Spaces:
Running on Zero
Running on Zero
File size: 34,152 Bytes
7970a56 e87beaf 2f3b407 7970a56 e87beaf 7970a56 2f3b407 b44b83f ac2d590 b44b83f ebaf191 b44b83f ac2d590 103b8ea b44b83f 103b8ea ac2d590 e87beaf 103b8ea e87beaf 103b8ea ac2d590 103b8ea ac2d590 e87beaf ac2d590 103b8ea ac2d590 e87beaf ac2d590 e87beaf b44b83f ac2d590 e87beaf ac2d590 b44b83f ac2d590 b44b83f ac2d590 b44b83f ac2d590 b44b83f ac2d590 b44b83f ac2d590 b44b83f ac2d590 b44b83f ac2d590 b44b83f ac2d590 7970a56 27b09f6 103b8ea e87beaf 103b8ea e87beaf 103b8ea e87beaf 103b8ea e87beaf 103b8ea 7970a56 e87beaf 7970a56 e87beaf 103b8ea 7970a56 e87beaf 7970a56 e87beaf 7970a56 e87beaf 7970a56 e87beaf 7970a56 103b8ea 7970a56 e87beaf 103b8ea e87beaf 103b8ea e87beaf 103b8ea 7970a56 e87beaf 7970a56 e87beaf 7970a56 e87beaf 7970a56 e87beaf 7970a56 e87beaf 7970a56 e87beaf 7970a56 e87beaf 7970a56 e87beaf 7970a56 de44604 e87beaf 27b09f6 7970a56 452f0d5 27b09f6 7970a56 1b93f28 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 | import os; os.system('pip install --upgrade --no-deps spaces')
import spaces
import shutil
import subprocess
import sys
import copy
import random
import tempfile
import warnings
import time
import gc
import uuid
from tqdm import tqdm
import cv2
import numpy as np
import torch
import torch._dynamo
from torch.nn import functional as F
from PIL import Image
import gradio as gr
from diffusers import (
FlowMatchEulerDiscreteScheduler,
SASolverScheduler,
DEISMultistepScheduler,
DPMSolverMultistepInverseScheduler,
UniPCMultistepScheduler,
DPMSolverMultistepScheduler,
DPMSolverSinglestepScheduler,
)
from diffusers.pipelines.wan.pipeline_wan_i2v import WanImageToVideoPipeline
from diffusers.utils.export_utils import export_to_video
from torchao.quantization import quantize_, Float8DynamicActivationFloat8WeightConfig, Int8WeightOnlyConfig
import aoti
import lora_loader
os.environ["TOKENIZERS_PARALLELISM"] = "true"
warnings.filterwarnings("ignore")
IS_ZERO_GPU = bool(os.getenv("SPACES_ZERO_GPU"))
# if IS_ZERO_GPU:
# print("Loading...")
# subprocess.run("rm -rf /data-nvme/zerogpu-offload/*", env={}, shell=True)
# --- FRAME EXTRACTION JS & LOGIC ---
# JS to grab timestamp from the output video
get_timestamp_js = """
function() {
// Select the video element specifically inside the component with id 'generated-video'
const video = document.querySelector('#generated-video video');
if (video) {
console.log("Video found! Time: " + video.currentTime);
return video.currentTime;
} else {
console.log("No video element found.");
return 0;
}
}
"""
def extract_frame(video_path, timestamp):
# Safety check: if no video is present
if not video_path:
return None
print(f"Extracting frame at timestamp: {timestamp}")
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
return None
# Calculate frame number
fps = cap.get(cv2.CAP_PROP_FPS)
target_frame_num = int(float(timestamp) * fps)
# Cap total frames to prevent errors at the very end of video
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
if target_frame_num >= total_frames:
target_frame_num = total_frames - 1
# Set position
cap.set(cv2.CAP_PROP_POS_FRAMES, target_frame_num)
ret, frame = cap.read()
cap.release()
if ret:
# Convert from BGR (OpenCV) to RGB (Gradio)
# Gradio Image component handles Numpy array -> PIL conversion automatically
return cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
return None
# --- END FRAME EXTRACTION LOGIC ---
def clear_vram():
gc.collect()
torch.cuda.empty_cache()
# RIFE
if not os.path.exists("RIFEv4.26_0921.zip"):
print("Downloading RIFE Model...")
subprocess.run([
"wget", "-q",
"https://huggingface.co/thornmaze/RIFE/resolve/main/RIFEv4.26_0921.zip",
"-O", "RIFEv4.26_0921.zip"
], check=True)
subprocess.run(["unzip", "-o", "RIFEv4.26_0921.zip"], check=True)
# sys.path.append(os.getcwd())
from train_log.RIFE_HDv3 import Model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
rife_model = Model()
rife_model.load_model("train_log", -1)
rife_model.eval()
@torch.no_grad()
def interpolate_bits(frames_np, multiplier=2, scale=1.0):
"""
Interpolation maintaining Numpy Float 0-1 format.
Args:
frames_np: Numpy Array (Time, Height, Width, Channels) - Float32 [0.0, 1.0]
multiplier: int (2, 4, 8)
Returns:
List of Numpy Arrays (Height, Width, Channels) - Float32 [0.0, 1.0]
"""
# Handle input shape
if isinstance(frames_np, list):
# Convert list of arrays to one big array for easier shape handling if needed,
# but here we just grab dims from first frame
T = len(frames_np)
H, W, C = frames_np[0].shape
else:
T, H, W, C = frames_np.shape
# 1. No Interpolation Case
if multiplier < 2:
# Just convert 4D array to list of 3D arrays
if isinstance(frames_np, np.ndarray):
return list(frames_np)
return frames_np
n_interp = multiplier - 1
# Pre-calc padding for RIFE (requires dimensions divisible by 32/scale)
tmp = max(128, int(128 / scale))
ph = ((H - 1) // tmp + 1) * tmp
pw = ((W - 1) // tmp + 1) * tmp
padding = (0, pw - W, 0, ph - H)
# Helper: Numpy (H, W, C) Float -> Tensor (1, C, H, W) Half
def to_tensor(frame_np):
# frame_np is float32 0-1
t = torch.from_numpy(frame_np).to(device)
# HWC -> CHW
t = t.permute(2, 0, 1).unsqueeze(0)
return F.pad(t, padding).half()
# Helper: Tensor (1, C, H, W) Half -> Numpy (H, W, C) Float
def from_tensor(tensor):
# Crop padding
t = tensor[0, :, :H, :W]
# CHW -> HWC
t = t.permute(1, 2, 0)
# Keep as float32, range 0-1
return t.float().cpu().numpy()
def make_inference(I0, I1, n):
if rife_model.version >= 3.9:
res = []
for i in range(n):
res.append(rife_model.inference(I0, I1, (i+1) * 1. / (n+1), scale))
return res
else:
middle = rife_model.inference(I0, I1, scale)
if n == 1:
return [middle]
first_half = make_inference(I0, middle, n=n//2)
second_half = make_inference(middle, I1, n=n//2)
if n % 2:
return [*first_half, middle, *second_half]
else:
return [*first_half, *second_half]
output_frames = []
# Process Frames
# Load first frame into GPU
I1 = to_tensor(frames_np[0])
total_steps = T - 1
with tqdm(total=total_steps, desc="Interpolating", unit="frame") as pbar:
for i in range(total_steps):
I0 = I1
# Add original frame to output
output_frames.append(from_tensor(I0))
# Load next frame
I1 = to_tensor(frames_np[i+1])
# Generate intermediate frames
mid_tensors = make_inference(I0, I1, n_interp)
# Append intermediate frames
for mid in mid_tensors:
output_frames.append(from_tensor(mid))
if (i + 1) % 50 == 0:
pbar.update(50)
pbar.update(total_steps % 50)
# Add the very last frame
output_frames.append(from_tensor(I1))
# Cleanup
del I0, I1, mid_tensors
torch.cuda.empty_cache()
return output_frames
# WAN
MODEL_ID = "thornmaze/WAMU_v3_WAN2.2_I2V_LIGHTNING"
LORA_MODELS = []
MAX_DIM = 832
MIN_DIM = 480
SQUARE_DIM = 640
MULTIPLE_OF = 16
MAX_SEED = np.iinfo(np.int32).max
FIXED_FPS = 16
MIN_FRAMES_MODEL = 8
MAX_FRAMES_MODEL = 321
MIN_DURATION = round(MIN_FRAMES_MODEL / FIXED_FPS, 1)
MAX_DURATION = round(MAX_FRAMES_MODEL / FIXED_FPS, 1)
SCHEDULER_MAP = {
"FlowMatchEulerDiscrete": FlowMatchEulerDiscreteScheduler,
"SASolver": SASolverScheduler,
"DEISMultistep": DEISMultistepScheduler,
"DPMSolverMultistepInverse": DPMSolverMultistepInverseScheduler,
"UniPCMultistep": UniPCMultistepScheduler,
"DPMSolverMultistep": DPMSolverMultistepScheduler,
"DPMSolverSinglestep": DPMSolverSinglestepScheduler,
}
pipe = WanImageToVideoPipeline.from_pretrained(
MODEL_ID,
torch_dtype=torch.bfloat16,
).to('cuda')
original_scheduler = copy.deepcopy(pipe.scheduler)
for i, lora in enumerate(LORA_MODELS):
name_high_tr = lora["high_tr"].split(".")[0].split("/")[-1] + "Hh"
name_low_tr = lora["low_tr"].split(".")[0].split("/")[-1] + "Ll"
try:
pipe.load_lora_weights(
lora["repo_id"],
weight_name=lora["high_tr"],
adapter_name=name_high_tr
)
kwargs_lora = {"load_into_transformer_2": True}
pipe.load_lora_weights(
lora["repo_id"],
weight_name=lora["low_tr"],
adapter_name=name_low_tr,
**kwargs_lora
)
pipe.set_adapters([name_high_tr, name_low_tr], adapter_weights=[1.0, 1.0])
pipe.fuse_lora(adapter_names=[name_high_tr], lora_scale=lora["high_scale"], components=["transformer"])
pipe.fuse_lora(adapter_names=[name_low_tr], lora_scale=lora["low_scale"], components=["transformer_2"])
pipe.unload_lora_weights()
print(f"Applied: {lora['high_tr']}, hs={lora['high_scale']}/ls={lora['low_scale']}, {i+1}/{len(LORA_MODELS)}")
except Exception as e:
print("Error:", str(e))
print("Failed LoRA:", name_high_tr)
pipe.unload_lora_weights()
# if os.path.exists(CACHE_DIR):
# shutil.rmtree(CACHE_DIR)
# print("Deleted Hugging Face cache.")
# else:
# print("No hub cache found.")
quantize_(pipe.text_encoder, Int8WeightOnlyConfig())
torch._dynamo.reset()
quantize_(pipe.transformer, Float8DynamicActivationFloat8WeightConfig())
torch._dynamo.reset()
quantize_(pipe.transformer_2, Float8DynamicActivationFloat8WeightConfig())
torch._dynamo.reset()
spaces.aoti_load(
module=pipe.transformer,
repo_id='thornmaze/WanTransformer3DModel-sm120-cu130-raa',
)
spaces.aoti_load(
module=pipe.transformer_2,
repo_id='thornmaze/WanTransformer3DModel-sm120-cu130-raa',
)
# pipe.vae.enable_slicing()
# pipe.vae.enable_tiling()
default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
default_negative_prompt = "่ฒ่ฐ่ณไธฝ, ่ฟๆ, ้ๆ, ็ป่ๆจก็ณไธๆธ
, ๅญๅน, ้ฃๆ ผ, ไฝๅ, ็ปไฝ, ็ป้ข, ้ๆญข, ๆดไฝๅ็ฐ, ๆๅทฎ่ดจ้, ไฝ่ดจ้, JPEGๅ็ผฉๆฎ็, ไธ้็, ๆฎ็ผบ็, ๅคไฝ็ๆๆ, ็ปๅพไธๅฅฝ็ๆ้จ, ็ปๅพไธๅฅฝ็่ธ้จ, ็ธๅฝข็, ๆฏๅฎน็, ๅฝขๆ็ธๅฝข็่ขไฝ, ๆๆ่ๅ, ้ๆญขไธๅจ็็ป้ข, ๆไนฑ็่ๆฏ, ไธๆก่
ฟ, ่ๆฏไบบๅพๅค, ๅ็่ตฐ"
def model_title():
return "## Wan 2.2 I2V 14B Lightning โ NSFW"
def resize_image(image: Image.Image) -> Image.Image:
width, height = image.size
if width == height:
return image.resize((SQUARE_DIM, SQUARE_DIM), Image.LANCZOS)
aspect_ratio = width / height
MAX_ASPECT_RATIO = MAX_DIM / MIN_DIM
MIN_ASPECT_RATIO = MIN_DIM / MAX_DIM
image_to_resize = image
if aspect_ratio > MAX_ASPECT_RATIO:
target_w, target_h = MAX_DIM, MIN_DIM
crop_width = int(round(height * MAX_ASPECT_RATIO))
left = (width - crop_width) // 2
image_to_resize = image.crop((left, 0, left + crop_width, height))
elif aspect_ratio < MIN_ASPECT_RATIO:
target_w, target_h = MIN_DIM, MAX_DIM
crop_height = int(round(width / MIN_ASPECT_RATIO))
top = (height - crop_height) // 2
image_to_resize = image.crop((0, top, width, top + crop_height))
else:
if width > height:
target_w = MAX_DIM
target_h = int(round(target_w / aspect_ratio))
else:
target_h = MAX_DIM
target_w = int(round(target_h * aspect_ratio))
final_w = round(target_w / MULTIPLE_OF) * MULTIPLE_OF
final_h = round(target_h / MULTIPLE_OF) * MULTIPLE_OF
final_w = max(MIN_DIM, min(MAX_DIM, final_w))
final_h = max(MIN_DIM, min(MAX_DIM, final_h))
return image_to_resize.resize((final_w, final_h), Image.LANCZOS)
def resize_and_crop_to_match(target_image, reference_image):
ref_width, ref_height = reference_image.size
target_width, target_height = target_image.size
scale = max(ref_width / target_width, ref_height / target_height)
new_width, new_height = int(target_width * scale), int(target_height * scale)
resized = target_image.resize((new_width, new_height), Image.Resampling.LANCZOS)
left, top = (new_width - ref_width) // 2, (new_height - ref_height) // 2
return resized.crop((left, top, left + ref_width, top + ref_height))
def get_num_frames(duration_seconds: float):
raw = int(round(duration_seconds * FIXED_FPS))
raw = max(MIN_FRAMES_MODEL, min(MAX_FRAMES_MODEL, raw))
return ((raw - 1) // 4) * 4 + 1
def get_inference_duration(
resized_image,
processed_last_image,
prompt,
steps,
negative_prompt,
num_frames,
guidance_scale,
guidance_scale_2,
current_seed,
scheduler_name,
flow_shift,
frame_multiplier,
quality,
duration_seconds,
safe_mode,
lora_groups,
progress
):
BASE_FRAMES_HEIGHT_WIDTH = 81 * 832 * 624
BASE_STEP_DURATION = 5.
width, height = resized_image.size
factor = num_frames * width * height / BASE_FRAMES_HEIGHT_WIDTH
step_duration = BASE_STEP_DURATION * factor ** 1.5
gen_time = int(steps) * step_duration
if guidance_scale > 1:
gen_time = gen_time * 2.4
frame_factor = frame_multiplier // FIXED_FPS
if frame_factor > 1:
total_out_frames = (num_frames * frame_factor) - num_frames
inter_time = (total_out_frames * 0.02)
gen_time += inter_time
total_time = 15 + gen_time
if safe_mode:
total_time = total_time * 1.30
return total_time
@spaces.GPU(duration=get_inference_duration, size='xlarge')
def run_inference(
resized_image,
processed_last_image,
prompt,
steps,
negative_prompt,
num_frames,
guidance_scale,
guidance_scale_2,
current_seed,
scheduler_name,
flow_shift,
frame_multiplier,
quality,
duration_seconds,
safe_mode=False,
lora_groups=None,
progress=gr.Progress(track_tqdm=True),
):
scheduler_class = SCHEDULER_MAP.get(scheduler_name)
if scheduler_class.__name__ != pipe.scheduler.config._class_name or flow_shift != pipe.scheduler.config.get("flow_shift", "shift"):
config = copy.deepcopy(original_scheduler.config)
if scheduler_class == FlowMatchEulerDiscreteScheduler:
config['shift'] = flow_shift
else:
config['flow_shift'] = flow_shift
pipe.scheduler = scheduler_class.from_config(config)
clear_vram()
task_name = str(uuid.uuid4())[:8]
print(f"Generating {num_frames} frames, task: {task_name}, {duration_seconds}, {resized_image.size}, lora={lora_groups}")
start = time.time()
lora_loaded = False
lora_adapter_names = []
if lora_groups:
try:
for idx, name in enumerate(lora_groups):
if name and name != "(None)":
lora_adapter_names += lora_loader.load_lora_to_pipe(
pipe, name, adapter_name=f"lora_{idx}"
)
# Activate ALL loaded adapters together (HIGH in transformer, LOW in
# transformer_2); each load_lora_weights alone would leave only the
# last adapter active.
if lora_adapter_names:
pipe.set_adapters(lora_adapter_names, adapter_weights=[1.0] * len(lora_adapter_names))
lora_loaded = True
print(f"LoRA loaded: {lora_groups}")
except Exception as e:
print(f"LoRA warning: {e}")
lora_loader.unload_lora(pipe)
# LoRA prompt + user prompt combined โ lora triggers first for activation
try:
additions = lora_loader.get_lora_prompt_additions(lora_groups)
except Exception:
additions = ""
if additions:
prompt = f"{additions} {prompt}"
print(f"Combined prompt: {prompt}")
result = pipe(
image=resized_image,
last_image=processed_last_image,
prompt=prompt,
negative_prompt=negative_prompt,
height=resized_image.height,
width=resized_image.width,
num_frames=num_frames,
guidance_scale=float(guidance_scale),
guidance_scale_2=float(guidance_scale_2),
num_inference_steps=int(steps),
generator=torch.Generator(device="cuda").manual_seed(current_seed),
output_type="np"
)
if lora_loaded:
lora_loader.unload_lora(pipe)
print("gen time passed:", time.time() - start)
raw_frames_np = result.frames[0] # Returns (T, H, W, C) float32
pipe.scheduler = original_scheduler
frame_factor = frame_multiplier // FIXED_FPS
if frame_factor > 1:
start = time.time()
print(f"Processing frames (RIFE Multiplier: {frame_factor}x)...")
rife_model.device()
rife_model.flownet = rife_model.flownet.half()
final_frames = interpolate_bits(raw_frames_np, multiplier=int(frame_factor))
print("Interpolation time passed:", time.time() - start)
else:
final_frames = list(raw_frames_np)
final_fps = FIXED_FPS * int(frame_factor)
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
video_path = tmpfile.name
start = time.time()
with tqdm(total=3, desc="Rendering Media", unit="clip") as pbar:
pbar.update(2)
export_to_video(final_frames, video_path, fps=final_fps, quality=quality)
pbar.update(1)
print(f"Export time passed, {final_fps} FPS:", time.time() - start)
return video_path, task_name
def generate_video(
input_image,
last_image,
prompt,
steps=4,
negative_prompt=default_negative_prompt,
duration_seconds=MAX_DURATION,
guidance_scale=1,
guidance_scale_2=1,
seed=42,
randomize_seed=False,
quality=5,
scheduler="UniPCMultistep",
flow_shift=6.0,
frame_multiplier=16,
safe_mode=False,
lora_groups=None,
auto_lora_enabled=True,
video_component=True,
progress=gr.Progress(track_tqdm=True),
):
"""
Generate a video from an input image using the Wan 2.2 14B I2V model with Lightning LoRA.
This function takes an input image and generates a video animation based on the provided
prompt and parameters. It uses an FP8 qunatized Wan 2.2 14B Image-to-Video model in with Lightning LoRA
for fast generation in 4-8 steps.
Args:
input_image (PIL.Image): The input image to animate. Will be resized to target dimensions.
last_image (PIL.Image, optional): The optional last image for the video.
prompt (str): Text prompt describing the desired animation or motion.
steps (int, optional): Number of inference steps. More steps = higher quality but slower.
Defaults to 4. Range: 1-30.
negative_prompt (str, optional): Negative prompt to avoid unwanted elements.
Defaults to default_negative_prompt (contains unwanted visual artifacts).
duration_seconds (float, optional): Duration of the generated video in seconds.
Defaults to 2. Clamped between MIN_FRAMES_MODEL/FIXED_FPS and MAX_FRAMES_MODEL/FIXED_FPS.
guidance_scale (float, optional): Controls adherence to the prompt. Higher values = more adherence.
Defaults to 1.0. Range: 0.0-20.0.
guidance_scale_2 (float, optional): Controls adherence to the prompt. Higher values = more adherence.
Defaults to 1.0. Range: 0.0-20.0.
seed (int, optional): Random seed for reproducible results. Defaults to 42.
Range: 0 to MAX_SEED (2147483647).
randomize_seed (bool, optional): Whether to use a random seed instead of the provided seed.
Defaults to False.
quality (float, optional): Video output quality. Default is 5. Uses variable bit rate.
Highest quality is 10, lowest is 1.
scheduler (str, optional): The name of the scheduler to use for inference. Defaults to "UniPCMultistep".
flow_shift (float, optional): The flow shift value for compatible schedulers. Defaults to 6.0.
frame_multiplier (int, optional): The int value for fps enhancer
video_component(bool, optional): Show video player in output.
Defaults to True.
progress (gr.Progress, optional): Gradio progress tracker. Defaults to gr.Progress(track_tqdm=True).
Returns:
tuple: A tuple containing:
- video_path (str): Path for the video component.
- video_path (str): Path for the file download component. Attempt to avoid reconversion in video component.
- current_seed (int): The seed used for generation.
Raises:
gr.Error: If input_image is None (no image uploaded).
Note:
- Frame count is calculated as duration_seconds * FIXED_FPS (24)
- Output dimensions are adjusted to be multiples of MOD_VALUE (32)
- The function uses GPU acceleration via the @spaces.GPU decorator
- Generation time varies based on steps and duration (see get_duration function)
"""
if input_image is None:
raise gr.Error("Please upload an input image.")
num_frames = get_num_frames(duration_seconds)
current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
resized_image = resize_image(input_image)
processed_last_image = None
if last_image:
processed_last_image = resize_and_crop_to_match(last_image, resized_image)
# Auto-activate LoRAs from the prompt when enabled, merging with manual picks
if auto_lora_enabled and prompt:
try:
auto_loras = lora_loader.auto_activate_loras(prompt)
except Exception as e:
auto_loras = []
print(f"Auto-LoRA detection error: {e}")
if auto_loras:
print(f"Auto-activated LoRAs: {auto_loras}")
manual = list(lora_groups or [])
merged = list(auto_loras)
for m in manual:
if m not in merged:
merged.append(m)
lora_groups = merged
video_path, task_n = run_inference(
resized_image,
processed_last_image,
prompt,
steps,
negative_prompt,
num_frames,
guidance_scale,
guidance_scale_2,
current_seed,
scheduler,
flow_shift,
frame_multiplier,
quality,
duration_seconds,
safe_mode,
lora_groups,
progress,
)
print(f"GPU complete: {task_n}")
return (video_path if video_component else None), video_path, current_seed
def _update_auto_lora_preview(prompt, auto_lora_enabled):
if not auto_lora_enabled or not prompt:
return ""
try:
details = lora_loader.auto_activation_details(prompt)
except Exception as e:
return f"<span style='color:#888'>auto-detection error: {e}</span>"
if not details:
return "<span style='color:#888'>No LoRAs matched.</span>"
items = []
for d in details:
phrases = ", ".join(d.get("matched", [])) or "โ"
items.append(f"<li><b>{d['group']}</b> <span style='color:#888'>(score {d['score']})</span><br><small>{phrases}</small></li>")
return "<b>Auto-activated:</b><ul style='margin:2px 0;padding-left:18px'>" + "".join(items) + "</ul>"
GLASS_CSS = """
/* โโ Hidden timestamp (preserved) โโ */
#hidden-timestamp {
opacity: 0; height: 0px; width: 0px; margin: 0px; padding: 0px;
overflow: hidden; position: absolute; pointer-events: none;
}
/* โโ BACKGROUND + AMBIENT GLOW โ radial-gradient on body โโ */
body {
background: #0a0e1a !important;
background-image:
radial-gradient(ellipse 700px 700px at -5% 30%, rgba(0, 201, 167, 0.38), transparent 70%),
radial-gradient(ellipse 800px 800px at 98% 10%, rgba(168, 85, 247, 0.32), transparent 70%),
radial-gradient(ellipse 500px 500px at 50% 80%, rgba(0, 132, 227, 0.12), transparent 70%) !important;
background-attachment: fixed !important;
}
.gradio-container {
background: transparent !important;
min-height: 100vh !important;
}
/* โโ TITLE / SUBTITLE โโ */
.glass-title {
font-size: 28px !important;
font-weight: 700 !important;
color: #ffffff !important;
margin-bottom: 6px !important;
line-height: 1.2 !important;
}
.glass-subtitle {
font-size: 14px !important;
color: rgba(255, 255, 255, 0.45) !important;
margin-bottom: 24px !important;
line-height: 1.5 !important;
}
/* โโ SECTION LABELS (gr.HTML injected) โโ */
.glass-label {
font-size: 11px !important;
font-weight: 700 !important;
text-transform: uppercase !important;
letter-spacing: 1.5px !important;
color: #5eead4 !important;
margin-bottom: 8px !important;
display: block !important;
}
/* โโ LORA PREVIEW โโ */
#auto-lora-preview { color: rgba(255,255,255,0.5) !important; }
#auto-lora-preview b { color: #5eead4 !important; }
/* โโ STATUS TEXT โโ */
.glass-status { color: rgba(255,255,255,0.40) !important; font-size: 13px !important; text-align: center !important; margin-top: 12px !important; }
.glass-status strong { color: rgba(255,255,255,0.70) !important; }
/* โโ INPUT PANEL glass โโ */
#input-panel {
background: rgba(15, 20, 35, 0.55) !important;
border: 1px solid rgba(255, 255, 255, 0.10) !important;
border-radius: 16px !important;
box-shadow: 0 8px 32px rgba(0, 0, 0, 0.35) !important;
padding: 28px !important;
}
/* โโ HIDE footer โโ */
.gradio-container footer { display: none !important; }
"""
GLASS_JS = """
<script>
(function(){
function glassApply(){
var teal='rgba(94,234,212,0.85)';
var tealDark='rgba(94,234,212,0.95)';
/* Gradio 6 labels: <span data-testid="block-info"> (Slider, Dropdown, Textbox) */
document.querySelectorAll('[data-testid="block-info"]').forEach(function(el){
el.style.cssText+='text-transform:uppercase!important;font-size:11px!important;font-weight:700!important;letter-spacing:1px!important;color:'+teal+'!important';
});
/* Gradio 6 accordion: <button class="label-wrap"> with <span> child */
document.querySelectorAll('button.label-wrap').forEach(function(el){
el.style.cssText+='text-transform:uppercase!important;color:'+teal+'!important;font-weight:600!important;letter-spacing:1px!important;font-size:12px!important;background:rgba(255,255,255,0.04)!important;border:1px solid rgba(255,255,255,0.08)!important;border-radius:12px!important;padding:12px 16px!important';
});
/* Gradio 6 checkbox labels: <label data-testid="block-label"> */
document.querySelectorAll('[data-testid="block-label"]').forEach(function(el){
el.style.cssText+='text-transform:uppercase!important;font-size:11px!important;font-weight:700!important;letter-spacing:1px!important;color:'+tealDark+'!important';
});
/* Generate button โ gradient override */
document.querySelectorAll('.gradio-container button').forEach(function(el){
if(el.textContent.trim()==='Generate Video' || el.classList.contains('primary') || el.getAttribute('data-variant')==='primary'){
el.style.cssText+='background:linear-gradient(135deg,#00b894,#0984e3,#6c5ce7)!important;color:#fff!important;font-weight:700!important;font-size:16px!important;border:none!important;border-radius:12px!important;padding:14px 32px!important;width:100%!important;box-shadow:0 4px 20px rgba(0,184,148,0.35),0 0 40px rgba(108,92,231,0.15)!important;cursor:pointer!important';
}
});
/* Glass panels โ input panel */
document.querySelectorAll('#input-panel').forEach(function(el){
el.style.cssText+='background:rgba(15,20,35,0.55)!important;border:1px solid rgba(255,255,255,0.10)!important;border-radius:16px!important;box-shadow:0 8px 32px rgba(0,0,0,0.35)!important;padding:28px!important';
});
}
if(document.readyState==='loading'){
document.addEventListener('DOMContentLoaded',function(){setTimeout(glassApply,500);});
} else {
setTimeout(glassApply,500);
}
setInterval(glassApply,2000);
})();
</script>
"""
with gr.Blocks(delete_cache=(3600, 10800)) as demo:
# โโ TITLE & SUBTITLE โ float freely โโ
gr.HTML("<div class='glass-title'>Wan 2.2 I2V 14B Lightning โ NSFW</div>")
gr.HTML("<div class='glass-subtitle'>Run Wan 2.2 in just 4-8 steps, fp8 quantization & AoT compilation โ compatible with diffusers and ZeroGPU</div>")
with gr.Row(equal_height=False):
# โโโโโโโโโโโโโโโ LEFT COLUMN โ GLASS INPUT PANEL โโโโโโโโโโโโโโโ
with gr.Column(scale=4):
gr.HTML("<div id='input-panel'><div class='glass-label'>Prompt</div>")
prompt_input = gr.Textbox(
label="", value=default_prompt_i2v, lines=4,
show_label=False, placeholder="Describe the motion you want...",
)
auto_lora_preview = gr.HTML(
value="", elem_id="auto-lora-preview",
)
gr.HTML("<div class='glass-label'>Input Image</div>")
input_image_component = gr.Image(
type="pil", label="",
sources=["upload", "clipboard"], show_label=False,
)
with gr.Row():
auto_lora_checkbox = gr.Checkbox(
label="Auto-activate LoRAs",
value=True, interactive=True,
)
safe_mode_checkbox = gr.Checkbox(
label="Safe Mode",
value=True, interactive=True,
)
duration_seconds_input = gr.Slider(
minimum=MIN_DURATION, maximum=MAX_DURATION,
step=0.1, value=3.5, label="Duration",
)
frame_multi = gr.Dropdown(
choices=[FIXED_FPS, FIXED_FPS*2, FIXED_FPS*4, FIXED_FPS*8],
value=FIXED_FPS, label="Fluidity (FPS)",
)
# Generate button โ above accordion, inside glass panel
generate_button = gr.Button("Generate Video", variant="primary")
with gr.Accordion("Advanced Settings", open=False):
last_image_component = gr.Image(
type="pil", label="Last Image (Optional)",
sources=["upload", "clipboard"],
)
negative_prompt_input = gr.Textbox(
label="Negative Prompt", value=default_negative_prompt, lines=2,
)
with gr.Row():
quality_slider = gr.Slider(minimum=1, maximum=10, step=1, value=6, label="Quality")
seed_input = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42, interactive=True)
randomize_seed_checkbox = gr.Checkbox(label="Randomize seed", value=True, interactive=True)
with gr.Row():
steps_slider = gr.Slider(minimum=1, maximum=30, step=1, value=6, label="Steps")
flow_shift_slider = gr.Slider(minimum=0.5, maximum=15.0, step=0.1, value=3.0, label="Flow Shift")
guidance_scale_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=1, label="Guidance Scale โ high noise")
guidance_scale_2_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=1, label="Guidance Scale 2 โ low noise")
scheduler_dropdown = gr.Dropdown(
label="Scheduler", choices=list(SCHEDULER_MAP.keys()),
value="UniPCMultistep",
)
lora_dropdown = gr.Dropdown(
choices=lora_loader.get_lora_choices(),
label="LoRA (NSFW)", multiselect=True,
)
play_result_video = gr.Checkbox(label="Display result", value=True, interactive=True)
gr.HTML("</div>") # close #input-panel
# โโโโโโโโโโโโโโโ RIGHT COLUMN โ PREVIEW โโโโโโโโโโโโโโโ
with gr.Column(scale=6):
gr.HTML("<div class='glass-label'>Preview</div>")
video_output = gr.Video(
label="", autoplay=True, sources=["upload"],
buttons=["download", "share"], interactive=True,
elem_id="generated-video",
)
gr.HTML("<div class='glass-status'><strong>Ready to create</strong><br>Upload an image and click Generate</div>")
with gr.Row():
grab_frame_btn = gr.Button("Use Current Frame as Input", variant="secondary")
timestamp_box = gr.Number(value=0, label="Timestamp", visible=True, elem_id="hidden-timestamp")
file_output = gr.File(label="Download")
# โโโโโโโโโโโโโโโ EVENT WIRING (identical to app.py) โโโโโโโโโโโโโโโ
ui_inputs = [
input_image_component, last_image_component, prompt_input, steps_slider,
negative_prompt_input, duration_seconds_input,
guidance_scale_input, guidance_scale_2_input, seed_input, randomize_seed_checkbox,
quality_slider, scheduler_dropdown, flow_shift_slider, frame_multi,
safe_mode_checkbox,
lora_dropdown,
auto_lora_checkbox,
play_result_video
]
generate_button.click(
fn=generate_video,
inputs=ui_inputs,
outputs=[video_output, file_output, seed_input]
)
prompt_input.change(
fn=_update_auto_lora_preview,
inputs=[prompt_input, auto_lora_checkbox],
outputs=[auto_lora_preview],
)
auto_lora_checkbox.change(
fn=_update_auto_lora_preview,
inputs=[prompt_input, auto_lora_checkbox],
outputs=[auto_lora_preview],
)
grab_frame_btn.click(
fn=None, inputs=None, outputs=[timestamp_box],
js=get_timestamp_js
)
timestamp_box.change(
fn=extract_frame,
inputs=[video_output, timestamp_box],
outputs=[input_image_component]
)
if __name__ == "__main__":
theme = gr.themes.Default().set(
slider_color="#5eead4",
slider_color_dark="#5eead4",
checkbox_background_color_selected="#5eead4",
checkbox_background_color_selected_dark="#5eead4",
checkbox_border_color_selected="#5eead4",
checkbox_border_color_focus="#5eead4",
)
demo.queue().launch(
mcp_server=True,
head=f"<style>{GLASS_CSS}</style>{GLASS_JS}",
theme=theme,
show_error=True,
) |