight:680;
letter-spacing:-.025em; line-height:1.2; color:var(--genrf-ink); }
.genrf-hero p { margin:0; max-width:80ch; font-size:13.5px; line-height:1.65; color:var(--genrf-ink-2); }
.genrf-badge { display:inline-block; padding:4px 11px; border-radius:999px; color:#fff;
background:linear-gradient(135deg,var(--genrf-accent),var(--genrf-accent-2));
font-size:11px; font-weight:700; letter-spacing:.08em; }
.genrf-badge.ghost { margin-left:6px; background:var(--genrf-surface); color:var(--genrf-accent);
border:1px solid var(--genrf-border-strong); font-weight:600; letter-spacing:0; }
.genrf-tags { margin-top:14px; display:flex; gap:8px; flex-wrap:wrap; }
.genrf-tags span { font-size:11px; padding:3px 9px; border-radius:6px;
background:var(--genrf-accent-soft); color:var(--genrf-accent); border:1px solid var(--genrf-border); }
.genrf-tags span.key { color:#fff; font-weight:650; border-color:transparent;
background:linear-gradient(135deg,var(--genrf-accent),var(--genrf-accent-2)); }
.genrf-panel { background:var(--genrf-surface) !important; border:1px solid var(--genrf-border) !important;
border-radius:18px !important; padding:18px !important; box-shadow:var(--genrf-shadow); }
.genrf-sec { font-size:13.5px; font-weight:650; line-height:1.4; color:var(--genrf-ink);
margin:0 0 14px; padding-left:11px; border-left:3px solid var(--genrf-accent); }
.genrf-sec small { display:block; margin:3px 0 0; font-weight:400; font-size:11px; color:var(--genrf-muted); }
.genrf-panel .html-container, .genrf-panel .html-container .prose, .genrf-instr-head,
.genrf-panel .styler, .genrf-panel .image-container, .genrf-panel .image-frame {
background:transparent !important; border:0 !important; box-shadow:none !important; }
.genrf-panel .styler { padding:0 !important; }
.genrf-panel .html-container { padding:0 !important; }
.genrf-sec-inline { margin:0 !important; }
.genrf-hint { font-size:11.5px; line-height:1.7; color:var(--genrf-ink-2); }
.genrf-note { margin-top:10px; font-size:11.5px; line-height:1.7; color:var(--genrf-muted); }
.genrf-instr-head { align-items:center !important; gap:14px !important; margin-bottom:8px; }
#genrf-preset { flex:0 0 250px !important; width:250px !important; min-width:250px !important; }
#genrf-preset .wrap { border:1px solid var(--genrf-border-strong) !important;
border-radius:999px !important; background:var(--genrf-surface-2) !important;
box-shadow:none !important; transition:border-color .15s ease, box-shadow .15s ease; }
#genrf-preset .wrap:hover { border-color:var(--genrf-accent) !important; }
#genrf-preset .wrap:focus-within { border-color:var(--genrf-accent) !important;
box-shadow:0 0 0 3px var(--genrf-accent-soft) !important; }
#genrf-preset .wrap-inner { padding:4px 26px 4px 12px !important; }
#genrf-preset input { font-size:12px !important; font-weight:600 !important;
color:var(--genrf-ink) !important; height:auto !important; }
#genrf-preset .icon-wrap { right:9px !important; width:14px !important; color:var(--genrf-muted) !important; }
#genrf-prompt textarea { font-size:13px !important; line-height:1.65 !important; }
.genrf-params [data-testid="block-info"] { background:transparent !important; border:0 !important;
padding:0 0 0 2px !important; color:var(--genrf-ink-2) !important; font-weight:600 !important;
font-size:11.5px !important; }
#genrf-reset { background:var(--genrf-surface-2) !important; border:1px solid var(--genrf-border-strong) !important;
border-radius:999px !important; padding:0 14px !important;
color:var(--genrf-ink-2) !important; font-weight:600 !important; font-size:12px !important;
box-shadow:none !important; transition:border-color .15s ease, color .15s ease; }
#genrf-reset:hover { border-color:var(--genrf-accent) !important; color:var(--genrf-accent) !important; }
.genrf-group { display:flex; align-items:center; gap:10px; margin:16px 0 8px; padding-bottom:8px;
border-bottom:1px solid var(--genrf-border); font-size:12.5px; color:var(--genrf-ink-2); }
.genrf-tagcode { font-family:ui-monospace,SFMono-Regular,Consolas,monospace; font-size:11px;
font-weight:700; color:var(--genrf-accent); background:var(--genrf-accent-soft);
border:1px solid var(--genrf-border); padding:2px 8px; border-radius:6px; }
.genrf-count { margin-left:auto; color:var(--genrf-muted); font-size:11px; }
.genrf-foot { margin-top:14px; padding:12px 4px; border-top:1px solid var(--genrf-border);
font-size:11.5px; color:var(--genrf-muted); line-height:1.75; }
#genrf-run { min-height:46px !important; font-size:14px !important; font-weight:650 !important;
letter-spacing:.01em; border:0 !important; color:#fff !important;
background:linear-gradient(135deg,var(--genrf-accent),var(--genrf-accent-2)) !important;
box-shadow:0 6px 18px rgba(79,70,229,.22) !important;
transition:transform .15s ease, box-shadow .15s ease, filter .15s ease !important; }
#genrf-run:hover:not(:disabled) { transform:translateY(-2px);
box-shadow:0 10px 26px rgba(79,70,229,.34) !important; }
#genrf-run:disabled, #genrf-run.disabled { filter:grayscale(.55); opacity:.55;
box-shadow:none !important; transform:none !important; }
.genrf-status { border-radius:12px; padding:11px 14px; font-size:12.5px; line-height:1.65;
background:var(--genrf-surface-2); border:1px solid var(--genrf-border);
border-left:3px solid var(--genrf-border-strong); color:var(--genrf-ink); }
.genrf-status b { display:block; margin-bottom:2px; font-size:12.5px; }
.genrf-status span { color:var(--genrf-ink-2); }
.genrf-status[data-kind="idle"] { border-left-color:#94a3b8; }
.genrf-status[data-kind="archive"] { border-left-color:var(--genrf-accent); }
.genrf-status[data-kind="running"], .genrf-status[data-kind="gallery"] { border-left-color:#f59e0b; }
.genrf-status[data-kind="done"] { border-left-color:#16a34a; }
.genrf-status[data-kind="error"] { border-left-color:#dc2626; }
.genrf-legend { display:grid; grid-template-columns:repeat(3,minmax(0,1fr));
column-gap:9px; margin:10px 0 8px; padding:0 2px; font-size:10.5px; letter-spacing:.02em;
text-transform:uppercase; font-weight:650; color:var(--genrf-muted); }
.genrf-legend span:last-child { color:var(--genrf-accent); }
@media (max-width:560px) { .genrf-legend { column-gap:6px; font-size:9.5px; } }
.genrf-gallery .grid-wrap, .genrf-gallery .grid-wrap.fixed-height {
min-height:0 !important; max-height:none !important; overflow-y:visible !important; }
.genrf-gallery .grid-container { grid-template-columns:repeat(3,minmax(0,1fr)) !important;
grid-template-rows:none !important; grid-auto-rows:auto !important;
column-gap:9px !important; row-gap:18px !important; align-items:start !important; }
@media (max-width:560px) { .genrf-gallery .grid-container { column-gap:6px !important; row-gap:12px !important; } }
.genrf-gallery .gallery-item { aspect-ratio:auto !important; height:auto !important; }
.genrf-gallery .thumbnail-item { aspect-ratio:auto !important; height:auto !important; }
.genrf-gallery .thumbnail-item img { width:100% !important; height:auto !important;
object-fit:fill !important; display:block; }
.genrf-gallery .gallery-item { border-radius:10px !important; overflow:hidden;
border:1px solid var(--genrf-border) !important;
transition:transform .15s ease, box-shadow .15s ease; }
.genrf-gallery .gallery-item:hover { transform:translateY(-2px);
box-shadow:0 8px 20px rgba(15,23,42,.16); }
.genrf-panel label[data-testid="block-label"] { background:var(--genrf-accent-soft) !important;
color:var(--genrf-accent) !important; border:1px solid var(--genrf-border) !important;
border-radius:8px !important; font-weight:600 !important; }
#genrf-result { background:var(--genrf-surface) !important; border-color:var(--genrf-border) !important; }
#genrf-result img { object-position:top center !important; }
#genrf-result .image-container { align-items:flex-start !important; }
footer { display:none !important; }
#genrf-run:focus-visible { outline:2px solid var(--genrf-accent); outline-offset:2px; }
@media (max-width:820px) {
.genrf-hero { padding:18px 20px; }
.genrf-links { margin-left:0; }
.genrf-panel { padding:14px !important; }
.genrf-instr-head { flex-wrap:wrap !important; }
}
@media (prefers-reduced-motion:reduce) {
.genrf-hero, #genrf-run, .genrf-gallery .gallery-item, #genrf-preset .wrap { transition:none !important; }
}
"""
# The arXiv entry is not live yet, so Paper points at the repository for now.
PAPER_URL = "https://github.com/Linfeng-Tang/GenRF"
CODE_URL = "https://github.com/Linfeng-Tang/GenRF"
PAGE_URL = "https://linfeng-tang.github.io/GenRF/"
HERO = """
<div class="genrf-hero">
<div class="genrf-hero-main">
<span class="genrf-badge">GenRF</span>
<span class="genrf-badge ghost">Interactive Demo</span>
<h1>GenRF: Generative Registration-free Image Fusion</h1>
<p>Misaligned, degraded, partially overlapping captures go in —<br>
a high-quality, coherent fused image comes out.</p>
<div class="genrf-tags"><span class="key">Generative Image Fusion</span>
<span>Registration-free</span><span>IVIF</span><span>MEF</span><span>MFF</span></div>
</div>
<div class="genrf-links">
<a class="genrf-link" href="{paper}" target="_blank" rel="noopener">
<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round"><path d="M14 3H7a2 2 0 0 0-2 2v14a2 2 0 0 0 2 2h10a2 2 0 0 0 2-2V8z"/><path d="M14 3v5h5"/><path d="M9 12.5h6M9 16.5h4"/></svg>Paper</a>
<a class="genrf-link" href="{code}" target="_blank" rel="noopener">
<svg viewBox="0 0 24 24" fill="currentColor"><path d="M12 .3a12 12 0 0 0-3.8 23.4c.6.1.8-.3.8-.6v-2c-3.3.7-4-1.6-4-1.6-.6-1.4-1.4-1.8-1.4-1.8-1.1-.7.1-.7.1-.7 1.2.1 1.8 1.2 1.8 1.2 1.1 1.8 2.8 1.3 3.5 1 .1-.8.4-1.3.8-1.6-2.7-.3-5.5-1.3-5.5-5.9 0-1.3.5-2.4 1.2-3.2-.1-.3-.5-1.5.1-3.2 0 0 1-.3 3.3 1.2a11.5 11.5 0 0 1 6 0C17.3 4.7 18.3 5 18.3 5c.6 1.7.2 2.9.1 3.2.8.8 1.2 1.9 1.2 3.2 0 4.6-2.8 5.6-5.5 5.9.4.4.8 1.1.8 2.2v3.3c0 .3.2.7.8.6A12 12 0 0 0 12 .3"/></svg>Code</a>
<a class="genrf-link" href="{page}" target="_blank" rel="noopener">
<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round"><circle cx="12" cy="12" r="9"/><path d="M3.2 9h17.6M3.2 15h17.6"/><path d="M12 3c2.5 2.6 3.8 5.7 3.8 9S14.5 18.4 12 21c-2.5-2.6-3.8-5.7-3.8-9S9.5 5.6 12 3z"/></svg>Project page</a>
</div>
</div>
""".format(paper=PAPER_URL, code=CODE_URL, page=PAGE_URL)
FOOTER = """
<div class="genrf-foot">
Live inference requires CUDA and a read-only HF_TOKEN for the gated FLUX.3 Klein Base model.
</div>
"""
STATUS_STYLE = {
# Quiet states render nothing: a persistent hint adds noise without information.
"idle": ("", ""),
"archive": ("", ""),
"running": ("Running", "Fusion in progress. Please wait..."),
"done": ("Done", "Fusion completed."),
"error": ("Error", ""),
"gallery": ("Gallery mode",
"CUDA hardware or HF_TOKEN is unavailable, so live fusion is disabled. Archived examples below remain available."),
}
def status_html(kind, detail="", note=""):
if kind in ("idle", "archive"):
return ""
title, text = STATUS_STYLE[kind]
if kind == "done":
text = "Fusion completed in {}.".format(html.escape(str(detail)))
if note:
text += " " + html.escape(note)
elif kind == "error":
text = html.escape(str(detail)) or "Fusion failed. Check the input images and the message above."
return (
'<div class="genrf-status" data-kind="{}"><b>{}</b><span>{}</span></div>'
).format(kind, title, text)
def section(title, inline=False):
return '<div class="genrf-sec{}">{}</div>'.format(" genrf-sec-inline" if inline else "", title)
def html_block(value):
"""Bare HTML block: no Gradio padding panel, so copy sits on the card itself."""
return gr.HTML(value, padding=False)
def legend(first_label, second_label):
return (
'<div class="genrf-legend"><span>{}</span><span>{}</span><span>GenRF result</span></div>'
).format(first_label, second_label)
def group_title(code, label, count):
return (
'<div class="genrf-group"><span class="genrf-tagcode">{}</span>'
'<span>{}</span><span class="genrf-count">{} cases</span></div>'
).format(code, label, count)
THUMB_W = 560
@lru_cache(maxsize=None)
def preview(path_str, mtime):
"""560 px wide JPEG preview, generated once per process into the temp dir."""
source = Path(path_str)
out_dir = Path(tempfile.gettempdir()) / "genrf_previews" / source.parent.parent.name / source.parent.name
out_dir.mkdir(parents=True, exist_ok=True)
out = out_dir / (source.stem + ".jpg")
if not out.exists():
with Image.open(source) as image:
image = image.convert("RGB")
if image.width > THUMB_W:
height = round(image.height * THUMB_W / image.width)
image = image.resize((THUMB_W, height), Image.LANCZOS)
image.save(out, "JPEG", quality=82, optimize=True, progressive=True)
return str(out)
def preview_of(path):
path = Path(path)
return preview(str(path), path.stat().st_mtime)
def gallery_items(task):
"""One row per case: the two captured inputs, then the fused result."""
base = ROOT / "examples" / task.lower()
items = []
for name in scenes(task):
folder = base / name
for stem, _ in INPUTS[task]:
source = next(p for p in folder.glob(stem + ".*") if p.is_file())
items.append((preview_of(source), ""))
fused = folder / "thumb.jpg"
items.append((preview_of(fused if fused.exists() else folder / "genrf.png"), ""))
return items
LIVE_READY = bool(os.getenv("HF_TOKEN"))
LIVE_OK = LIVE_READY and (ZERO_GPU or (_HAS_TORCH and torch.cuda.is_available()))
if ZERO_GPU and LIVE_READY and _HAS_TORCH:
pipeline()
with gr.Blocks(title=PAGE_TITLE, theme=THEME, css=CSS) as demo:
html_block(HERO)
# ---- Row 1: source images on the left, fusion result on the right ----
with gr.Row(equal_height=True):
with gr.Column(scale=1, elem_classes=["genrf-panel"]):
html_block(section("Source images"))
first = gr.Image(type="pil", label="Source A - Infrared", sources=["upload"], height=230)
second = gr.Image(type="pil", label="Source B - Visible", sources=["upload"], height=230)
run = gr.Button("Run Fusion", variant="primary", elem_id="genrf-run",
interactive=LIVE_OK)
with gr.Column(scale=1, elem_classes=["genrf-panel"]):
html_block(section("Fusion result"))
result = gr.Image(label="", show_label=False, interactive=False, format="png",
height=300, elem_id="genrf-result")
html_block('<div class="genrf-note">Returned in the visible reference geometry, '
'at the full resolution of the padded pair.</div>')
status = html_block(status_html("gallery" if not LIVE_OK else "idle"))
# ---- Row 2: instruction with an inline task preset, then tunable parameters ----
with gr.Group(elem_classes=["genrf-panel"]):
with gr.Row(elem_classes=["genrf-instr-head"]):
html_block(section("Fusion instruction", inline=True))
preset = gr.Dropdown([MODES[k] for k in PROMPTS], value=MODES["IVIF"],
show_label=False, container=False, filterable=False,
scale=0, min_width=250, elem_id="genrf-preset")
reset = gr.Button("Reset", size="sm", variant="secondary",
scale=0, min_width=86, elem_id="genrf-reset")
prompt = gr.Textbox(value=PROMPTS["IVIF"], show_label=False, lines=4,
placeholder="Describe how the two sources should be fused.",
elem_id="genrf-prompt")
with gr.Row(elem_classes=["genrf-params"]):
steps = gr.Slider(10, 60, value=DEFAULTS["steps"], step=1, label="Denoising steps")
cfg = gr.Slider(1.0, 8.0, value=DEFAULTS["cfg"], step=0.1, label="CFG / guidance")
seed = gr.Number(value=DEFAULTS["seed"], precision=0, label="Random seed")
with gr.Accordion("Notes & defaults", open=False):
html_block(
'<div class="genrf-hint">'
'Defaults: 30 denoising steps, guidance 3.5, seed 2026.<br>'
'Differing input sizes are aligned automatically: the first (infrared / reference) image '
'is zero-padded to the second (visible) image\'s aspect ratio, and the second image is '
'resampled to the padded size. Long edges above 1280 px are downsampled and padded to a '
'multiple of 16.<br>'
'Editing the instruction reloads the Qwen3-VL text encoder on first use. GenRF generalizes '
'to misaligned multi-exposure and multi-focus fusion without fine-tuning.'
'</div>'
)
# ---- Row 3: example gallery, grouped by task with text labels ----
with gr.Group(elem_classes=["genrf-panel"]):
html_block('<div class="genrf-sec">Examples<small>One row per case: the two captured '
'inputs, then the fused result. Click any tile to load that case above.</small></div>')
for code, label in GROUPS:
names = scenes(code)
html_block(group_title(code, label, len(names)))
first_label, second_label = (label for _, label in INPUTS[code])
html_block(legend(first_label, second_label))
gallery = gr.Gallery(
value=gallery_items(code), label="", show_label=False,
columns=3, rows=len(names), object_fit="contain",
allow_preview=False, elem_classes=["genrf-gallery"],
)
def make_picker(task=code, choices=names):
def _pick(evt: gr.SelectData):
idx = evt.index
if isinstance(idx, (list, tuple)):
idx = idx[0]
idx = int(idx) // 3 # three tiles per case: input A, input B, result
mode = MODES[task]
a, b, fused, _ = example(mode, choices[int(idx)])
return (a, b, fused, mode, PROMPTS[task],
status_html("archive"))
return _pick
gallery.select(make_picker(), None,
[first, second, result, preset, prompt, status])
html_block(FOOTER)
# ---- Callbacks ----
def set_preset(mode):
labels = INPUTS[task_name(mode)]
return (
PROMPTS[task_name(mode)],
gr.update(label="Source A - " + labels[0][1]),
gr.update(label="Source B - " + labels[1][1]),
)
def initial():
a, b, fused, _ = example(MODES["IVIF"], scenes(MODES["IVIF"])[0])
return a, b, fused, status_html("gallery" if not LIVE_OK else "archive")
def reset_params():
return (DEFAULTS["steps"], DEFAULTS["cfg"], DEFAULTS["seed"],
PROMPTS["IVIF"], MODES["IVIF"])
def busy():
return status_html("running"), gr.update(interactive=False)
def run_fusion(a, b, text, n_steps, guidance, seed_value):
aligned = a is not None and b is not None and a.size != b.size
started = time.perf_counter()
try:
output = fuse(a, b, text, n_steps, guidance, seed_value)
except Exception as exc:
message = str(exc)
gr.Warning(message)
return gr.update(), status_html("error", message), gr.update(interactive=LIVE_OK)
elapsed = time.perf_counter() - started
note = ("Inputs had different sizes: the first image was zero-padded to the second "
"image's aspect ratio, and the second image was resampled to the padded size."
if aligned else "")
return (output, status_html("done", "{:.1f} s".format(elapsed), note),
gr.update(interactive=LIVE_OK))
preset.change(set_preset, preset, [prompt, first, second])
reset.click(reset_params, None, [steps, cfg, seed, prompt, preset])
demo.load(initial, None, [first, second, result, status])
run.click(busy, None, [status, run]).then(
run_fusion, [first, second, prompt, steps, cfg, seed],
[result, status, run], concurrency_limit=1
)
if __name__ == "__main__":
demo.queue(max_size=12).launch()
|