Update app.py
Browse files
app.py
CHANGED
|
@@ -1,34 +1,32 @@
|
|
| 1 |
import os
|
| 2 |
-
import gc
|
| 3 |
-
import torch
|
| 4 |
-
import gradio as gr
|
| 5 |
-
from PIL import Image
|
| 6 |
-
from pathlib import Path
|
| 7 |
-
import shutil
|
| 8 |
import json
|
| 9 |
-
|
| 10 |
-
import
|
| 11 |
-
import
|
| 12 |
-
import
|
| 13 |
-
|
| 14 |
-
# Training imports
|
| 15 |
-
from peft import LoraConfig, get_peft_model
|
| 16 |
-
from tqdm.auto import tqdm
|
| 17 |
import numpy as np
|
|
|
|
|
|
|
| 18 |
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
# Global state
|
| 23 |
-
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 24 |
-
DTYPE = torch.bfloat16 if torch.cuda.is_available() else torch.float32
|
| 25 |
|
| 26 |
-
|
| 27 |
-
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
-
|
| 30 |
-
|
|
|
|
|
|
|
|
|
|
| 31 |
|
|
|
|
|
|
|
| 32 |
|
| 33 |
# ============================================
|
| 34 |
# Comic Style CSS
|
|
@@ -202,15 +200,6 @@ textarea:focus, input[type="text"]:focus {
|
|
| 202 |
color: #1F2937 !important;
|
| 203 |
}
|
| 204 |
|
| 205 |
-
.result-box textarea {
|
| 206 |
-
background: #1F2937 !important;
|
| 207 |
-
color: #10B981 !important;
|
| 208 |
-
font-family: 'Courier New', monospace !important;
|
| 209 |
-
border: 3px solid #10B981 !important;
|
| 210 |
-
border-radius: 8px !important;
|
| 211 |
-
box-shadow: 4px 4px 0 #10B981 !important;
|
| 212 |
-
}
|
| 213 |
-
|
| 214 |
label, .gr-input-label, .gr-block-label {
|
| 215 |
color: #1F2937 !important;
|
| 216 |
font-family: 'Comic Neue', cursive !important;
|
|
@@ -224,19 +213,6 @@ label, .gr-input-label, .gr-block-label {
|
|
| 224 |
box-shadow: 4px 4px 0 #1F2937 !important;
|
| 225 |
}
|
| 226 |
|
| 227 |
-
.tab-nav button {
|
| 228 |
-
font-family: 'Comic Neue', cursive !important;
|
| 229 |
-
font-weight: 700 !important;
|
| 230 |
-
border: 2px solid #1F2937 !important;
|
| 231 |
-
margin: 2px !important;
|
| 232 |
-
}
|
| 233 |
-
|
| 234 |
-
.tab-nav button.selected {
|
| 235 |
-
background: #3B82F6 !important;
|
| 236 |
-
color: #FFF !important;
|
| 237 |
-
box-shadow: 3px 3px 0 #1F2937 !important;
|
| 238 |
-
}
|
| 239 |
-
|
| 240 |
.footer-comic {
|
| 241 |
text-align: center;
|
| 242 |
padding: 20px;
|
|
@@ -284,11 +260,6 @@ input[type="range"] {
|
|
| 284 |
accent-color: #3B82F6;
|
| 285 |
}
|
| 286 |
|
| 287 |
-
.gr-slider input[type="range"]::-webkit-slider-thumb {
|
| 288 |
-
background: #EF4444 !important;
|
| 289 |
-
border: 2px solid #1F2937 !important;
|
| 290 |
-
}
|
| 291 |
-
|
| 292 |
/* Image/Gallery Container */
|
| 293 |
.gr-image, .gr-gallery {
|
| 294 |
border: 3px solid #1F2937 !important;
|
|
@@ -296,736 +267,495 @@ input[type="range"] {
|
|
| 296 |
box-shadow: 4px 4px 0 #1F2937 !important;
|
| 297 |
}
|
| 298 |
|
| 299 |
-
/*
|
| 300 |
-
|
| 301 |
-
|
| 302 |
-
|
| 303 |
-
|
| 304 |
-
|
| 305 |
-
|
| 306 |
-
|
| 307 |
-
|
| 308 |
-
|
| 309 |
-
|
| 310 |
-
}
|
| 311 |
-
|
| 312 |
-
|
| 313 |
-
|
| 314 |
-
margin: 0 auto;
|
| 315 |
-
}
|
| 316 |
-
|
| 317 |
-
/* Hide Hugging Face elements */
|
| 318 |
-
.huggingface-space-link,
|
| 319 |
-
a[href*="huggingface.co/spaces"],
|
| 320 |
-
button[class*="share"],
|
| 321 |
-
.share-button,
|
| 322 |
-
[class*="hf-logo"],
|
| 323 |
-
.gr-share-btn,
|
| 324 |
-
#hf-logo,
|
| 325 |
-
.hf-icon,
|
| 326 |
-
svg[class*="hf"],
|
| 327 |
-
div[class*="huggingface"],
|
| 328 |
-
a[class*="huggingface"],
|
| 329 |
-
.svelte-1rjryqp,
|
| 330 |
-
header a[href*="huggingface"],
|
| 331 |
-
.space-header,
|
| 332 |
-
div.absolute.right-0 a[href*="huggingface"],
|
| 333 |
-
.gr-group > a[href*="huggingface"],
|
| 334 |
-
a[target="_blank"][href*="huggingface.co"] {
|
| 335 |
-
display: none !important;
|
| 336 |
-
visibility: hidden !important;
|
| 337 |
-
opacity: 0 !important;
|
| 338 |
-
pointer-events: none !important;
|
| 339 |
-
width: 0 !important;
|
| 340 |
-
height: 0 !important;
|
| 341 |
-
overflow: hidden !important;
|
| 342 |
-
}
|
| 343 |
-
|
| 344 |
-
/* Training specific styles */
|
| 345 |
-
.training-section {
|
| 346 |
-
background: linear-gradient(135deg, #E0F2FE 0%, #DBEAFE 100%) !important;
|
| 347 |
-
border: 3px solid #1F2937 !important;
|
| 348 |
-
border-radius: 12px !important;
|
| 349 |
-
padding: 15px !important;
|
| 350 |
-
margin: 10px 0 !important;
|
| 351 |
-
box-shadow: 4px 4px 0 #1F2937 !important;
|
| 352 |
-
}
|
| 353 |
-
|
| 354 |
-
.tips-box {
|
| 355 |
-
background: linear-gradient(135deg, #D1FAE5 0%, #A7F3D0 100%) !important;
|
| 356 |
-
border: 3px solid #1F2937 !important;
|
| 357 |
-
border-radius: 8px !important;
|
| 358 |
-
padding: 12px 15px !important;
|
| 359 |
-
margin: 10px 0 !important;
|
| 360 |
-
box-shadow: 4px 4px 0 #1F2937 !important;
|
| 361 |
-
font-family: 'Comic Neue', cursive !important;
|
| 362 |
-
font-weight: 700 !important;
|
| 363 |
-
color: #1F2937 !important;
|
| 364 |
-
}
|
| 365 |
"""
|
| 366 |
|
| 367 |
-
|
| 368 |
-
|
| 369 |
-
|
| 370 |
-
|
| 371 |
-
gc.collect()
|
| 372 |
-
if torch.cuda.is_available():
|
| 373 |
-
torch.cuda.empty_cache()
|
| 374 |
-
torch.cuda.synchronize()
|
| 375 |
-
torch.cuda.reset_peak_memory_stats()
|
| 376 |
-
torch.cuda.reset_accumulated_memory_stats()
|
| 377 |
-
|
| 378 |
-
|
| 379 |
-
def get_gpu_memory_info():
|
| 380 |
-
"""Get current GPU memory status"""
|
| 381 |
-
if torch.cuda.is_available():
|
| 382 |
-
allocated = torch.cuda.memory_allocated(0) / 1e9
|
| 383 |
-
reserved = torch.cuda.memory_reserved(0) / 1e9
|
| 384 |
-
total = torch.cuda.get_device_properties(0).total_memory / 1e9
|
| 385 |
-
free = total - allocated
|
| 386 |
-
return f"GPU: {allocated:.1f}GB allocated, {reserved:.1f}GB reserved, {free:.1f}GB free of {total:.1f}GB"
|
| 387 |
-
return "No GPU"
|
| 388 |
-
|
| 389 |
-
|
| 390 |
-
def check_gpu():
|
| 391 |
-
"""Check GPU availability and memory"""
|
| 392 |
-
if torch.cuda.is_available():
|
| 393 |
-
gpu_name = torch.cuda.get_device_name(0)
|
| 394 |
-
gpu_mem = torch.cuda.get_device_properties(0).total_memory / 1e9
|
| 395 |
-
return f"✅ GPU: {gpu_name} ({gpu_mem:.1f}GB total)"
|
| 396 |
-
return "❌ No GPU detected"
|
| 397 |
-
|
| 398 |
-
|
| 399 |
-
def check_hf_token():
|
| 400 |
-
"""Check if HF_TOKEN is configured"""
|
| 401 |
-
if HF_TOKEN:
|
| 402 |
-
try:
|
| 403 |
-
api = HfApi(token=HF_TOKEN)
|
| 404 |
-
user_info = api.whoami()
|
| 405 |
-
return f"✅ Logged in as: {user_info['name']}"
|
| 406 |
-
except Exception as e:
|
| 407 |
-
return f"⚠️ Token invalid: {str(e)}"
|
| 408 |
-
return "❌ HF_TOKEN not set"
|
| 409 |
-
|
| 410 |
-
|
| 411 |
-
def get_hf_username():
|
| 412 |
-
if HF_TOKEN:
|
| 413 |
-
try:
|
| 414 |
-
api = HfApi(token=HF_TOKEN)
|
| 415 |
-
return api.whoami()['name']
|
| 416 |
-
except:
|
| 417 |
-
return None
|
| 418 |
-
return None
|
| 419 |
-
|
| 420 |
-
|
| 421 |
-
# ============================================
|
| 422 |
-
# SUBPROCESS-BASED CAPTIONING
|
| 423 |
-
# ============================================
|
| 424 |
-
|
| 425 |
-
def _caption_worker(image_paths_queue, results_queue, trigger_word, is_person):
|
| 426 |
-
"""Worker process for Florence-2 captioning - completely isolated GPU context"""
|
| 427 |
-
import torch
|
| 428 |
-
from PIL import Image
|
| 429 |
-
|
| 430 |
-
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'
|
| 431 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 432 |
-
|
| 433 |
-
try:
|
| 434 |
-
from transformers import AutoProcessor, AutoModelForCausalLM
|
| 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 |
-
worker.start()
|
| 539 |
-
image_paths_queue.put(image_paths)
|
| 540 |
-
|
| 541 |
-
try:
|
| 542 |
-
captions = results_queue.get(timeout=900) # 15 min timeout
|
| 543 |
-
except Exception as e:
|
| 544 |
-
print(f"[Main] Captioning error: {e}")
|
| 545 |
-
captions = [trigger_word] * len(image_paths)
|
| 546 |
-
|
| 547 |
-
worker.join(timeout=30)
|
| 548 |
-
if worker.is_alive():
|
| 549 |
-
worker.terminate()
|
| 550 |
-
worker.join()
|
| 551 |
-
|
| 552 |
-
print(f"[Main] After captioning: {get_gpu_memory_info()}")
|
| 553 |
-
return captions
|
| 554 |
-
|
| 555 |
|
| 556 |
-
|
| 557 |
-
|
| 558 |
-
|
| 559 |
-
|
| 560 |
-
|
| 561 |
-
image_paths = []
|
| 562 |
|
| 563 |
-
|
| 564 |
-
|
| 565 |
-
|
| 566 |
-
|
| 567 |
-
|
| 568 |
-
|
| 569 |
-
|
| 570 |
-
|
| 571 |
-
|
| 572 |
-
|
| 573 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 574 |
else:
|
| 575 |
-
|
| 576 |
|
| 577 |
-
|
| 578 |
-
|
| 579 |
-
|
| 580 |
-
|
| 581 |
-
|
| 582 |
-
|
| 583 |
-
|
| 584 |
-
|
| 585 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 586 |
else:
|
| 587 |
-
|
| 588 |
-
|
| 589 |
-
|
| 590 |
-
|
| 591 |
-
|
|
|
|
|
|
|
| 592 |
|
| 593 |
-
|
| 594 |
-
|
| 595 |
-
|
| 596 |
-
def compute_flow_matching_loss(model_output, target, timesteps):
|
| 597 |
-
"""Compute Rectified Flow loss"""
|
| 598 |
-
loss = torch.nn.functional.mse_loss(model_output, target, reduction="none")
|
| 599 |
-
loss = loss.mean(dim=list(range(1, len(loss.shape))))
|
| 600 |
-
return loss.mean()
|
| 601 |
|
| 602 |
-
|
| 603 |
-
|
| 604 |
-
"""Upload to HF Hub"""
|
| 605 |
-
if not HF_TOKEN:
|
| 606 |
-
return False, "HF_TOKEN not configured"
|
| 607 |
|
| 608 |
-
|
| 609 |
-
|
| 610 |
-
|
| 611 |
-
if not username:
|
| 612 |
-
return False, "Could not get username"
|
| 613 |
-
|
| 614 |
-
repo_id = f"{username}/{repo_name}"
|
| 615 |
-
|
| 616 |
-
try:
|
| 617 |
-
create_repo(repo_id=repo_id, token=HF_TOKEN, private=True, repo_type="model", exist_ok=True)
|
| 618 |
-
except:
|
| 619 |
-
pass
|
| 620 |
|
| 621 |
-
|
| 622 |
-
|
| 623 |
-
|
| 624 |
-
|
| 625 |
-
|
| 626 |
-
|
|
|
|
|
|
|
| 627 |
|
| 628 |
-
|
| 629 |
-
|
| 630 |
-
|
| 631 |
-
|
| 632 |
-
|
| 633 |
-
|
| 634 |
-
|
| 635 |
-
|
| 636 |
-
|
| 637 |
-
|
| 638 |
-
|
| 639 |
-
|
| 640 |
-
|
| 641 |
-
|
| 642 |
-
|
| 643 |
-
|
| 644 |
-
|
| 645 |
-
|
| 646 |
-
|
| 647 |
-
resolution, batch_size, upload_to_hub_flag, hub_repo_name,
|
| 648 |
-
use_auto_caption, is_person_training, progress=gr.Progress()
|
| 649 |
-
):
|
| 650 |
-
"""Train LoRA with CORRECT ZImageTransformer2DModel forward signature"""
|
| 651 |
-
|
| 652 |
-
if not torch.cuda.is_available():
|
| 653 |
-
return None, "❌ No GPU available"
|
| 654 |
-
|
| 655 |
-
if not images or len(images) < 3:
|
| 656 |
-
return None, "❌ Please upload at least 3 images"
|
| 657 |
-
|
| 658 |
-
if not trigger_word:
|
| 659 |
-
return None, "❌ Please specify a trigger word"
|
| 660 |
-
|
| 661 |
-
if not output_name:
|
| 662 |
-
output_name = "z_image_lora"
|
| 663 |
-
output_name = output_name.replace(" ", "_").lower()
|
| 664 |
-
|
| 665 |
-
if upload_to_hub_flag and not HF_TOKEN:
|
| 666 |
-
return None, "❌ HF_TOKEN not configured"
|
| 667 |
-
|
| 668 |
-
progress(0, desc="Initializing...")
|
| 669 |
-
print(f"[Train] Start: {get_gpu_memory_info()}")
|
| 670 |
-
aggressive_cleanup()
|
| 671 |
-
|
| 672 |
-
with tempfile.TemporaryDirectory() as tmpdir:
|
| 673 |
-
try:
|
| 674 |
-
# ============================================
|
| 675 |
-
# PHASE 1: Captioning (Subprocess)
|
| 676 |
-
# ============================================
|
| 677 |
-
progress(0.02, desc="Running Florence-2 captioning (subprocess)...")
|
| 678 |
-
|
| 679 |
-
image_paths, captions, dataset_dir = prepare_dataset(
|
| 680 |
-
images, trigger_word, tmpdir, use_auto_caption, is_person_training
|
| 681 |
-
)
|
| 682 |
-
|
| 683 |
-
if len(image_paths) < 3:
|
| 684 |
-
return None, "❌ Not enough valid images"
|
| 685 |
-
|
| 686 |
-
progress(0.12, desc=f"Captioning done: {len(image_paths)} images")
|
| 687 |
-
aggressive_cleanup()
|
| 688 |
-
print(f"[Train] After captioning cleanup: {get_gpu_memory_info()}")
|
| 689 |
-
|
| 690 |
-
# ============================================
|
| 691 |
-
# PHASE 2: Load Pipeline for Text Encoding
|
| 692 |
-
# ============================================
|
| 693 |
-
progress(0.15, desc="Loading pipeline for encoding...")
|
| 694 |
-
print(f"[Train] Before pipeline: {get_gpu_memory_info()}")
|
| 695 |
-
|
| 696 |
-
from diffusers import ZImagePipeline
|
| 697 |
-
|
| 698 |
-
# Load pipeline to CPU first
|
| 699 |
-
pipe = ZImagePipeline.from_pretrained(
|
| 700 |
-
MODEL_REPO,
|
| 701 |
-
torch_dtype=DTYPE,
|
| 702 |
-
)
|
| 703 |
-
|
| 704 |
-
# Get VAE scaling factor
|
| 705 |
-
vae_scaling_factor = pipe.vae.config.scaling_factor
|
| 706 |
-
|
| 707 |
-
# ============================================
|
| 708 |
-
# PHASE 3: Encode Captions with Text Encoder
|
| 709 |
-
# ============================================
|
| 710 |
-
progress(0.20, desc="Encoding captions...")
|
| 711 |
-
|
| 712 |
-
# Move text encoder to GPU
|
| 713 |
-
pipe.text_encoder.to(DEVICE)
|
| 714 |
-
|
| 715 |
-
cached_text_embeddings = []
|
| 716 |
-
|
| 717 |
-
with torch.no_grad():
|
| 718 |
-
for idx, caption in enumerate(captions):
|
| 719 |
-
text_inputs = pipe.tokenizer(
|
| 720 |
-
caption,
|
| 721 |
-
padding="max_length",
|
| 722 |
-
max_length=256,
|
| 723 |
-
truncation=True,
|
| 724 |
-
return_tensors="pt"
|
| 725 |
-
).to(DEVICE)
|
| 726 |
-
|
| 727 |
-
text_emb = pipe.text_encoder(**text_inputs)[0]
|
| 728 |
-
cached_text_embeddings.append(text_emb.cpu())
|
| 729 |
-
|
| 730 |
-
del text_inputs, text_emb
|
| 731 |
-
|
| 732 |
-
if idx % 2 == 0:
|
| 733 |
-
torch.cuda.empty_cache()
|
| 734 |
-
progress(0.20 + 0.10 * (idx / len(captions)),
|
| 735 |
-
desc=f"Encoding caption {idx+1}/{len(captions)}")
|
| 736 |
-
|
| 737 |
-
# Free text encoder
|
| 738 |
-
pipe.text_encoder.to("cpu")
|
| 739 |
-
del pipe.text_encoder
|
| 740 |
-
aggressive_cleanup()
|
| 741 |
-
print(f"[Train] After text encoding: {get_gpu_memory_info()}")
|
| 742 |
-
|
| 743 |
-
# ============================================
|
| 744 |
-
# PHASE 4: Encode Images with VAE
|
| 745 |
-
# ============================================
|
| 746 |
-
progress(0.32, desc="Encoding images with VAE...")
|
| 747 |
-
|
| 748 |
-
pipe.vae.to(DEVICE)
|
| 749 |
-
|
| 750 |
-
cached_latents = []
|
| 751 |
-
|
| 752 |
-
with torch.no_grad():
|
| 753 |
-
for idx, img_path in enumerate(image_paths):
|
| 754 |
-
img = Image.open(img_path).convert("RGB")
|
| 755 |
-
img = img.resize((int(resolution), int(resolution)), Image.LANCZOS)
|
| 756 |
-
img_tensor = torch.from_numpy(np.array(img)).permute(2, 0, 1).float() / 255.0
|
| 757 |
-
img_tensor = img_tensor.unsqueeze(0).to(DEVICE, dtype=DTYPE)
|
| 758 |
-
img_tensor = 2.0 * img_tensor - 1.0
|
| 759 |
-
|
| 760 |
-
latent = pipe.vae.encode(img_tensor).latent_dist.sample()
|
| 761 |
-
latent = latent * vae_scaling_factor
|
| 762 |
-
cached_latents.append(latent.cpu())
|
| 763 |
-
|
| 764 |
-
del img_tensor, latent, img
|
| 765 |
-
|
| 766 |
-
if idx % 2 == 0:
|
| 767 |
-
torch.cuda.empty_cache()
|
| 768 |
-
progress(0.32 + 0.08 * (idx / len(image_paths)),
|
| 769 |
-
desc=f"Encoding image {idx+1}/{len(image_paths)}")
|
| 770 |
-
|
| 771 |
-
# Free VAE
|
| 772 |
-
pipe.vae.to("cpu")
|
| 773 |
-
del pipe.vae
|
| 774 |
-
aggressive_cleanup()
|
| 775 |
-
print(f"[Train] After VAE encoding: {get_gpu_memory_info()}")
|
| 776 |
-
|
| 777 |
-
# ============================================
|
| 778 |
-
# PHASE 5: Setup Transformer with Training Adapter
|
| 779 |
-
# ============================================
|
| 780 |
-
progress(0.42, desc="Setting up transformer with training adapter...")
|
| 781 |
-
|
| 782 |
-
# Download training adapter
|
| 783 |
-
try:
|
| 784 |
-
adapter_path = hf_hub_download(
|
| 785 |
-
repo_id="ostris/zimage_turbo_training_adapter",
|
| 786 |
-
filename="zimage_turbo_training_adapter_v1.safetensors",
|
| 787 |
-
local_dir=tmpdir
|
| 788 |
-
)
|
| 789 |
-
print(f"[Train] Training adapter downloaded: {adapter_path}")
|
| 790 |
-
except Exception as e:
|
| 791 |
-
return None, f"❌ Could not download training adapter: {e}"
|
| 792 |
-
|
| 793 |
-
# Get transformer (still on CPU from pipeline)
|
| 794 |
-
transformer = pipe.transformer
|
| 795 |
-
|
| 796 |
-
# Load adapter via pipe's load_lora_weights
|
| 797 |
-
from safetensors.torch import load_file, save_file
|
| 798 |
-
|
| 799 |
try:
|
| 800 |
-
|
| 801 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 802 |
except Exception as e:
|
| 803 |
-
|
| 804 |
-
|
| 805 |
-
|
| 806 |
-
|
| 807 |
-
|
| 808 |
-
|
| 809 |
-
|
| 810 |
-
|
| 811 |
-
|
| 812 |
-
|
| 813 |
-
|
| 814 |
-
|
| 815 |
-
|
| 816 |
-
|
| 817 |
-
|
| 818 |
-
|
| 819 |
-
|
| 820 |
-
|
| 821 |
-
|
| 822 |
-
|
| 823 |
-
|
| 824 |
-
|
| 825 |
-
|
| 826 |
-
|
| 827 |
-
|
| 828 |
-
|
| 829 |
-
|
| 830 |
-
|
| 831 |
-
|
| 832 |
-
|
| 833 |
-
|
| 834 |
-
|
| 835 |
-
|
| 836 |
-
|
| 837 |
-
|
| 838 |
-
|
| 839 |
-
|
| 840 |
-
|
| 841 |
-
|
| 842 |
-
|
| 843 |
-
|
| 844 |
-
|
| 845 |
-
|
| 846 |
-
|
| 847 |
-
|
| 848 |
-
optimizer, T_0=max(1, int(num_steps - warmup_steps)), eta_min=learning_rate * 0.01
|
| 849 |
-
)
|
| 850 |
-
lr_scheduler = SequentialLR(
|
| 851 |
-
optimizer, [warmup_scheduler, cosine_scheduler], milestones=[warmup_steps]
|
| 852 |
-
)
|
| 853 |
-
|
| 854 |
-
progress(0.50, desc=f"Training with {len(cached_latents)} samples...")
|
| 855 |
-
|
| 856 |
-
# ============================================
|
| 857 |
-
# PHASE 6: Training Loop with CORRECT forward signature
|
| 858 |
-
# ZImageTransformer2DModel.forward(x, t, cap_feats, ...)
|
| 859 |
-
# x: List[Tensor] where each tensor is [C, F, H, W]
|
| 860 |
-
# cap_feats: List[Tensor] where each tensor is [seq_len, dim]
|
| 861 |
-
# ============================================
|
| 862 |
-
transformer.train()
|
| 863 |
-
losses = []
|
| 864 |
-
successful_steps = 0
|
| 865 |
-
|
| 866 |
-
for step in range(int(num_steps)):
|
| 867 |
-
try:
|
| 868 |
-
idx = np.random.randint(0, len(cached_latents))
|
| 869 |
-
|
| 870 |
-
# latents: [B, C, H, W] -> need [C, F, H, W] where F=1
|
| 871 |
-
latents = cached_latents[idx].to(DEVICE, dtype=DTYPE)
|
| 872 |
-
# Remove batch dim, add frame dim: [1, C, H, W] -> [C, 1, H, W]
|
| 873 |
-
latents = latents.squeeze(0).unsqueeze(1) # [C, 1, H, W]
|
| 874 |
-
|
| 875 |
-
# text_embeddings: [B, seq_len, dim] -> [seq_len, dim]
|
| 876 |
-
text_embeddings = cached_text_embeddings[idx].to(DEVICE, dtype=DTYPE)
|
| 877 |
-
text_embeddings = text_embeddings.squeeze(0) # [seq_len, dim]
|
| 878 |
-
|
| 879 |
-
# Timestep for flow matching (0 to 1)
|
| 880 |
-
timesteps = torch.rand(1, device=DEVICE, dtype=DTYPE)
|
| 881 |
-
|
| 882 |
-
# Create noisy latents using flow matching interpolation
|
| 883 |
-
noise = torch.randn_like(latents)
|
| 884 |
-
t = timesteps.view(-1, 1, 1, 1)
|
| 885 |
-
noisy_latents = (1 - t) * latents + t * noise
|
| 886 |
-
|
| 887 |
-
# Target is the velocity: noise - clean
|
| 888 |
-
target = noise - latents
|
| 889 |
-
|
| 890 |
-
# Scale timestep for model
|
| 891 |
-
t_input = timesteps * 1000
|
| 892 |
-
|
| 893 |
-
# CORRECT FORWARD CALL:
|
| 894 |
-
# x and cap_feats must be Lists!
|
| 895 |
-
with torch.amp.autocast('cuda', dtype=DTYPE):
|
| 896 |
-
output = transformer(
|
| 897 |
-
x=[noisy_latents], # List of [C, F, H, W]
|
| 898 |
-
t=t_input, # timestep
|
| 899 |
-
cap_feats=[text_embeddings], # List of [seq_len, dim]
|
| 900 |
-
return_dict=True
|
| 901 |
-
)
|
| 902 |
-
|
| 903 |
-
# Get model output - it will also be a list
|
| 904 |
-
if hasattr(output, 'sample'):
|
| 905 |
-
model_output = output.sample
|
| 906 |
-
if isinstance(model_output, list):
|
| 907 |
-
model_output = model_output[0]
|
| 908 |
-
elif isinstance(output, tuple):
|
| 909 |
-
model_output = output[0]
|
| 910 |
-
if isinstance(model_output, list):
|
| 911 |
-
model_output = model_output[0]
|
| 912 |
-
else:
|
| 913 |
-
model_output = output
|
| 914 |
-
if isinstance(model_output, list):
|
| 915 |
-
model_output = model_output[0]
|
| 916 |
-
|
| 917 |
-
loss = compute_flow_matching_loss(model_output, target, timesteps)
|
| 918 |
-
|
| 919 |
-
optimizer.zero_grad()
|
| 920 |
-
loss.backward()
|
| 921 |
-
torch.nn.utils.clip_grad_norm_(transformer.parameters(), 1.0)
|
| 922 |
-
optimizer.step()
|
| 923 |
-
lr_scheduler.step()
|
| 924 |
-
|
| 925 |
-
losses.append(loss.item())
|
| 926 |
-
successful_steps += 1
|
| 927 |
-
|
| 928 |
-
del latents, text_embeddings, noise, noisy_latents, target, model_output, loss, output
|
| 929 |
-
|
| 930 |
-
if step % 25 == 0:
|
| 931 |
-
avg_loss = np.mean(losses[-50:]) if len(losses) >= 50 else np.mean(losses) if losses else float('nan')
|
| 932 |
-
progress(
|
| 933 |
-
0.50 + 0.40 * (step / int(num_steps)),
|
| 934 |
-
desc=f"Step {step}/{int(num_steps)} | Loss: {avg_loss:.4f}"
|
| 935 |
-
)
|
| 936 |
-
print(f"[Train] Step {step}: Loss={avg_loss:.4f}")
|
| 937 |
-
|
| 938 |
-
if step % 100 == 0:
|
| 939 |
-
gc.collect()
|
| 940 |
-
torch.cuda.empty_cache()
|
| 941 |
-
|
| 942 |
-
except Exception as e:
|
| 943 |
-
if step < 5:
|
| 944 |
-
print(f"[Train] Error at step {step}: {e}")
|
| 945 |
-
import traceback
|
| 946 |
-
traceback.print_exc()
|
| 947 |
-
continue
|
| 948 |
-
|
| 949 |
-
if successful_steps == 0:
|
| 950 |
-
return None, "❌ Training failed - no successful steps. Check model forward signature."
|
| 951 |
-
|
| 952 |
-
# ============================================
|
| 953 |
-
# PHASE 7: Save LoRA
|
| 954 |
-
# ============================================
|
| 955 |
-
progress(0.92, desc="Saving LoRA...")
|
| 956 |
-
|
| 957 |
-
del cached_latents, cached_text_embeddings
|
| 958 |
-
aggressive_cleanup()
|
| 959 |
-
|
| 960 |
-
lora_state_dict = {}
|
| 961 |
-
for name, param in transformer.named_parameters():
|
| 962 |
-
if "lora" in name.lower() and param.requires_grad:
|
| 963 |
-
clean_name = name.replace("base_model.model.", "")
|
| 964 |
-
lora_state_dict[clean_name] = param.detach().cpu()
|
| 965 |
-
|
| 966 |
-
if not lora_state_dict:
|
| 967 |
-
return None, "❌ No LoRA weights found"
|
| 968 |
-
|
| 969 |
-
final_output = f"/tmp/{output_name}.safetensors"
|
| 970 |
-
save_file(lora_state_dict, final_output)
|
| 971 |
-
|
| 972 |
-
file_size = os.path.getsize(final_output) / (1024 * 1024)
|
| 973 |
-
avg_final_loss = np.mean(losses[-100:]) if len(losses) >= 100 else np.mean(losses) if losses else float('nan')
|
| 974 |
-
|
| 975 |
-
training_info = f"""
|
| 976 |
-
- Images: {len(image_paths)}
|
| 977 |
-
- Steps: {successful_steps}
|
| 978 |
-
- Final Loss: {avg_final_loss:.4f}
|
| 979 |
-
- LR: {learning_rate}, Rank: {int(lora_rank)}, Resolution: {int(resolution)}
|
| 980 |
-
"""
|
| 981 |
-
|
| 982 |
-
hub_result = ""
|
| 983 |
-
if upload_to_hub_flag:
|
| 984 |
-
progress(0.94, desc="Uploading to Hub...")
|
| 985 |
-
success, result = upload_to_hub(
|
| 986 |
-
final_output, hub_repo_name or output_name, trigger_word, training_info
|
| 987 |
-
)
|
| 988 |
-
hub_result = f"\n\n🚀 Uploaded: {result}" if success else f"\n\n⚠️ Upload failed: {result}"
|
| 989 |
-
|
| 990 |
-
del transformer
|
| 991 |
-
aggressive_cleanup()
|
| 992 |
-
progress(1.0, desc="Complete!")
|
| 993 |
-
|
| 994 |
-
sample_captions = "\n".join([f" - {c[:80]}..." for c in captions[:3]])
|
| 995 |
-
|
| 996 |
-
return final_output, f"""✅ Training complete!
|
| 997 |
-
|
| 998 |
-
📁 LoRA: {output_name}.safetensors ({file_size:.1f} MB)
|
| 999 |
-
🏷️ Trigger: {trigger_word}
|
| 1000 |
-
📊 Loss: {avg_final_loss:.4f}
|
| 1001 |
-
🖼️ Images: {len(image_paths)}
|
| 1002 |
-
⚙️ Steps: {successful_steps}
|
| 1003 |
-
|
| 1004 |
-
**Sample captions:**
|
| 1005 |
-
{sample_captions}{hub_result}
|
| 1006 |
-
|
| 1007 |
-
**Usage:**
|
| 1008 |
-
```python
|
| 1009 |
-
from diffusers import ZImagePipeline
|
| 1010 |
-
import torch
|
| 1011 |
-
|
| 1012 |
-
pipe = ZImagePipeline.from_pretrained("Tongyi-MAI/Z-Image-Turbo", torch_dtype=torch.bfloat16)
|
| 1013 |
-
pipe.load_lora_weights("{output_name}.safetensors")
|
| 1014 |
-
pipe.to("cuda")
|
| 1015 |
-
|
| 1016 |
-
image = pipe("{trigger_word}, your prompt here", num_inference_steps=8, guidance_scale=0.0).images[0]
|
| 1017 |
-
```"""
|
| 1018 |
-
|
| 1019 |
except Exception as e:
|
| 1020 |
-
|
| 1021 |
-
|
| 1022 |
-
|
|
|
|
| 1023 |
|
|
|
|
|
|
|
| 1024 |
|
| 1025 |
-
|
| 1026 |
-
|
| 1027 |
-
|
| 1028 |
-
with gr.Blocks(title="Z-IMAGE GEN/LORA") as demo:
|
| 1029 |
|
| 1030 |
# HOME Button
|
| 1031 |
gr.HTML("""
|
|
@@ -1041,138 +771,90 @@ with gr.Blocks(title="Z-IMAGE GEN/LORA") as demo:
|
|
| 1041 |
gr.HTML("""
|
| 1042 |
<div class="header-container">
|
| 1043 |
<div class="header-title">🎨 Z-IMAGE GEN/LORA 🎨</div>
|
| 1044 |
-
<div class="header-subtitle">
|
| 1045 |
<div style="margin-top:12px">
|
| 1046 |
-
<span class="stats-badge"
|
| 1047 |
-
<span class="stats-badge"
|
| 1048 |
-
<span class="stats-badge"
|
| 1049 |
-
<span class="stats-badge"
|
| 1050 |
</div>
|
| 1051 |
</div>
|
| 1052 |
""")
|
| 1053 |
-
|
| 1054 |
-
# Status Row
|
| 1055 |
-
gr.HTML('<div class="info-box">📊 <b>System Status</b> - Check GPU and HuggingFace connection</div>')
|
| 1056 |
|
|
|
|
| 1057 |
with gr.Row():
|
| 1058 |
-
with gr.Column(scale=
|
| 1059 |
-
|
| 1060 |
-
with gr.Column(scale=1):
|
| 1061 |
-
|
| 1062 |
-
refresh_btn = gr.Button("🔄 Refresh", size="sm")
|
| 1063 |
-
refresh_btn.click(fn=lambda: (check_gpu(), check_hf_token()), outputs=[gpu_status, hf_status])
|
| 1064 |
-
|
| 1065 |
with gr.Row():
|
| 1066 |
-
|
| 1067 |
-
|
| 1068 |
-
gr.
|
| 1069 |
-
|
| 1070 |
-
label="
|
| 1071 |
-
|
| 1072 |
-
|
| 1073 |
-
|
| 1074 |
-
interactive=True
|
| 1075 |
)
|
| 1076 |
-
gr.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1077 |
|
| 1078 |
-
# Right Column - Settings
|
| 1079 |
-
with gr.Column(scale=1):
|
| 1080 |
-
gr.HTML('<div class="info-box">⚙️ <b>Training Settings</b> - Configure your LoRA training</div>')
|
| 1081 |
-
|
| 1082 |
-
trigger_word = gr.Textbox(
|
| 1083 |
-
label="🏷️ Trigger Word",
|
| 1084 |
-
placeholder="ohwx_person",
|
| 1085 |
-
info="Use unique token like 'ohwx' to avoid conflicts"
|
| 1086 |
-
)
|
| 1087 |
-
output_name = gr.Textbox(
|
| 1088 |
-
label="📁 Output Name",
|
| 1089 |
-
placeholder="my_lora"
|
| 1090 |
-
)
|
| 1091 |
-
|
| 1092 |
-
with gr.Row():
|
| 1093 |
-
use_auto_caption = gr.Checkbox(label="🔤 Auto-Caption (Florence-2)", value=True)
|
| 1094 |
-
is_person_training = gr.Checkbox(label="👤 Person/Face Training", value=True)
|
| 1095 |
-
|
| 1096 |
-
with gr.Row():
|
| 1097 |
-
num_steps = gr.Slider(500, 5000, 1500, step=100, label="🔢 Steps")
|
| 1098 |
-
learning_rate = gr.Slider(1e-5, 5e-4, 5e-5, step=1e-5, label="📈 Learning Rate")
|
| 1099 |
-
|
| 1100 |
-
with gr.Row():
|
| 1101 |
-
lora_rank = gr.Slider(4, 64, 32, step=4, label="🎚️ LoRA Rank")
|
| 1102 |
-
resolution = gr.Slider(512, 1024, 1024, step=128, label="📐 Resolution")
|
| 1103 |
-
|
| 1104 |
-
batch_size = gr.Slider(1, 4, 1, step=1, visible=False)
|
| 1105 |
-
|
| 1106 |
-
gr.HTML('<div class="info-box">🚀 <b>Hub Upload</b> - Upload trained LoRA to HuggingFace</div>')
|
| 1107 |
-
|
| 1108 |
-
with gr.Row():
|
| 1109 |
-
upload_to_hub_flag = gr.Checkbox(label="📤 Upload to HF Hub (Private)", value=False)
|
| 1110 |
-
hub_repo_name = gr.Textbox(label="📦 Repo Name", placeholder="my-zimage-lora")
|
| 1111 |
-
|
| 1112 |
-
# Train Button
|
| 1113 |
-
with gr.Row():
|
| 1114 |
-
train_btn = gr.Button("🚀 START TRAINING!", variant="primary", size="lg")
|
| 1115 |
-
|
| 1116 |
-
# Output
|
| 1117 |
with gr.Row():
|
| 1118 |
-
with gr.
|
| 1119 |
-
|
| 1120 |
-
|
| 1121 |
-
|
| 1122 |
-
|
| 1123 |
-
|
| 1124 |
-
|
| 1125 |
-
|
| 1126 |
-
|
| 1127 |
-
|
| 1128 |
-
|
| 1129 |
-
|
| 1130 |
-
|
| 1131 |
-
|
| 1132 |
-
|
| 1133 |
-
|
| 1134 |
-
|
| 1135 |
-
<div class="info-box">
|
| 1136 |
-
📋 <b>Recommended Settings by Use Case</b>
|
| 1137 |
-
<table style="width:100%; margin-top:10px; border-collapse: collapse;">
|
| 1138 |
-
<tr style="background:#3B82F6; color:white;">
|
| 1139 |
-
<th style="padding:8px; border:2px solid #1F2937;">Use Case</th>
|
| 1140 |
-
<th style="padding:8px; border:2px solid #1F2937;">Steps</th>
|
| 1141 |
-
<th style="padding:8px; border:2px solid #1F2937;">LR</th>
|
| 1142 |
-
<th style="padding:8px; border:2px solid #1F2937;">Rank</th>
|
| 1143 |
-
</tr>
|
| 1144 |
-
<tr style="background:#FEF9C3;">
|
| 1145 |
-
<td style="padding:8px; border:2px solid #1F2937;">👤 Person</td>
|
| 1146 |
-
<td style="padding:8px; border:2px solid #1F2937;">1500</td>
|
| 1147 |
-
<td style="padding:8px; border:2px solid #1F2937;">5e-5</td>
|
| 1148 |
-
<td style="padding:8px; border:2px solid #1F2937;">32</td>
|
| 1149 |
-
</tr>
|
| 1150 |
-
<tr style="background:#FFF;">
|
| 1151 |
-
<td style="padding:8px; border:2px solid #1F2937;">🎨 Style</td>
|
| 1152 |
-
<td style="padding:8px; border:2px solid #1F2937;">2000</td>
|
| 1153 |
-
<td style="padding:8px; border:2px solid #1F2937;">1e-4</td>
|
| 1154 |
-
<td style="padding:8px; border:2px solid #1F2937;">16</td>
|
| 1155 |
-
</tr>
|
| 1156 |
-
<tr style="background:#FEF9C3;">
|
| 1157 |
-
<td style="padding:8px; border:2px solid #1F2937;">📦 Object</td>
|
| 1158 |
-
<td style="padding:8px; border:2px solid #1F2937;">1200</td>
|
| 1159 |
-
<td style="padding:8px; border:2px solid #1F2937;">8e-5</td>
|
| 1160 |
-
<td style="padding:8px; border:2px solid #1F2937;">24</td>
|
| 1161 |
-
</tr>
|
| 1162 |
-
</table>
|
| 1163 |
-
</div>
|
| 1164 |
-
""")
|
| 1165 |
|
| 1166 |
# Footer
|
| 1167 |
gr.HTML("""
|
| 1168 |
<div class="footer-comic">
|
| 1169 |
<p style="font-family:'Bangers',cursive;font-size:1.5rem;letter-spacing:2px">🎨 Z-IMAGE GEN/LORA 🎨</p>
|
| 1170 |
-
<p>Powered by Z-Image Turbo +
|
| 1171 |
-
<p
|
| 1172 |
<p style="margin-top:10px"><a href="https://www.ginigen.com" target="_blank" style="color:#FACC15;text-decoration:none;font-weight:bold;">🏠 www.ginigen.com</a></p>
|
| 1173 |
</div>
|
| 1174 |
""")
|
| 1175 |
|
| 1176 |
-
|
| 1177 |
-
|
| 1178 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
import json
|
| 3 |
+
import copy
|
| 4 |
+
import time
|
| 5 |
+
import requests
|
| 6 |
+
import random
|
| 7 |
+
import logging
|
|
|
|
|
|
|
|
|
|
| 8 |
import numpy as np
|
| 9 |
+
import spaces
|
| 10 |
+
from typing import Any, Dict, List, Optional, Union
|
| 11 |
|
| 12 |
+
import torch
|
| 13 |
+
from PIL import Image
|
| 14 |
+
import gradio as gr
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
+
from diffusers import (
|
| 17 |
+
DiffusionPipeline,
|
| 18 |
+
AutoencoderKL,
|
| 19 |
+
ZImagePipeline
|
| 20 |
+
)
|
| 21 |
|
| 22 |
+
from huggingface_hub import (
|
| 23 |
+
hf_hub_download,
|
| 24 |
+
HfFileSystem,
|
| 25 |
+
ModelCard,
|
| 26 |
+
snapshot_download)
|
| 27 |
|
| 28 |
+
from diffusers.utils import load_image
|
| 29 |
+
from typing import Iterable
|
| 30 |
|
| 31 |
# ============================================
|
| 32 |
# Comic Style CSS
|
|
|
|
| 200 |
color: #1F2937 !important;
|
| 201 |
}
|
| 202 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 203 |
label, .gr-input-label, .gr-block-label {
|
| 204 |
color: #1F2937 !important;
|
| 205 |
font-family: 'Comic Neue', cursive !important;
|
|
|
|
| 213 |
box-shadow: 4px 4px 0 #1F2937 !important;
|
| 214 |
}
|
| 215 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 216 |
.footer-comic {
|
| 217 |
text-align: center;
|
| 218 |
padding: 20px;
|
|
|
|
| 260 |
accent-color: #3B82F6;
|
| 261 |
}
|
| 262 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 263 |
/* Image/Gallery Container */
|
| 264 |
.gr-image, .gr-gallery {
|
| 265 |
border: 3px solid #1F2937 !important;
|
|
|
|
| 267 |
box-shadow: 4px 4px 0 #1F2937 !important;
|
| 268 |
}
|
| 269 |
|
| 270 |
+
/* Original CSS additions */
|
| 271 |
+
#gen_btn{height: 100%}
|
| 272 |
+
#gen_column{align-self: stretch}
|
| 273 |
+
#title{text-align: center}
|
| 274 |
+
#title h1{font-size: 3em; display:inline-flex; align-items:center}
|
| 275 |
+
#title img{width: 100px; margin-right: 0.5em}
|
| 276 |
+
#gallery .grid-wrap{height: 10vh}
|
| 277 |
+
#lora_list{background: var(--block-background-fill);padding: 0 1em .3em; font-size: 90%}
|
| 278 |
+
.card_internal{display: flex;height: 100px;margin-top: .5em}
|
| 279 |
+
.card_internal img{margin-right: 1em}
|
| 280 |
+
.styler{--form-gap-width: 0px !important}
|
| 281 |
+
#progress{height:30px}
|
| 282 |
+
#progress .generating{display:none}
|
| 283 |
+
.progress-container {width: 100%;height: 30px;background-color: #f0f0f0;border-radius: 15px;overflow: hidden;margin-bottom: 20px}
|
| 284 |
+
.progress-bar {height: 100%;background-color: #4f46e5;width: calc(var(--current) / var(--total) * 100%);transition: width 0.5s ease-in-out}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 285 |
"""
|
| 286 |
|
| 287 |
+
loras = [
|
| 288 |
+
# 로컬 jimin LoRA (app.py와 같은 디렉토리에 jimin.safetensors 필요)
|
| 289 |
+
{
|
| 290 |
+
"image": "https://i.namu.wiki/i/umL8EZtn0hs-nMRYeFxIrkGrMe-R1u5c9fJE8ufrLjvXz52VcSIbG7TT9QJoL2rR7vsFww1lLrE4bwfn5uOBzfq9a90HGdNdlTLmr_KoqOchTovbVC3RDzhDbp7FI-Wq-esCu7_BYIptqethL4onBg.webp",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 291 |
|
| 292 |
+
"title": "jimin Style",
|
| 293 |
+
"repo": "./", # 로컬 경로
|
| 294 |
+
"weights": "jimin.safetensors",
|
| 295 |
+
"trigger_word": "jimin"
|
| 296 |
+
},
|
| 297 |
+
{
|
| 298 |
+
"image": "https://huggingface.co/strangerzonehf/Flux-Ultimate-LoRA-Collection/resolve/main/images/1111111111.png",
|
| 299 |
+
"title": "AWPortrait Z",
|
| 300 |
+
"repo": "Shakker-Labs/AWPortrait-Z", #1
|
| 301 |
+
"weights": "AWPortrait-Z.safetensors",
|
| 302 |
+
"trigger_word": "Portrait"
|
| 303 |
+
},
|
| 304 |
+
|
| 305 |
+
{
|
| 306 |
+
"image": "https://cdn-uploads.huggingface.co/production/uploads/653cd3049107029eb004f968/DLCGlF9uUnFo5zxR5uyx6.png",
|
| 307 |
+
"title": "50s Western",
|
| 308 |
+
"repo": "neph1/50s_western_lora_zit",
|
| 309 |
+
"weights": "50s_western_z_100.safetensors",
|
| 310 |
+
"trigger_word": "50s_western"
|
| 311 |
+
},
|
| 312 |
+
{
|
| 313 |
+
"image": "https://huggingface.co/Sumitc13/Z-image-Turbo_LogC4_lora/resolve/main/images/1764464517272__000005000_1.jpg",
|
| 314 |
+
"title": "LogC4",
|
| 315 |
+
"repo": "Sumitc13/Z-image-Turbo_LogC4_lora", #30
|
| 316 |
+
"weights": "z-image-logc4_000005000.safetensors",
|
| 317 |
+
"trigger_word": "LogC4"
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"image": "https://huggingface.co/neph1/80s_scifi_lora_zit/resolve/main/images/ComfyUI_10288_.png",
|
| 321 |
+
"title": "80s Scifi",
|
| 322 |
+
"repo": "neph1/80s_scifi_lora_zit",
|
| 323 |
+
"weights": "80s_scifi_z_80.safetensors",
|
| 324 |
+
"trigger_word": "80s_scifi"
|
| 325 |
+
},
|
| 326 |
+
|
| 327 |
+
# --------------------------------------------------------------------------------------------------------------------------------------
|
| 328 |
+
{
|
| 329 |
+
"image": "https://huggingface.co/Ttio2/Z-Image-Turbo-pencil-sketch/resolve/main/images/z-image_00097_.png",
|
| 330 |
+
"title": "Turbo Pencil",
|
| 331 |
+
"repo": "Ttio2/Z-Image-Turbo-pencil-sketch", #0
|
| 332 |
+
"weights": "Zimage_pencil_sketch.safetensors",
|
| 333 |
+
"trigger_word": "pencil sketch"
|
| 334 |
+
},
|
| 335 |
+
{
|
| 336 |
+
"image": "https://huggingface.co/neph1/50s_scifi_lora_zit/resolve/main/images/ComfyUI_08067_.png",
|
| 337 |
+
"title": "50s Scifi",
|
| 338 |
+
"repo": "neph1/50s_scifi_lora_zit",
|
| 339 |
+
"weights": "50s_scifi_z_80.safetensors",
|
| 340 |
+
"trigger_word": "50s_scifi"
|
| 341 |
+
},
|
| 342 |
+
{
|
| 343 |
+
"image": "https://huggingface.co/strangerzonehf/Flux-Ultimate-LoRA-Collection/resolve/main/images/cookie-mons.png",
|
| 344 |
+
"title": "Yarn Art Style",
|
| 345 |
+
"repo": "linoyts/yarn-art-style", #28
|
| 346 |
+
"weights": "yarn-art-style_000001250.safetensors",
|
| 347 |
+
"trigger_word": "yarn art style"
|
| 348 |
+
},
|
| 349 |
+
{
|
| 350 |
+
"image": "https://huggingface.co/Quorlen/Z-Image-Turbo-Behind-Reeded-Glass-Lora/resolve/main/images/ComfyUI_00391_.png",
|
| 351 |
+
"title": "Behind Reeded Glass",
|
| 352 |
+
"repo": "Quorlen/Z-Image-Turbo-Behind-Reeded-Glass-Lora", #26
|
| 353 |
+
"weights": "Z_Image_Turbo_Behind_Reeded_Glass_Lora_TAV2_000002750.safetensors",
|
| 354 |
+
"trigger_word": "Act1vate!, Behind reeded glass"
|
| 355 |
+
},
|
| 356 |
+
{
|
| 357 |
+
"image": "https://huggingface.co/ostris/z_image_turbo_childrens_drawings/resolve/main/images/1764433619736__000003000_9.jpg",
|
| 358 |
+
"title": "Childrens Drawings",
|
| 359 |
+
"repo": "ostris/z_image_turbo_childrens_drawings", #2
|
| 360 |
+
"weights": "z_image_turbo_childrens_drawings.safetensors",
|
| 361 |
+
"trigger_word": "Children Drawings"
|
| 362 |
+
},
|
| 363 |
+
{
|
| 364 |
+
"image": "https://huggingface.co/strangerzonehf/Flux-Ultimate-LoRA-Collection/resolve/main/images/xcxc.png",
|
| 365 |
+
"title": "Tarot Z",
|
| 366 |
+
"repo": "multimodalart/tarot-z-image-lora", #22
|
| 367 |
+
"weights": "tarot-z-image_000001250.safetensors",
|
| 368 |
+
"trigger_word": "trtcrd"
|
| 369 |
+
},
|
| 370 |
+
{
|
| 371 |
+
"image": "https://huggingface.co/renderartist/Technically-Color-Z-Image-Turbo/resolve/main/images/ComfyUI_00917_.png",
|
| 372 |
+
"title": "Technically Color Z",
|
| 373 |
+
"repo": "renderartist/Technically-Color-Z-Image-Turbo", #3
|
| 374 |
+
"weights": "Technically_Color_Z_Image_Turbo_v1_renderartist_2000.safetensors",
|
| 375 |
+
"trigger_word": "t3chnic4lly"
|
| 376 |
+
},
|
| 377 |
+
{
|
| 378 |
+
"image": "https://huggingface.co/SkyAsl/Tattoo-artist-Z/resolve/main/images/a%20dragon%20with%20flames.png",
|
| 379 |
+
"title": "Tattoo-artist-Z",
|
| 380 |
+
"repo": "SkyAsl/Tattoo-artist-Z", #31
|
| 381 |
+
"weights": "adapter_model.safetensors",
|
| 382 |
+
"trigger_word": "a tattoo design"
|
| 383 |
+
},
|
| 384 |
+
{
|
| 385 |
+
"image": "https://huggingface.co/strangerzonehf/Flux-Ultimate-LoRA-Collection/resolve/main/images/z-image_00147_.png",
|
| 386 |
+
"title": "Turbo Ghibli",
|
| 387 |
+
"repo": "Ttio2/Z-Image-Turbo-Ghibli-Style", #19
|
| 388 |
+
"weights": "ghibli_zimage_finetune.safetensors",
|
| 389 |
+
"trigger_word": "Ghibli Style"
|
| 390 |
+
},
|
| 391 |
+
{
|
| 392 |
+
"image": "https://huggingface.co/tarn59/pixel_art_style_lora_z_image_turbo/resolve/main/images/ComfyUI_00273_.png",
|
| 393 |
+
"title": "Pixel Art",
|
| 394 |
+
"repo": "tarn59/pixel_art_style_lora_z_image_turbo", #4
|
| 395 |
+
"weights": "pixel_art_style_z_image_turbo.safetensors",
|
| 396 |
+
"trigger_word": "Pixel art style."
|
| 397 |
+
},
|
| 398 |
+
{
|
| 399 |
+
"image": "https://huggingface.co/renderartist/Saturday-Morning-Z-Image-Turbo/resolve/main/images/Saturday_Morning_Z_15.png",
|
| 400 |
+
"title": "Saturday Morning",
|
| 401 |
+
"repo": "renderartist/Saturday-Morning-Z-Image-Turbo", #5
|
| 402 |
+
"weights": "Saturday_Morning_Z_Image_Turbo_v1_renderartist_1250.safetensors",
|
| 403 |
+
"trigger_word": "saturd4ym0rning"
|
| 404 |
+
},
|
| 405 |
+
{
|
| 406 |
+
"image": "https://huggingface.co/AIImageStudio/ReversalFilmGravure_z_Image_turbo/resolve/main/images/2025-12-01_173047-z_image_z_image_turbo_bf16-435125750859057-euler_10_hires.png",
|
| 407 |
+
"title": "ReversalFilmGravure",
|
| 408 |
+
"repo": "AIImageStudio/ReversalFilmGravure_z_Image_turbo", #6
|
| 409 |
+
"weights": "z_image_turbo_ReversalFilmGravure_v1.0.safetensors",
|
| 410 |
+
"trigger_word": "Reversal Film Gravure, analog film photography"
|
| 411 |
+
},
|
| 412 |
+
{
|
| 413 |
+
"image": "https://huggingface.co/renderartist/Coloring-Book-Z-Image-Turbo-LoRA/resolve/main/images/CBZ_00274_.png",
|
| 414 |
+
"title": "Coloring Book Z",
|
| 415 |
+
"repo": "renderartist/Coloring-Book-Z-Image-Turbo-LoRA", #7
|
| 416 |
+
"weights": "Coloring_Book_Z_Image_Turbo_v1_renderartist_2000.safetensors",
|
| 417 |
+
"trigger_word": "c0l0ringb00k"
|
| 418 |
+
},
|
| 419 |
+
{
|
| 420 |
+
"image": "https://huggingface.co/damnthatai/1950s_American_Dream/resolve/main/images/ZImage_20251129163459_135x_00001_.jpg",
|
| 421 |
+
"title": "1950s American Dream",
|
| 422 |
+
"repo": "damnthatai/1950s_American_Dream", #8
|
| 423 |
+
"weights": "5os4m3r1c4n4_z.safetensors",
|
| 424 |
+
"trigger_word": "5os4m3r1c4n4, 1950s, painting, a painting of"
|
| 425 |
+
},
|
| 426 |
+
{
|
| 427 |
+
"image": "https://huggingface.co/wcde/Z-Image-Turbo-DeJPEG-Lora/resolve/main/images/01.png",
|
| 428 |
+
"title": "DeJPEG",
|
| 429 |
+
"repo": "wcde/Z-Image-Turbo-DeJPEG-Lora", #9
|
| 430 |
+
"weights": "dejpeg_v3.safetensors",
|
| 431 |
+
"trigger_word": ""
|
| 432 |
+
},
|
| 433 |
+
{
|
| 434 |
+
"image": "https://huggingface.co/suayptalha/Z-Image-Turbo-Realism-LoRA/resolve/main/images/n4aSpqa-YFXYo4dtcIg4W.png",
|
| 435 |
+
"title": "DeJPEG",
|
| 436 |
+
"repo": "suayptalha/Z-Image-Turbo-Realism-LoRA", #10
|
| 437 |
+
"weights": "pytorch_lora_weights.safetensors",
|
| 438 |
+
"trigger_word": "Realism"
|
| 439 |
+
},
|
| 440 |
+
{
|
| 441 |
+
"image": "https://huggingface.co/renderartist/Classic-Painting-Z-Image-Turbo-LoRA/resolve/main/images/Classic_Painting_Z_00247_.png",
|
| 442 |
+
"title": "Classic Painting Z",
|
| 443 |
+
"repo": "renderartist/Classic-Painting-Z-Image-Turbo-LoRA", #11
|
| 444 |
+
"weights": "Classic_Painting_Z_Image_Turbo_v1_renderartist_1750.safetensors",
|
| 445 |
+
"trigger_word": "class1cpa1nt"
|
| 446 |
+
},
|
| 447 |
+
{
|
| 448 |
+
"image": "https://huggingface.co/DK9/3D_MMORPG_style_z-image-turbo_lora/resolve/main/images/10_with_lora.png",
|
| 449 |
+
"title": "3D MMORPG",
|
| 450 |
+
"repo": "DK9/3D_MMORPG_style_z-image-turbo_lora", #12
|
| 451 |
+
"weights": "lostark_v1.safetensors",
|
| 452 |
+
"trigger_word": ""
|
| 453 |
+
},
|
| 454 |
+
{
|
| 455 |
+
"image": "https://huggingface.co/Danrisi/Olympus_UltraReal_ZImage/resolve/main/images/Z-Image_01011_.png",
|
| 456 |
+
"title": "Olympus UltraReal",
|
| 457 |
+
"repo": "Danrisi/Olympus_UltraReal_ZImage", #13
|
| 458 |
+
"weights": "Olympus.safetensors",
|
| 459 |
+
"trigger_word": "digital photography, early 2000s compact camera aesthetic, amateur candid shot, digital photography, early 2000s compact camera aesthetic, amateur candid shot, direct flash lighting, hard flash shadow, specular highlights, overexposed highlights"
|
| 460 |
+
},
|
| 461 |
+
{
|
| 462 |
+
"image": "https://huggingface.co/AiAF/D-ART_Z-Image-Turbo_LoRA/resolve/main/images/example_l3otpwzaz.png",
|
| 463 |
+
"title": "D ART Z Image",
|
| 464 |
+
"repo": "AiAF/D-ART_Z-Image-Turbo_LoRA", #14
|
| 465 |
+
"weights": "D-ART_Z-Image-Turbo.safetensors",
|
| 466 |
+
"trigger_word": "D-ART"
|
| 467 |
+
},
|
| 468 |
+
{
|
| 469 |
+
"image": "https://huggingface.co/AlekseyCalvin/Marionette_Modernism_Z-image-Turbo_LoRA/resolve/main/bluebirdmandoll.webp",
|
| 470 |
+
"title": "Marionette Modernism",
|
| 471 |
+
"repo": "AlekseyCalvin/Marionette_Modernism_Z-image-Turbo_LoRA", #15
|
| 472 |
+
"weights": "ZImageDadadoll_000003600.safetensors",
|
| 473 |
+
"trigger_word": "DADADOLL style"
|
| 474 |
+
},
|
| 475 |
+
{
|
| 476 |
+
"image": "https://huggingface.co/AlekseyCalvin/HistoricColor_Z-image-Turbo-LoRA/resolve/main/HSTZgen2.webp",
|
| 477 |
+
"title": "Historic Color Z",
|
| 478 |
+
"repo": "AlekseyCalvin/HistoricColor_Z-image-Turbo-LoRA", #16
|
| 479 |
+
"weights": "ZImage1HST_000004000.safetensors",
|
| 480 |
+
"trigger_word": "HST style"
|
| 481 |
+
},
|
| 482 |
+
{
|
| 483 |
+
"image": "https://huggingface.co/tarn59/80s_air_brush_style_z_image_turbo/resolve/main/images/ComfyUI_00707_.png",
|
| 484 |
+
"title": "80s Air Brush",
|
| 485 |
+
"repo": "tarn59/80s_air_brush_style_z_image_turbo", #17
|
| 486 |
+
"weights": "80s_air_brush_style_v2_z_image_turbo.safetensors",
|
| 487 |
+
"trigger_word": "80s Air Brush style."
|
| 488 |
+
},
|
| 489 |
+
{
|
| 490 |
+
"image": "https://huggingface.co/CedarC/Z-Image_360/resolve/main/images/1765505225357__000006750_6.jpg",
|
| 491 |
+
"title": "360panorama",
|
| 492 |
+
"repo": "CedarC/Z-Image_360", #18
|
| 493 |
+
"weights": "Z-Image_360.safetensors",
|
| 494 |
+
"trigger_word": "360panorama"
|
| 495 |
+
},
|
| 496 |
+
{
|
| 497 |
+
"image": "https://huggingface.co/HAV0X1014/Z-Image-Turbo-KF-Bat-Eared-Fox-LoRA/resolve/main/images/ComfyUI_00132_.png",
|
| 498 |
+
"title": "KF-Bat-Eared",
|
| 499 |
+
"repo": "HAV0X1014/Z-Image-Turbo-KF-Bat-Eared-Fox-LoRA", #21
|
| 500 |
+
"weights": "z-image-turbo-bat_eared_fox.safetensors",
|
| 501 |
+
"trigger_word": "bat_eared_fox_kemono_friends"
|
| 502 |
+
},
|
| 503 |
+
{
|
| 504 |
+
"image": "https://cdn-uploads.huggingface.co/production/uploads/653cd3049107029eb004f968/IHttgddXu6ZBMo7eyy8p6.png",
|
| 505 |
+
"title": "80s Horror",
|
| 506 |
+
"repo": "neph1/80s_horror_movies_lora_zit", #23
|
| 507 |
+
"weights": "80s_horror_z_80.safetensors",
|
| 508 |
+
"trigger_word": "80s_horror"
|
| 509 |
+
},
|
| 510 |
+
{
|
| 511 |
+
"image": "https://huggingface.co/Quorlen/z_image_turbo_Sunbleached_Protograph_Style_Lora/resolve/main/images/ComfyUI_00024_.png",
|
| 512 |
+
"title": "Sunbleached Protograph",
|
| 513 |
+
"repo": "Quorlen/z_image_turbo_Sunbleached_Protograph_Style_Lora", #24
|
| 514 |
+
"weights": "zimageturbo_Sunbleach_Photograph_Style_Lora_TAV2_000002750.safetensors",
|
| 515 |
+
"trigger_word": "Act1vate!"
|
| 516 |
+
},
|
| 517 |
+
{
|
| 518 |
+
"image": "https://huggingface.co/bunnycore/Z-Art-2.1/resolve/main/images/ComfyUI_00069_.png",
|
| 519 |
+
"title": "Z-Art-2.1",
|
| 520 |
+
"repo": "bunnycore/Z-Art-2.1", #25
|
| 521 |
+
"weights": "Z-Image-Art2.1.safetensors",
|
| 522 |
+
"trigger_word": "anime art"
|
| 523 |
+
},
|
| 524 |
+
{
|
| 525 |
+
"image": "https://huggingface.co/cactusfriend/longfurby-z/resolve/main/images/1764658860954__000003000_1.jpg",
|
| 526 |
+
"title": "Longfurby",
|
| 527 |
+
"repo": "cactusfriend/longfurby-z", #27
|
| 528 |
+
"weights": "longfurbyZ.safetensors",
|
| 529 |
+
"trigger_word": ""
|
| 530 |
+
},
|
| 531 |
+
{
|
| 532 |
+
"image": "https://huggingface.co/SkyAsl/Pixel-artist-Z/resolve/main/pixel-art-result.png",
|
| 533 |
+
"title": "Pixel Art",
|
| 534 |
+
"repo": "SkyAsl/Pixel-artist-Z", #29
|
| 535 |
+
"weights": "adapter_model.safetensors",
|
| 536 |
+
"trigger_word": "a pixel art character"
|
| 537 |
+
},
|
| 538 |
+
]
|
| 539 |
+
|
| 540 |
+
dtype = torch.bfloat16
|
| 541 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 542 |
+
base_model = "Tongyi-MAI/Z-Image-Turbo"
|
| 543 |
+
|
| 544 |
+
print(f"Loading {base_model} pipeline...")
|
| 545 |
+
|
| 546 |
+
# Initialize Pipeline
|
| 547 |
+
pipe = ZImagePipeline.from_pretrained(
|
| 548 |
+
base_model,
|
| 549 |
+
torch_dtype=dtype,
|
| 550 |
+
low_cpu_mem_usage=False,
|
| 551 |
+
).to(device)
|
| 552 |
+
|
| 553 |
+
# ======== AoTI compilation + FA3 ========
|
| 554 |
+
# As per reference for optimization
|
| 555 |
+
try:
|
| 556 |
+
print("Applying AoTI compilation and FA3...")
|
| 557 |
+
pipe.transformer.layers._repeated_blocks = ["ZImageTransformerBlock"]
|
| 558 |
+
spaces.aoti_blocks_load(pipe.transformer.layers, "zerogpu-aoti/Z-Image", variant="fa3")
|
| 559 |
+
print("Optimization applied successfully.")
|
| 560 |
+
except Exception as e:
|
| 561 |
+
print(f"Optimization warning: {e}. Continuing with standard pipeline.")
|
| 562 |
+
|
| 563 |
+
MAX_SEED = np.iinfo(np.int32).max
|
| 564 |
+
|
| 565 |
+
class calculateDuration:
|
| 566 |
+
def __init__(self, activity_name=""):
|
| 567 |
+
self.activity_name = activity_name
|
| 568 |
+
|
| 569 |
+
def __enter__(self):
|
| 570 |
+
self.start_time = time.time()
|
| 571 |
+
return self
|
| 572 |
|
| 573 |
+
def __exit__(self, exc_type, exc_value, traceback):
|
| 574 |
+
self.end_time = time.time()
|
| 575 |
+
self.elapsed_time = self.end_time - self.start_time
|
| 576 |
+
if self.activity_name:
|
| 577 |
+
print(f"Elapsed time for {self.activity_name}: {self.elapsed_time:.6f} seconds")
|
| 578 |
+
else:
|
| 579 |
+
print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
|
| 580 |
+
|
| 581 |
+
def update_selection(evt: gr.SelectData, width, height):
|
| 582 |
+
selected_lora = loras[evt.index]
|
| 583 |
+
new_placeholder = f"Type a prompt for {selected_lora['title']}"
|
| 584 |
+
lora_repo = selected_lora["repo"]
|
| 585 |
+
# 로컬 LoRA 처리
|
| 586 |
+
if lora_repo == "./":
|
| 587 |
+
updated_text = f"### Selected: Local LoRA - {selected_lora['title']} ✅"
|
| 588 |
+
else:
|
| 589 |
+
updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}) ✅"
|
| 590 |
+
if "aspect" in selected_lora:
|
| 591 |
+
if selected_lora["aspect"] == "portrait":
|
| 592 |
+
width = 768
|
| 593 |
+
height = 1024
|
| 594 |
+
elif selected_lora["aspect"] == "landscape":
|
| 595 |
+
width = 1024
|
| 596 |
+
height = 768
|
| 597 |
+
else:
|
| 598 |
+
width = 1024
|
| 599 |
+
height = 1024
|
| 600 |
+
return (
|
| 601 |
+
gr.update(placeholder=new_placeholder),
|
| 602 |
+
updated_text,
|
| 603 |
+
evt.index,
|
| 604 |
+
width,
|
| 605 |
+
height,
|
| 606 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 607 |
|
| 608 |
+
@spaces.GPU
|
| 609 |
+
def run_lora(prompt, image_input, image_strength, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale, progress=gr.Progress(track_tqdm=True)):
|
| 610 |
+
# Clean up previous LoRAs in both cases
|
| 611 |
+
with calculateDuration("Unloading LoRA"):
|
| 612 |
+
pipe.unload_lora_weights()
|
|
|
|
| 613 |
|
| 614 |
+
# Check if a LoRA is selected
|
| 615 |
+
if selected_index is not None and selected_index < len(loras):
|
| 616 |
+
selected_lora = loras[selected_index]
|
| 617 |
+
lora_path = selected_lora["repo"]
|
| 618 |
+
trigger_word = selected_lora["trigger_word"]
|
| 619 |
+
|
| 620 |
+
# Prepare Prompt with Trigger Word
|
| 621 |
+
if(trigger_word):
|
| 622 |
+
if "trigger_position" in selected_lora:
|
| 623 |
+
if selected_lora["trigger_position"] == "prepend":
|
| 624 |
+
prompt_mash = f"{trigger_word} {prompt}"
|
| 625 |
+
else:
|
| 626 |
+
prompt_mash = f"{prompt} {trigger_word}"
|
| 627 |
+
else:
|
| 628 |
+
prompt_mash = f"{trigger_word} {prompt}"
|
| 629 |
else:
|
| 630 |
+
prompt_mash = prompt
|
| 631 |
|
| 632 |
+
# Load LoRA
|
| 633 |
+
with calculateDuration(f"Loading LoRA weights for {selected_lora['title']}"):
|
| 634 |
+
weight_name = selected_lora.get("weights", None)
|
| 635 |
+
try:
|
| 636 |
+
pipe.load_lora_weights(
|
| 637 |
+
lora_path,
|
| 638 |
+
weight_name=weight_name,
|
| 639 |
+
adapter_name="default",
|
| 640 |
+
low_cpu_mem_usage=True
|
| 641 |
+
)
|
| 642 |
+
# Set adapter scale
|
| 643 |
+
pipe.set_adapters(["default"], adapter_weights=[lora_scale])
|
| 644 |
+
except Exception as e:
|
| 645 |
+
print(f"Error loading LoRA: {e}")
|
| 646 |
+
gr.Warning("Failed to load LoRA weights. Generating with base model.")
|
| 647 |
else:
|
| 648 |
+
# Base Model Case
|
| 649 |
+
print("No LoRA selected. Running with Base Model.")
|
| 650 |
+
prompt_mash = prompt
|
| 651 |
+
|
| 652 |
+
with calculateDuration("Randomizing seed"):
|
| 653 |
+
if randomize_seed:
|
| 654 |
+
seed = random.randint(0, MAX_SEED)
|
| 655 |
|
| 656 |
+
generator = torch.Generator(device=device).manual_seed(seed)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 657 |
|
| 658 |
+
# Note: Z-Image-Turbo is strictly T2I in this reference implementation.
|
| 659 |
+
# Img2Img via image_input is disabled/ignored for this pipeline update.
|
|
|
|
|
|
|
|
|
|
| 660 |
|
| 661 |
+
with calculateDuration("Generating image"):
|
| 662 |
+
# For Turbo models, guidance_scale is typically 0.0
|
| 663 |
+
forced_guidance = 0.0 # Turbo mode
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 664 |
|
| 665 |
+
final_image = pipe(
|
| 666 |
+
prompt=prompt_mash,
|
| 667 |
+
height=int(height),
|
| 668 |
+
width=int(width),
|
| 669 |
+
num_inference_steps=int(steps),
|
| 670 |
+
guidance_scale=forced_guidance,
|
| 671 |
+
generator=generator,
|
| 672 |
+
).images[0]
|
| 673 |
|
| 674 |
+
yield final_image, seed, gr.update(visible=False)
|
| 675 |
+
|
| 676 |
+
def get_huggingface_safetensors(link):
|
| 677 |
+
split_link = link.split("/")
|
| 678 |
+
if(len(split_link) == 2):
|
| 679 |
+
model_card = ModelCard.load(link)
|
| 680 |
+
base_model = model_card.data.get("base_model")
|
| 681 |
+
print(base_model)
|
| 682 |
+
|
| 683 |
+
# Relaxed check to allow Z-Image or Flux or others, assuming user knows what they are doing
|
| 684 |
+
# or specifically check for Z-Image-Turbo
|
| 685 |
+
if base_model not in ["Tongyi-MAI/Z-Image-Turbo", "black-forest-labs/FLUX.1-dev"]:
|
| 686 |
+
# Just a warning instead of error to allow experimentation
|
| 687 |
+
print("Warning: Base model might not match.")
|
| 688 |
+
|
| 689 |
+
image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None)
|
| 690 |
+
trigger_word = model_card.data.get("instance_prompt", "")
|
| 691 |
+
image_url = f"https://huggingface.co/{link}/resolve/main/{image_path}" if image_path else None
|
| 692 |
+
fs = HfFileSystem()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 693 |
try:
|
| 694 |
+
list_of_files = fs.ls(link, detail=False)
|
| 695 |
+
for file in list_of_files:
|
| 696 |
+
if(file.endswith(".safetensors")):
|
| 697 |
+
safetensors_name = file.split("/")[-1]
|
| 698 |
+
if (not image_url and file.lower().endswith((".jpg", ".jpeg", ".png", ".webp"))):
|
| 699 |
+
image_elements = file.split("/")
|
| 700 |
+
image_url = f"https://huggingface.co/{link}/resolve/main/{image_elements[-1]}"
|
| 701 |
except Exception as e:
|
| 702 |
+
print(e)
|
| 703 |
+
gr.Warning(f"You didn't include a link neither a valid Hugging Face repository with a *.safetensors LoRA")
|
| 704 |
+
raise Exception(f"You didn't include a link neither a valid Hugging Face repository with a *.safetensors LoRA")
|
| 705 |
+
return split_link[1], link, safetensors_name, trigger_word, image_url
|
| 706 |
+
|
| 707 |
+
def check_custom_model(link):
|
| 708 |
+
if(link.startswith("https://")):
|
| 709 |
+
if(link.startswith("https://huggingface.co") or link.startswith("https://www.huggingface.co")):
|
| 710 |
+
link_split = link.split("huggingface.co/")
|
| 711 |
+
return get_huggingface_safetensors(link_split[1])
|
| 712 |
+
else:
|
| 713 |
+
return get_huggingface_safetensors(link)
|
| 714 |
+
|
| 715 |
+
def add_custom_lora(custom_lora):
|
| 716 |
+
global loras
|
| 717 |
+
if(custom_lora):
|
| 718 |
+
try:
|
| 719 |
+
title, repo, path, trigger_word, image = check_custom_model(custom_lora)
|
| 720 |
+
print(f"Loaded custom LoRA: {repo}")
|
| 721 |
+
card = f'''
|
| 722 |
+
<div class="custom_lora_card">
|
| 723 |
+
<span>Loaded custom LoRA:</span>
|
| 724 |
+
<div class="card_internal">
|
| 725 |
+
<img src="{image}" />
|
| 726 |
+
<div>
|
| 727 |
+
<h3>{title}</h3>
|
| 728 |
+
<small>{"Using: <code><b>"+trigger_word+"</code></b> as the trigger word" if trigger_word else "No trigger word found. If there's a trigger word, include it in your prompt"}<br></small>
|
| 729 |
+
</div>
|
| 730 |
+
</div>
|
| 731 |
+
</div>
|
| 732 |
+
'''
|
| 733 |
+
existing_item_index = next((index for (index, item) in enumerate(loras) if item['repo'] == repo), None)
|
| 734 |
+
if(not existing_item_index):
|
| 735 |
+
new_item = {
|
| 736 |
+
"image": image,
|
| 737 |
+
"title": title,
|
| 738 |
+
"repo": repo,
|
| 739 |
+
"weights": path,
|
| 740 |
+
"trigger_word": trigger_word
|
| 741 |
+
}
|
| 742 |
+
print(new_item)
|
| 743 |
+
existing_item_index = len(loras)
|
| 744 |
+
loras.append(new_item)
|
| 745 |
+
|
| 746 |
+
return gr.update(visible=True, value=card), gr.update(visible=True), gr.Gallery(selected_index=None), f"Custom: {path}", existing_item_index, trigger_word
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 747 |
except Exception as e:
|
| 748 |
+
gr.Warning(f"Invalid LoRA: either you entered an invalid link, or a non-supported LoRA")
|
| 749 |
+
return gr.update(visible=True, value=f"Invalid LoRA: either you entered an invalid link, a non-supported LoRA"), gr.update(visible=False), gr.update(), "", None, ""
|
| 750 |
+
else:
|
| 751 |
+
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
| 752 |
|
| 753 |
+
def remove_custom_lora():
|
| 754 |
+
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
| 755 |
|
| 756 |
+
run_lora.zerogpu = True
|
| 757 |
+
|
| 758 |
+
with gr.Blocks(title="Z-IMAGE GEN/LORA", delete_cache=(60, 60)) as demo:
|
|
|
|
| 759 |
|
| 760 |
# HOME Button
|
| 761 |
gr.HTML("""
|
|
|
|
| 771 |
gr.HTML("""
|
| 772 |
<div class="header-container">
|
| 773 |
<div class="header-title">🎨 Z-IMAGE GEN/LORA 🎨</div>
|
| 774 |
+
<div class="header-subtitle">Generate amazing images with Z-Image Turbo and various LoRA styles!</div>
|
| 775 |
<div style="margin-top:12px">
|
| 776 |
+
<span class="stats-badge">⚡ Turbo Speed</span>
|
| 777 |
+
<span class="stats-badge">🎭 30+ LoRAs</span>
|
| 778 |
+
<span class="stats-badge">🖼️ High Quality</span>
|
| 779 |
+
<span class="stats-badge">🔧 Custom LoRA</span>
|
| 780 |
</div>
|
| 781 |
</div>
|
| 782 |
""")
|
|
|
|
|
|
|
|
|
|
| 783 |
|
| 784 |
+
selected_index = gr.State(None)
|
| 785 |
with gr.Row():
|
| 786 |
+
with gr.Column(scale=3):
|
| 787 |
+
prompt = gr.Textbox(label="Enter Prompt", lines=1, placeholder="✦︎ Choose the LoRA and type the prompt (LoRA = None → Base Model = Active)")
|
| 788 |
+
with gr.Column(scale=1, elem_id="gen_column"):
|
| 789 |
+
generate_button = gr.Button("🚀 Generate", variant="primary", elem_id="gen_btn")
|
|
|
|
|
|
|
|
|
|
| 790 |
with gr.Row():
|
| 791 |
+
with gr.Column():
|
| 792 |
+
selected_info = gr.Markdown("### No LoRA Selected (Base Model)")
|
| 793 |
+
gallery = gr.Gallery(
|
| 794 |
+
[(item["image"], item["title"]) for item in loras],
|
| 795 |
+
label="Z-Image LoRAs",
|
| 796 |
+
allow_preview=False,
|
| 797 |
+
columns=3,
|
| 798 |
+
elem_id="gallery",
|
|
|
|
| 799 |
)
|
| 800 |
+
with gr.Group():
|
| 801 |
+
custom_lora = gr.Textbox(label="Enter Custom LoRA", placeholder="Paste the LoRA path and press Enter (e.g., Shakker-Labs/AWPortrait-Z).")
|
| 802 |
+
gr.Markdown("[Check the list of Z-Image LoRA's](https://huggingface.co/models?other=base_model:adapter:Tongyi-MAI/Z-Image-Turbo)", elem_id="lora_list")
|
| 803 |
+
custom_lora_info = gr.HTML(visible=False)
|
| 804 |
+
custom_lora_button = gr.Button("Remove custom LoRA", visible=False)
|
| 805 |
+
with gr.Column():
|
| 806 |
+
progress_bar = gr.Markdown(elem_id="progress",visible=False)
|
| 807 |
+
result = gr.Image(label="Generated Image", format="png", height=630)
|
| 808 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 809 |
with gr.Row():
|
| 810 |
+
with gr.Accordion("Advanced Settings", open=False):
|
| 811 |
+
with gr.Row():
|
| 812 |
+
input_image = gr.Image(label="Input image (Ignored for Z-Image-Turbo)", type="filepath", visible=False)
|
| 813 |
+
image_strength = gr.Slider(label="Denoise Strength", info="Ignored for Z-Image-Turbo", minimum=0.1, maximum=1.0, step=0.01, value=0.75, visible=False)
|
| 814 |
+
with gr.Column():
|
| 815 |
+
with gr.Row():
|
| 816 |
+
cfg_scale = gr.Slider(label="CFG Scale", info="Forced to 0.0 for Turbo", minimum=0, maximum=20, step=0.5, value=0.0, interactive=False)
|
| 817 |
+
steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=25)
|
| 818 |
+
|
| 819 |
+
with gr.Row():
|
| 820 |
+
width = gr.Slider(label="Width", minimum=256, maximum=1536, step=64, value=1536)
|
| 821 |
+
height = gr.Slider(label="Height", minimum=256, maximum=1536, step=64, value=1536)
|
| 822 |
+
|
| 823 |
+
with gr.Row():
|
| 824 |
+
randomize_seed = gr.Checkbox(True, label="Randomize seed")
|
| 825 |
+
seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
|
| 826 |
+
lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=3, step=0.01, value=0.95)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 827 |
|
| 828 |
# Footer
|
| 829 |
gr.HTML("""
|
| 830 |
<div class="footer-comic">
|
| 831 |
<p style="font-family:'Bangers',cursive;font-size:1.5rem;letter-spacing:2px">🎨 Z-IMAGE GEN/LORA 🎨</p>
|
| 832 |
+
<p>Powered by Z-Image Turbo + LoRA Adapters</p>
|
| 833 |
+
<p>⚡ Fast Generation • 🎭 Multiple Styles • 🖼️ High Quality</p>
|
| 834 |
<p style="margin-top:10px"><a href="https://www.ginigen.com" target="_blank" style="color:#FACC15;text-decoration:none;font-weight:bold;">🏠 www.ginigen.com</a></p>
|
| 835 |
</div>
|
| 836 |
""")
|
| 837 |
|
| 838 |
+
gallery.select(
|
| 839 |
+
update_selection,
|
| 840 |
+
inputs=[width, height],
|
| 841 |
+
outputs=[prompt, selected_info, selected_index, width, height]
|
| 842 |
+
)
|
| 843 |
+
custom_lora.input(
|
| 844 |
+
add_custom_lora,
|
| 845 |
+
inputs=[custom_lora],
|
| 846 |
+
outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, prompt]
|
| 847 |
+
)
|
| 848 |
+
custom_lora_button.click(
|
| 849 |
+
remove_custom_lora,
|
| 850 |
+
outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, custom_lora]
|
| 851 |
+
)
|
| 852 |
+
gr.on(
|
| 853 |
+
triggers=[generate_button.click, prompt.submit],
|
| 854 |
+
fn=run_lora,
|
| 855 |
+
inputs=[prompt, input_image, image_strength, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale],
|
| 856 |
+
outputs=[result, seed, progress_bar]
|
| 857 |
+
)
|
| 858 |
+
|
| 859 |
+
demo.queue()
|
| 860 |
+
demo.launch(css=COMIC_CSS, ssr_mode=False, show_error=True)
|