Spaces:
Running on Zero
Running on Zero
UI: blue→green→purple gradient theme with frosted glass panels
Browse files
app.py
CHANGED
|
@@ -1,800 +1,1120 @@
|
|
| 1 |
-
import os; os.system('pip install --upgrade --no-deps spaces')
|
| 2 |
-
import spaces
|
| 3 |
-
import shutil
|
| 4 |
-
import subprocess
|
| 5 |
-
import sys
|
| 6 |
-
import copy
|
| 7 |
-
import random
|
| 8 |
-
import tempfile
|
| 9 |
-
import warnings
|
| 10 |
-
import time
|
| 11 |
-
import gc
|
| 12 |
-
import uuid
|
| 13 |
-
from tqdm import tqdm
|
| 14 |
-
import cv2
|
| 15 |
-
import numpy as np
|
| 16 |
-
import torch
|
| 17 |
-
import torch._dynamo
|
| 18 |
-
from torch.nn import functional as F
|
| 19 |
-
from PIL import Image
|
| 20 |
-
|
| 21 |
-
import gradio as gr
|
| 22 |
-
from diffusers import (
|
| 23 |
-
FlowMatchEulerDiscreteScheduler,
|
| 24 |
-
SASolverScheduler,
|
| 25 |
-
DEISMultistepScheduler,
|
| 26 |
-
DPMSolverMultistepInverseScheduler,
|
| 27 |
-
UniPCMultistepScheduler,
|
| 28 |
-
DPMSolverMultistepScheduler,
|
| 29 |
-
DPMSolverSinglestepScheduler,
|
| 30 |
-
)
|
| 31 |
-
from diffusers.pipelines.wan.pipeline_wan_i2v import WanImageToVideoPipeline
|
| 32 |
-
from diffusers.utils.export_utils import export_to_video
|
| 33 |
-
|
| 34 |
-
from torchao.quantization import quantize_, Float8DynamicActivationFloat8WeightConfig, Int8WeightOnlyConfig
|
| 35 |
-
import aoti
|
| 36 |
-
import lora_loader
|
| 37 |
-
|
| 38 |
-
os.environ["TOKENIZERS_PARALLELISM"] = "true"
|
| 39 |
-
warnings.filterwarnings("ignore")
|
| 40 |
-
IS_ZERO_GPU = bool(os.getenv("SPACES_ZERO_GPU"))
|
| 41 |
-
|
| 42 |
-
# if IS_ZERO_GPU:
|
| 43 |
-
# print("Loading...")
|
| 44 |
-
# subprocess.run("rm -rf /data-nvme/zerogpu-offload/*", env={}, shell=True)
|
| 45 |
-
|
| 46 |
-
# --- FRAME EXTRACTION JS & LOGIC ---
|
| 47 |
-
|
| 48 |
-
# JS to grab timestamp from the output video
|
| 49 |
-
get_timestamp_js = """
|
| 50 |
-
function() {
|
| 51 |
-
// Select the video element specifically inside the component with id 'generated-video'
|
| 52 |
-
const video = document.querySelector('#generated-video video');
|
| 53 |
-
|
| 54 |
-
if (video) {
|
| 55 |
-
console.log("Video found! Time: " + video.currentTime);
|
| 56 |
-
return video.currentTime;
|
| 57 |
-
} else {
|
| 58 |
-
console.log("No video element found.");
|
| 59 |
-
return 0;
|
| 60 |
-
}
|
| 61 |
-
}
|
| 62 |
-
"""
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
def extract_frame(video_path, timestamp):
|
| 66 |
-
# Safety check: if no video is present
|
| 67 |
-
if not video_path:
|
| 68 |
-
return None
|
| 69 |
-
|
| 70 |
-
print(f"Extracting frame at timestamp: {timestamp}")
|
| 71 |
-
|
| 72 |
-
cap = cv2.VideoCapture(video_path)
|
| 73 |
-
|
| 74 |
-
if not cap.isOpened():
|
| 75 |
-
return None
|
| 76 |
-
|
| 77 |
-
# Calculate frame number
|
| 78 |
-
fps = cap.get(cv2.CAP_PROP_FPS)
|
| 79 |
-
target_frame_num = int(float(timestamp) * fps)
|
| 80 |
-
|
| 81 |
-
# Cap total frames to prevent errors at the very end of video
|
| 82 |
-
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
| 83 |
-
if target_frame_num >= total_frames:
|
| 84 |
-
target_frame_num = total_frames - 1
|
| 85 |
-
|
| 86 |
-
# Set position
|
| 87 |
-
cap.set(cv2.CAP_PROP_POS_FRAMES, target_frame_num)
|
| 88 |
-
ret, frame = cap.read()
|
| 89 |
-
cap.release()
|
| 90 |
-
|
| 91 |
-
if ret:
|
| 92 |
-
# Convert from BGR (OpenCV) to RGB (Gradio)
|
| 93 |
-
# Gradio Image component handles Numpy array -> PIL conversion automatically
|
| 94 |
-
return cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
| 95 |
-
|
| 96 |
-
return None
|
| 97 |
-
|
| 98 |
-
# --- END FRAME EXTRACTION LOGIC ---
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
def clear_vram():
|
| 102 |
-
gc.collect()
|
| 103 |
-
torch.cuda.empty_cache()
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
# RIFE
|
| 107 |
-
if not os.path.exists("RIFEv4.26_0921.zip"):
|
| 108 |
-
print("Downloading RIFE Model...")
|
| 109 |
-
subprocess.run([
|
| 110 |
-
"wget", "-q",
|
| 111 |
-
"https://huggingface.co/thornmaze/RIFE/resolve/main/RIFEv4.26_0921.zip",
|
| 112 |
-
"-O", "RIFEv4.26_0921.zip"
|
| 113 |
-
], check=True)
|
| 114 |
-
subprocess.run(["unzip", "-o", "RIFEv4.26_0921.zip"], check=True)
|
| 115 |
-
|
| 116 |
-
# sys.path.append(os.getcwd())
|
| 117 |
-
|
| 118 |
-
from train_log.RIFE_HDv3 import Model
|
| 119 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 120 |
-
rife_model = Model()
|
| 121 |
-
rife_model.load_model("train_log", -1)
|
| 122 |
-
rife_model.eval()
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
@torch.no_grad()
|
| 126 |
-
def interpolate_bits(frames_np, multiplier=2, scale=1.0):
|
| 127 |
-
"""
|
| 128 |
-
Interpolation maintaining Numpy Float 0-1 format.
|
| 129 |
-
Args:
|
| 130 |
-
frames_np: Numpy Array (Time, Height, Width, Channels) - Float32 [0.0, 1.0]
|
| 131 |
-
multiplier: int (2, 4, 8)
|
| 132 |
-
Returns:
|
| 133 |
-
List of Numpy Arrays (Height, Width, Channels) - Float32 [0.0, 1.0]
|
| 134 |
-
"""
|
| 135 |
-
|
| 136 |
-
# Handle input shape
|
| 137 |
-
if isinstance(frames_np, list):
|
| 138 |
-
# Convert list of arrays to one big array for easier shape handling if needed,
|
| 139 |
-
# but here we just grab dims from first frame
|
| 140 |
-
T = len(frames_np)
|
| 141 |
-
H, W, C = frames_np[0].shape
|
| 142 |
-
else:
|
| 143 |
-
T, H, W, C = frames_np.shape
|
| 144 |
-
|
| 145 |
-
# 1. No Interpolation Case
|
| 146 |
-
if multiplier < 2:
|
| 147 |
-
# Just convert 4D array to list of 3D arrays
|
| 148 |
-
if isinstance(frames_np, np.ndarray):
|
| 149 |
-
return list(frames_np)
|
| 150 |
-
return frames_np
|
| 151 |
-
|
| 152 |
-
n_interp = multiplier - 1
|
| 153 |
-
|
| 154 |
-
# Pre-calc padding for RIFE (requires dimensions divisible by 32/scale)
|
| 155 |
-
tmp = max(128, int(128 / scale))
|
| 156 |
-
ph = ((H - 1) // tmp + 1) * tmp
|
| 157 |
-
pw = ((W - 1) // tmp + 1) * tmp
|
| 158 |
-
padding = (0, pw - W, 0, ph - H)
|
| 159 |
-
|
| 160 |
-
# Helper: Numpy (H, W, C) Float -> Tensor (1, C, H, W) Half
|
| 161 |
-
def to_tensor(frame_np):
|
| 162 |
-
# frame_np is float32 0-1
|
| 163 |
-
t = torch.from_numpy(frame_np).to(device)
|
| 164 |
-
# HWC -> CHW
|
| 165 |
-
t = t.permute(2, 0, 1).unsqueeze(0)
|
| 166 |
-
return F.pad(t, padding).half()
|
| 167 |
-
|
| 168 |
-
# Helper: Tensor (1, C, H, W) Half -> Numpy (H, W, C) Float
|
| 169 |
-
def from_tensor(tensor):
|
| 170 |
-
# Crop padding
|
| 171 |
-
t = tensor[0, :, :H, :W]
|
| 172 |
-
# CHW -> HWC
|
| 173 |
-
t = t.permute(1, 2, 0)
|
| 174 |
-
# Keep as float32, range 0-1
|
| 175 |
-
return t.float().cpu().numpy()
|
| 176 |
-
|
| 177 |
-
def make_inference(I0, I1, n):
|
| 178 |
-
if rife_model.version >= 3.9:
|
| 179 |
-
res = []
|
| 180 |
-
for i in range(n):
|
| 181 |
-
res.append(rife_model.inference(I0, I1, (i+1) * 1. / (n+1), scale))
|
| 182 |
-
return res
|
| 183 |
-
else:
|
| 184 |
-
middle = rife_model.inference(I0, I1, scale)
|
| 185 |
-
if n == 1:
|
| 186 |
-
return [middle]
|
| 187 |
-
first_half = make_inference(I0, middle, n=n//2)
|
| 188 |
-
second_half = make_inference(middle, I1, n=n//2)
|
| 189 |
-
if n % 2:
|
| 190 |
-
return [*first_half, middle, *second_half]
|
| 191 |
-
else:
|
| 192 |
-
return [*first_half, *second_half]
|
| 193 |
-
|
| 194 |
-
output_frames = []
|
| 195 |
-
|
| 196 |
-
# Process Frames
|
| 197 |
-
# Load first frame into GPU
|
| 198 |
-
I1 = to_tensor(frames_np[0])
|
| 199 |
-
|
| 200 |
-
total_steps = T - 1
|
| 201 |
-
|
| 202 |
-
with tqdm(total=total_steps, desc="Interpolating", unit="frame") as pbar:
|
| 203 |
-
|
| 204 |
-
for i in range(total_steps):
|
| 205 |
-
I0 = I1
|
| 206 |
-
# Add original frame to output
|
| 207 |
-
output_frames.append(from_tensor(I0))
|
| 208 |
-
|
| 209 |
-
# Load next frame
|
| 210 |
-
I1 = to_tensor(frames_np[i+1])
|
| 211 |
-
|
| 212 |
-
# Generate intermediate frames
|
| 213 |
-
mid_tensors = make_inference(I0, I1, n_interp)
|
| 214 |
-
|
| 215 |
-
# Append intermediate frames
|
| 216 |
-
for mid in mid_tensors:
|
| 217 |
-
output_frames.append(from_tensor(mid))
|
| 218 |
-
|
| 219 |
-
if (i + 1) % 50 == 0:
|
| 220 |
-
pbar.update(50)
|
| 221 |
-
pbar.update(total_steps % 50)
|
| 222 |
-
|
| 223 |
-
# Add the very last frame
|
| 224 |
-
output_frames.append(from_tensor(I1))
|
| 225 |
-
|
| 226 |
-
# Cleanup
|
| 227 |
-
del I0, I1, mid_tensors
|
| 228 |
-
torch.cuda.empty_cache()
|
| 229 |
-
|
| 230 |
-
return output_frames
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
# WAN
|
| 234 |
-
|
| 235 |
-
MODEL_ID = "thornmaze/WAMU_v3_WAN2.2_I2V_LIGHTNING"
|
| 236 |
-
|
| 237 |
-
LORA_MODELS = []
|
| 238 |
-
|
| 239 |
-
MAX_DIM = 832
|
| 240 |
-
MIN_DIM = 480
|
| 241 |
-
SQUARE_DIM = 640
|
| 242 |
-
MULTIPLE_OF = 16
|
| 243 |
-
MAX_SEED = np.iinfo(np.int32).max
|
| 244 |
-
|
| 245 |
-
FIXED_FPS = 16
|
| 246 |
-
MIN_FRAMES_MODEL = 8
|
| 247 |
-
MAX_FRAMES_MODEL = 321
|
| 248 |
-
|
| 249 |
-
MIN_DURATION = round(MIN_FRAMES_MODEL / FIXED_FPS, 1)
|
| 250 |
-
MAX_DURATION = round(MAX_FRAMES_MODEL / FIXED_FPS, 1)
|
| 251 |
-
|
| 252 |
-
SCHEDULER_MAP = {
|
| 253 |
-
"FlowMatchEulerDiscrete": FlowMatchEulerDiscreteScheduler,
|
| 254 |
-
"SASolver": SASolverScheduler,
|
| 255 |
-
"DEISMultistep": DEISMultistepScheduler,
|
| 256 |
-
"DPMSolverMultistepInverse": DPMSolverMultistepInverseScheduler,
|
| 257 |
-
"UniPCMultistep": UniPCMultistepScheduler,
|
| 258 |
-
"DPMSolverMultistep": DPMSolverMultistepScheduler,
|
| 259 |
-
"DPMSolverSinglestep": DPMSolverSinglestepScheduler,
|
| 260 |
-
}
|
| 261 |
-
|
| 262 |
-
pipe = WanImageToVideoPipeline.from_pretrained(
|
| 263 |
-
MODEL_ID,
|
| 264 |
-
torch_dtype=torch.bfloat16,
|
| 265 |
-
).to('cuda')
|
| 266 |
-
original_scheduler = copy.deepcopy(pipe.scheduler)
|
| 267 |
-
|
| 268 |
-
for i, lora in enumerate(LORA_MODELS):
|
| 269 |
-
name_high_tr = lora["high_tr"].split(".")[0].split("/")[-1] + "Hh"
|
| 270 |
-
name_low_tr = lora["low_tr"].split(".")[0].split("/")[-1] + "Ll"
|
| 271 |
-
|
| 272 |
-
try:
|
| 273 |
-
pipe.load_lora_weights(
|
| 274 |
-
lora["repo_id"],
|
| 275 |
-
weight_name=lora["high_tr"],
|
| 276 |
-
adapter_name=name_high_tr
|
| 277 |
-
)
|
| 278 |
-
|
| 279 |
-
kwargs_lora = {"load_into_transformer_2": True}
|
| 280 |
-
pipe.load_lora_weights(
|
| 281 |
-
lora["repo_id"],
|
| 282 |
-
weight_name=lora["low_tr"],
|
| 283 |
-
adapter_name=name_low_tr,
|
| 284 |
-
**kwargs_lora
|
| 285 |
-
)
|
| 286 |
-
|
| 287 |
-
pipe.set_adapters([name_high_tr, name_low_tr], adapter_weights=[1.0, 1.0])
|
| 288 |
-
|
| 289 |
-
pipe.fuse_lora(adapter_names=[name_high_tr], lora_scale=lora["high_scale"], components=["transformer"])
|
| 290 |
-
pipe.fuse_lora(adapter_names=[name_low_tr], lora_scale=lora["low_scale"], components=["transformer_2"])
|
| 291 |
-
|
| 292 |
-
pipe.unload_lora_weights()
|
| 293 |
-
|
| 294 |
-
print(f"Applied: {lora['high_tr']}, hs={lora['high_scale']}/ls={lora['low_scale']}, {i+1}/{len(LORA_MODELS)}")
|
| 295 |
-
except Exception as e:
|
| 296 |
-
print("Error:", str(e))
|
| 297 |
-
print("Failed LoRA:", name_high_tr)
|
| 298 |
-
pipe.unload_lora_weights()
|
| 299 |
-
|
| 300 |
-
# if os.path.exists(CACHE_DIR):
|
| 301 |
-
# shutil.rmtree(CACHE_DIR)
|
| 302 |
-
# print("Deleted Hugging Face cache.")
|
| 303 |
-
# else:
|
| 304 |
-
# print("No hub cache found.")
|
| 305 |
-
|
| 306 |
-
quantize_(pipe.text_encoder, Int8WeightOnlyConfig())
|
| 307 |
-
torch._dynamo.reset()
|
| 308 |
-
quantize_(pipe.transformer, Float8DynamicActivationFloat8WeightConfig())
|
| 309 |
-
torch._dynamo.reset()
|
| 310 |
-
quantize_(pipe.transformer_2, Float8DynamicActivationFloat8WeightConfig())
|
| 311 |
-
torch._dynamo.reset()
|
| 312 |
-
|
| 313 |
-
spaces.aoti_load(
|
| 314 |
-
module=pipe.transformer,
|
| 315 |
-
repo_id='thornmaze/WanTransformer3DModel-sm120-cu130-raa',
|
| 316 |
-
)
|
| 317 |
-
spaces.aoti_load(
|
| 318 |
-
module=pipe.transformer_2,
|
| 319 |
-
repo_id='thornmaze/WanTransformer3DModel-sm120-cu130-raa',
|
| 320 |
-
)
|
| 321 |
-
|
| 322 |
-
# pipe.vae.enable_slicing()
|
| 323 |
-
# pipe.vae.enable_tiling()
|
| 324 |
-
|
| 325 |
-
default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
|
| 326 |
-
default_negative_prompt = "色调艳丽, 过曝, 静态, 细节模糊不清, 字幕, 风格, 作品, 画作, 画面, 静止, 整体发灰, 最差质量, 低质量, JPEG压缩残留, 丑陋的, 残缺的, 多余的手指, 画得不好的手部, 画得不好的脸部, 畸形的, 毁容的, 形态畸形的肢体, 手指融合, 静止不动的画面, 杂乱的背景, 三条腿, 背景人很
|
| 327 |
-
|
| 328 |
-
|
| 329 |
-
def model_title():
|
| 330 |
-
return "## Wan 2.2 I2V 14B Lightning — NSFW"
|
| 331 |
-
|
| 332 |
-
|
| 333 |
-
def resize_image(image: Image.Image) -> Image.Image:
|
| 334 |
-
width, height = image.size
|
| 335 |
-
if width == height:
|
| 336 |
-
return image.resize((SQUARE_DIM, SQUARE_DIM), Image.LANCZOS)
|
| 337 |
-
|
| 338 |
-
aspect_ratio = width / height
|
| 339 |
-
MAX_ASPECT_RATIO = MAX_DIM / MIN_DIM
|
| 340 |
-
MIN_ASPECT_RATIO = MIN_DIM / MAX_DIM
|
| 341 |
-
|
| 342 |
-
image_to_resize = image
|
| 343 |
-
if aspect_ratio > MAX_ASPECT_RATIO:
|
| 344 |
-
target_w, target_h = MAX_DIM, MIN_DIM
|
| 345 |
-
crop_width = int(round(height * MAX_ASPECT_RATIO))
|
| 346 |
-
left = (width - crop_width) // 2
|
| 347 |
-
image_to_resize = image.crop((left, 0, left + crop_width, height))
|
| 348 |
-
elif aspect_ratio < MIN_ASPECT_RATIO:
|
| 349 |
-
target_w, target_h = MIN_DIM, MAX_DIM
|
| 350 |
-
crop_height = int(round(width / MIN_ASPECT_RATIO))
|
| 351 |
-
top = (height - crop_height) // 2
|
| 352 |
-
image_to_resize = image.crop((0, top, width, top + crop_height))
|
| 353 |
-
else:
|
| 354 |
-
if width > height:
|
| 355 |
-
target_w = MAX_DIM
|
| 356 |
-
target_h = int(round(target_w / aspect_ratio))
|
| 357 |
-
else:
|
| 358 |
-
target_h = MAX_DIM
|
| 359 |
-
target_w = int(round(target_h * aspect_ratio))
|
| 360 |
-
|
| 361 |
-
final_w = round(target_w / MULTIPLE_OF) * MULTIPLE_OF
|
| 362 |
-
final_h = round(target_h / MULTIPLE_OF) * MULTIPLE_OF
|
| 363 |
-
final_w = max(MIN_DIM, min(MAX_DIM, final_w))
|
| 364 |
-
final_h = max(MIN_DIM, min(MAX_DIM, final_h))
|
| 365 |
-
return image_to_resize.resize((final_w, final_h), Image.LANCZOS)
|
| 366 |
-
|
| 367 |
-
|
| 368 |
-
def resize_and_crop_to_match(target_image, reference_image):
|
| 369 |
-
ref_width, ref_height = reference_image.size
|
| 370 |
-
target_width, target_height = target_image.size
|
| 371 |
-
scale = max(ref_width / target_width, ref_height / target_height)
|
| 372 |
-
new_width, new_height = int(target_width * scale), int(target_height * scale)
|
| 373 |
-
resized = target_image.resize((new_width, new_height), Image.Resampling.LANCZOS)
|
| 374 |
-
left, top = (new_width - ref_width) // 2, (new_height - ref_height) // 2
|
| 375 |
-
return resized.crop((left, top, left + ref_width, top + ref_height))
|
| 376 |
-
|
| 377 |
-
|
| 378 |
-
def get_num_frames(duration_seconds: float):
|
| 379 |
-
raw = int(round(duration_seconds * FIXED_FPS))
|
| 380 |
-
raw = max(MIN_FRAMES_MODEL, min(MAX_FRAMES_MODEL, raw))
|
| 381 |
-
return ((raw - 1) // 4) * 4 + 1
|
| 382 |
-
|
| 383 |
-
|
| 384 |
-
def get_inference_duration(
|
| 385 |
-
resized_image,
|
| 386 |
-
processed_last_image,
|
| 387 |
-
prompt,
|
| 388 |
-
steps,
|
| 389 |
-
negative_prompt,
|
| 390 |
-
num_frames,
|
| 391 |
-
guidance_scale,
|
| 392 |
-
guidance_scale_2,
|
| 393 |
-
current_seed,
|
| 394 |
-
scheduler_name,
|
| 395 |
-
flow_shift,
|
| 396 |
-
frame_multiplier,
|
| 397 |
-
quality,
|
| 398 |
-
duration_seconds,
|
| 399 |
-
safe_mode,
|
| 400 |
-
lora_groups,
|
| 401 |
-
progress
|
| 402 |
-
):
|
| 403 |
-
BASE_FRAMES_HEIGHT_WIDTH = 81 * 832 * 624
|
| 404 |
-
BASE_STEP_DURATION = 5.
|
| 405 |
-
width, height = resized_image.size
|
| 406 |
-
factor = num_frames * width * height / BASE_FRAMES_HEIGHT_WIDTH
|
| 407 |
-
step_duration = BASE_STEP_DURATION * factor ** 1.5
|
| 408 |
-
gen_time = int(steps) * step_duration
|
| 409 |
-
|
| 410 |
-
if guidance_scale > 1:
|
| 411 |
-
gen_time = gen_time * 2.4
|
| 412 |
-
|
| 413 |
-
frame_factor = frame_multiplier // FIXED_FPS
|
| 414 |
-
if frame_factor > 1:
|
| 415 |
-
total_out_frames = (num_frames * frame_factor) - num_frames
|
| 416 |
-
inter_time = (total_out_frames * 0.02)
|
| 417 |
-
gen_time += inter_time
|
| 418 |
-
|
| 419 |
-
total_time = 15 + gen_time
|
| 420 |
-
if safe_mode:
|
| 421 |
-
total_time = total_time * 1.30
|
| 422 |
-
|
| 423 |
-
return total_time
|
| 424 |
-
|
| 425 |
-
|
| 426 |
-
@spaces.GPU(duration=get_inference_duration, size='xlarge')
|
| 427 |
-
def run_inference(
|
| 428 |
-
resized_image,
|
| 429 |
-
processed_last_image,
|
| 430 |
-
prompt,
|
| 431 |
-
steps,
|
| 432 |
-
negative_prompt,
|
| 433 |
-
num_frames,
|
| 434 |
-
guidance_scale,
|
| 435 |
-
guidance_scale_2,
|
| 436 |
-
current_seed,
|
| 437 |
-
scheduler_name,
|
| 438 |
-
flow_shift,
|
| 439 |
-
frame_multiplier,
|
| 440 |
-
quality,
|
| 441 |
-
duration_seconds,
|
| 442 |
-
safe_mode=False,
|
| 443 |
-
lora_groups=None,
|
| 444 |
-
progress=gr.Progress(track_tqdm=True),
|
| 445 |
-
):
|
| 446 |
-
scheduler_class = SCHEDULER_MAP.get(scheduler_name)
|
| 447 |
-
if scheduler_class.__name__ != pipe.scheduler.config._class_name or flow_shift != pipe.scheduler.config.get("flow_shift", "shift"):
|
| 448 |
-
config = copy.deepcopy(original_scheduler.config)
|
| 449 |
-
if scheduler_class == FlowMatchEulerDiscreteScheduler:
|
| 450 |
-
config['shift'] = flow_shift
|
| 451 |
-
else:
|
| 452 |
-
config['flow_shift'] = flow_shift
|
| 453 |
-
pipe.scheduler = scheduler_class.from_config(config)
|
| 454 |
-
|
| 455 |
-
clear_vram()
|
| 456 |
-
|
| 457 |
-
task_name = str(uuid.uuid4())[:8]
|
| 458 |
-
print(f"Generating {num_frames} frames, task: {task_name}, {duration_seconds}, {resized_image.size}, lora={lora_groups}")
|
| 459 |
-
start = time.time()
|
| 460 |
-
|
| 461 |
-
lora_loaded = False
|
| 462 |
-
lora_adapter_names = []
|
| 463 |
-
if lora_groups:
|
| 464 |
-
try:
|
| 465 |
-
for idx, name in enumerate(lora_groups):
|
| 466 |
-
if name and name != "(None)":
|
| 467 |
-
lora_adapter_names += lora_loader.load_lora_to_pipe(
|
| 468 |
-
pipe, name, adapter_name=f"lora_{idx}"
|
| 469 |
-
)
|
| 470 |
-
# Activate ALL loaded adapters together (HIGH in transformer, LOW in
|
| 471 |
-
# transformer_2); each load_lora_weights alone would leave only the
|
| 472 |
-
# last adapter active.
|
| 473 |
-
if lora_adapter_names:
|
| 474 |
-
pipe.set_adapters(lora_adapter_names, adapter_weights=[1.0] * len(lora_adapter_names))
|
| 475 |
-
lora_loaded = True
|
| 476 |
-
print(f"LoRA loaded: {lora_groups}")
|
| 477 |
-
except Exception as e:
|
| 478 |
-
print(f"LoRA warning: {e}")
|
| 479 |
-
lora_loader.unload_lora(pipe)
|
| 480 |
-
|
| 481 |
-
# LoRA prompt + user prompt combined — lora triggers first for activation
|
| 482 |
-
try:
|
| 483 |
-
additions = lora_loader.get_lora_prompt_additions(lora_groups)
|
| 484 |
-
except Exception:
|
| 485 |
-
additions = ""
|
| 486 |
-
if additions:
|
| 487 |
-
prompt = f"{additions} {prompt}"
|
| 488 |
-
print(f"Combined prompt: {prompt}")
|
| 489 |
-
|
| 490 |
-
result = pipe(
|
| 491 |
-
image=resized_image,
|
| 492 |
-
last_image=processed_last_image,
|
| 493 |
-
prompt=prompt,
|
| 494 |
-
negative_prompt=negative_prompt,
|
| 495 |
-
height=resized_image.height,
|
| 496 |
-
width=resized_image.width,
|
| 497 |
-
num_frames=num_frames,
|
| 498 |
-
guidance_scale=float(guidance_scale),
|
| 499 |
-
guidance_scale_2=float(guidance_scale_2),
|
| 500 |
-
num_inference_steps=int(steps),
|
| 501 |
-
generator=torch.Generator(device="cuda").manual_seed(current_seed),
|
| 502 |
-
output_type="np"
|
| 503 |
-
)
|
| 504 |
-
|
| 505 |
-
if lora_loaded:
|
| 506 |
-
lora_loader.unload_lora(pipe)
|
| 507 |
-
|
| 508 |
-
print("gen time passed:", time.time() - start)
|
| 509 |
-
|
| 510 |
-
raw_frames_np = result.frames[0] # Returns (T, H, W, C) float32
|
| 511 |
-
pipe.scheduler = original_scheduler
|
| 512 |
-
|
| 513 |
-
frame_factor = frame_multiplier // FIXED_FPS
|
| 514 |
-
if frame_factor > 1:
|
| 515 |
-
start = time.time()
|
| 516 |
-
print(f"Processing frames (RIFE Multiplier: {frame_factor}x)...")
|
| 517 |
-
rife_model.device()
|
| 518 |
-
rife_model.flownet = rife_model.flownet.half()
|
| 519 |
-
final_frames = interpolate_bits(raw_frames_np, multiplier=int(frame_factor))
|
| 520 |
-
print("Interpolation time passed:", time.time() - start)
|
| 521 |
-
else:
|
| 522 |
-
final_frames = list(raw_frames_np)
|
| 523 |
-
|
| 524 |
-
final_fps = FIXED_FPS * int(frame_factor)
|
| 525 |
-
|
| 526 |
-
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
|
| 527 |
-
video_path = tmpfile.name
|
| 528 |
-
|
| 529 |
-
start = time.time()
|
| 530 |
-
with tqdm(total=3, desc="Rendering Media", unit="clip") as pbar:
|
| 531 |
-
pbar.update(2)
|
| 532 |
-
export_to_video(final_frames, video_path, fps=final_fps, quality=quality)
|
| 533 |
-
pbar.update(1)
|
| 534 |
-
print(f"Export time passed, {final_fps} FPS:", time.time() - start)
|
| 535 |
-
|
| 536 |
-
return video_path, task_name
|
| 537 |
-
|
| 538 |
-
|
| 539 |
-
def generate_video(
|
| 540 |
-
input_image,
|
| 541 |
-
last_image,
|
| 542 |
-
prompt,
|
| 543 |
-
steps=4,
|
| 544 |
-
negative_prompt=default_negative_prompt,
|
| 545 |
-
duration_seconds=MAX_DURATION,
|
| 546 |
-
guidance_scale=1,
|
| 547 |
-
guidance_scale_2=1,
|
| 548 |
-
seed=42,
|
| 549 |
-
randomize_seed=False,
|
| 550 |
-
quality=5,
|
| 551 |
-
scheduler="UniPCMultistep",
|
| 552 |
-
flow_shift=6.0,
|
| 553 |
-
frame_multiplier=16,
|
| 554 |
-
safe_mode=False,
|
| 555 |
-
lora_groups=None,
|
| 556 |
-
auto_lora_enabled=True,
|
| 557 |
-
video_component=True,
|
| 558 |
-
progress=gr.Progress(track_tqdm=True),
|
| 559 |
-
):
|
| 560 |
-
"""
|
| 561 |
-
Generate a video from an input image using the Wan 2.2 14B I2V model with Lightning LoRA.
|
| 562 |
-
This function takes an input image and generates a video animation based on the provided
|
| 563 |
-
prompt and parameters. It uses an FP8 qunatized Wan 2.2 14B Image-to-Video model in with Lightning LoRA
|
| 564 |
-
for fast generation in 4-8 steps.
|
| 565 |
-
Args:
|
| 566 |
-
input_image (PIL.Image): The input image to animate. Will be resized to target dimensions.
|
| 567 |
-
last_image (PIL.Image, optional): The optional last image for the video.
|
| 568 |
-
prompt (str): Text prompt describing the desired animation or motion.
|
| 569 |
-
steps (int, optional): Number of inference steps. More steps = higher quality but slower.
|
| 570 |
-
Defaults to 4. Range: 1-30.
|
| 571 |
-
negative_prompt (str, optional): Negative prompt to avoid unwanted elements.
|
| 572 |
-
Defaults to default_negative_prompt (contains unwanted visual artifacts).
|
| 573 |
-
duration_seconds (float, optional): Duration of the generated video in seconds.
|
| 574 |
-
Defaults to 2. Clamped between MIN_FRAMES_MODEL/FIXED_FPS and MAX_FRAMES_MODEL/FIXED_FPS.
|
| 575 |
-
guidance_scale (float, optional): Controls adherence to the prompt. Higher values = more adherence.
|
| 576 |
-
Defaults to 1.0. Range: 0.0-20.0.
|
| 577 |
-
guidance_scale_2 (float, optional): Controls adherence to the prompt. Higher values = more adherence.
|
| 578 |
-
Defaults to 1.0. Range: 0.0-20.0.
|
| 579 |
-
seed (int, optional): Random seed for reproducible results. Defaults to 42.
|
| 580 |
-
Range: 0 to MAX_SEED (2147483647).
|
| 581 |
-
randomize_seed (bool, optional): Whether to use a random seed instead of the provided seed.
|
| 582 |
-
Defaults to False.
|
| 583 |
-
quality (float, optional): Video output quality. Default is 5. Uses variable bit rate.
|
| 584 |
-
Highest quality is 10, lowest is 1.
|
| 585 |
-
scheduler (str, optional): The name of the scheduler to use for inference. Defaults to "UniPCMultistep".
|
| 586 |
-
flow_shift (float, optional): The flow shift value for compatible schedulers. Defaults to 6.0.
|
| 587 |
-
frame_multiplier (int, optional): The int value for fps enhancer
|
| 588 |
-
video_component(bool, optional): Show video player in output.
|
| 589 |
-
Defaults to True.
|
| 590 |
-
progress (gr.Progress, optional): Gradio progress tracker. Defaults to gr.Progress(track_tqdm=True).
|
| 591 |
-
Returns:
|
| 592 |
-
tuple: A tuple containing:
|
| 593 |
-
- video_path (str): Path for the video component.
|
| 594 |
-
- video_path (str): Path for the file download component. Attempt to avoid reconversion in video component.
|
| 595 |
-
- current_seed (int): The seed used for generation.
|
| 596 |
-
Raises:
|
| 597 |
-
gr.Error: If input_image is None (no image uploaded).
|
| 598 |
-
Note:
|
| 599 |
-
- Frame count is calculated as duration_seconds * FIXED_FPS (24)
|
| 600 |
-
- Output dimensions are adjusted to be multiples of MOD_VALUE (32)
|
| 601 |
-
- The function uses GPU acceleration via the @spaces.GPU decorator
|
| 602 |
-
- Generation time varies based on steps and duration (see get_duration function)
|
| 603 |
-
"""
|
| 604 |
-
|
| 605 |
-
if input_image is None:
|
| 606 |
-
raise gr.Error("Please upload an input image.")
|
| 607 |
-
|
| 608 |
-
num_frames = get_num_frames(duration_seconds)
|
| 609 |
-
current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
|
| 610 |
-
resized_image = resize_image(input_image)
|
| 611 |
-
|
| 612 |
-
processed_last_image = None
|
| 613 |
-
if last_image:
|
| 614 |
-
processed_last_image = resize_and_crop_to_match(last_image, resized_image)
|
| 615 |
-
|
| 616 |
-
# Auto-activate LoRAs from the prompt when enabled, merging with manual picks
|
| 617 |
-
if auto_lora_enabled and prompt:
|
| 618 |
-
try:
|
| 619 |
-
auto_loras = lora_loader.auto_activate_loras(prompt)
|
| 620 |
-
except Exception as e:
|
| 621 |
-
auto_loras = []
|
| 622 |
-
print(f"Auto-LoRA detection error: {e}")
|
| 623 |
-
if auto_loras:
|
| 624 |
-
print(f"Auto-activated LoRAs: {auto_loras}")
|
| 625 |
-
manual = list(lora_groups or [])
|
| 626 |
-
merged = list(auto_loras)
|
| 627 |
-
for m in manual:
|
| 628 |
-
if m not in merged:
|
| 629 |
-
merged.append(m)
|
| 630 |
-
lora_groups = merged
|
| 631 |
-
|
| 632 |
-
video_path, task_n = run_inference(
|
| 633 |
-
resized_image,
|
| 634 |
-
processed_last_image,
|
| 635 |
-
prompt,
|
| 636 |
-
steps,
|
| 637 |
-
negative_prompt,
|
| 638 |
-
num_frames,
|
| 639 |
-
guidance_scale,
|
| 640 |
-
guidance_scale_2,
|
| 641 |
-
current_seed,
|
| 642 |
-
scheduler,
|
| 643 |
-
flow_shift,
|
| 644 |
-
frame_multiplier,
|
| 645 |
-
quality,
|
| 646 |
-
duration_seconds,
|
| 647 |
-
safe_mode,
|
| 648 |
-
lora_groups,
|
| 649 |
-
progress,
|
| 650 |
-
)
|
| 651 |
-
print(f"GPU complete: {task_n}")
|
| 652 |
-
|
| 653 |
-
return (video_path if video_component else None), video_path, current_seed
|
| 654 |
-
|
| 655 |
-
|
| 656 |
-
def _update_auto_lora_preview(prompt, auto_lora_enabled):
|
| 657 |
-
if not auto_lora_enabled or not prompt:
|
| 658 |
-
return ""
|
| 659 |
-
try:
|
| 660 |
-
details = lora_loader.auto_activation_details(prompt)
|
| 661 |
-
except Exception as e:
|
| 662 |
-
return f"<span style='color:#888'>auto-detection error: {e}</span>"
|
| 663 |
-
if not details:
|
| 664 |
-
return "<span style='color:#888'>No LoRAs matched.</span>"
|
| 665 |
-
items = []
|
| 666 |
-
for d in details:
|
| 667 |
-
phrases = ", ".join(d.get("matched", [])) or "—"
|
| 668 |
-
items.append(f"<li><b>{d['group']}</b> <span style='color:#888'>(score {d['score']})</span><br><small>{phrases}</small></li>")
|
| 669 |
-
return "<b>Auto-activated:</b><ul style='margin:2px 0;padding-left:18px'>" + "".join(items) + "</ul>"
|
| 670 |
-
|
| 671 |
-
|
| 672 |
-
CSS = """
|
| 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 |
)
|
|
|
|
| 1 |
+
import os; os.system('pip install --upgrade --no-deps spaces')
|
| 2 |
+
import spaces
|
| 3 |
+
import shutil
|
| 4 |
+
import subprocess
|
| 5 |
+
import sys
|
| 6 |
+
import copy
|
| 7 |
+
import random
|
| 8 |
+
import tempfile
|
| 9 |
+
import warnings
|
| 10 |
+
import time
|
| 11 |
+
import gc
|
| 12 |
+
import uuid
|
| 13 |
+
from tqdm import tqdm
|
| 14 |
+
import cv2
|
| 15 |
+
import numpy as np
|
| 16 |
+
import torch
|
| 17 |
+
import torch._dynamo
|
| 18 |
+
from torch.nn import functional as F
|
| 19 |
+
from PIL import Image
|
| 20 |
+
|
| 21 |
+
import gradio as gr
|
| 22 |
+
from diffusers import (
|
| 23 |
+
FlowMatchEulerDiscreteScheduler,
|
| 24 |
+
SASolverScheduler,
|
| 25 |
+
DEISMultistepScheduler,
|
| 26 |
+
DPMSolverMultistepInverseScheduler,
|
| 27 |
+
UniPCMultistepScheduler,
|
| 28 |
+
DPMSolverMultistepScheduler,
|
| 29 |
+
DPMSolverSinglestepScheduler,
|
| 30 |
+
)
|
| 31 |
+
from diffusers.pipelines.wan.pipeline_wan_i2v import WanImageToVideoPipeline
|
| 32 |
+
from diffusers.utils.export_utils import export_to_video
|
| 33 |
+
|
| 34 |
+
from torchao.quantization import quantize_, Float8DynamicActivationFloat8WeightConfig, Int8WeightOnlyConfig
|
| 35 |
+
import aoti
|
| 36 |
+
import lora_loader
|
| 37 |
+
|
| 38 |
+
os.environ["TOKENIZERS_PARALLELISM"] = "true"
|
| 39 |
+
warnings.filterwarnings("ignore")
|
| 40 |
+
IS_ZERO_GPU = bool(os.getenv("SPACES_ZERO_GPU"))
|
| 41 |
+
|
| 42 |
+
# if IS_ZERO_GPU:
|
| 43 |
+
# print("Loading...")
|
| 44 |
+
# subprocess.run("rm -rf /data-nvme/zerogpu-offload/*", env={}, shell=True)
|
| 45 |
+
|
| 46 |
+
# --- FRAME EXTRACTION JS & LOGIC ---
|
| 47 |
+
|
| 48 |
+
# JS to grab timestamp from the output video
|
| 49 |
+
get_timestamp_js = """
|
| 50 |
+
function() {
|
| 51 |
+
// Select the video element specifically inside the component with id 'generated-video'
|
| 52 |
+
const video = document.querySelector('#generated-video video');
|
| 53 |
+
|
| 54 |
+
if (video) {
|
| 55 |
+
console.log("Video found! Time: " + video.currentTime);
|
| 56 |
+
return video.currentTime;
|
| 57 |
+
} else {
|
| 58 |
+
console.log("No video element found.");
|
| 59 |
+
return 0;
|
| 60 |
+
}
|
| 61 |
+
}
|
| 62 |
+
"""
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
def extract_frame(video_path, timestamp):
|
| 66 |
+
# Safety check: if no video is present
|
| 67 |
+
if not video_path:
|
| 68 |
+
return None
|
| 69 |
+
|
| 70 |
+
print(f"Extracting frame at timestamp: {timestamp}")
|
| 71 |
+
|
| 72 |
+
cap = cv2.VideoCapture(video_path)
|
| 73 |
+
|
| 74 |
+
if not cap.isOpened():
|
| 75 |
+
return None
|
| 76 |
+
|
| 77 |
+
# Calculate frame number
|
| 78 |
+
fps = cap.get(cv2.CAP_PROP_FPS)
|
| 79 |
+
target_frame_num = int(float(timestamp) * fps)
|
| 80 |
+
|
| 81 |
+
# Cap total frames to prevent errors at the very end of video
|
| 82 |
+
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
| 83 |
+
if target_frame_num >= total_frames:
|
| 84 |
+
target_frame_num = total_frames - 1
|
| 85 |
+
|
| 86 |
+
# Set position
|
| 87 |
+
cap.set(cv2.CAP_PROP_POS_FRAMES, target_frame_num)
|
| 88 |
+
ret, frame = cap.read()
|
| 89 |
+
cap.release()
|
| 90 |
+
|
| 91 |
+
if ret:
|
| 92 |
+
# Convert from BGR (OpenCV) to RGB (Gradio)
|
| 93 |
+
# Gradio Image component handles Numpy array -> PIL conversion automatically
|
| 94 |
+
return cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
| 95 |
+
|
| 96 |
+
return None
|
| 97 |
+
|
| 98 |
+
# --- END FRAME EXTRACTION LOGIC ---
|
| 99 |
+
|
| 100 |
+
|
| 101 |
+
def clear_vram():
|
| 102 |
+
gc.collect()
|
| 103 |
+
torch.cuda.empty_cache()
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
# RIFE
|
| 107 |
+
if not os.path.exists("RIFEv4.26_0921.zip"):
|
| 108 |
+
print("Downloading RIFE Model...")
|
| 109 |
+
subprocess.run([
|
| 110 |
+
"wget", "-q",
|
| 111 |
+
"https://huggingface.co/thornmaze/RIFE/resolve/main/RIFEv4.26_0921.zip",
|
| 112 |
+
"-O", "RIFEv4.26_0921.zip"
|
| 113 |
+
], check=True)
|
| 114 |
+
subprocess.run(["unzip", "-o", "RIFEv4.26_0921.zip"], check=True)
|
| 115 |
+
|
| 116 |
+
# sys.path.append(os.getcwd())
|
| 117 |
+
|
| 118 |
+
from train_log.RIFE_HDv3 import Model
|
| 119 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 120 |
+
rife_model = Model()
|
| 121 |
+
rife_model.load_model("train_log", -1)
|
| 122 |
+
rife_model.eval()
|
| 123 |
+
|
| 124 |
+
|
| 125 |
+
@torch.no_grad()
|
| 126 |
+
def interpolate_bits(frames_np, multiplier=2, scale=1.0):
|
| 127 |
+
"""
|
| 128 |
+
Interpolation maintaining Numpy Float 0-1 format.
|
| 129 |
+
Args:
|
| 130 |
+
frames_np: Numpy Array (Time, Height, Width, Channels) - Float32 [0.0, 1.0]
|
| 131 |
+
multiplier: int (2, 4, 8)
|
| 132 |
+
Returns:
|
| 133 |
+
List of Numpy Arrays (Height, Width, Channels) - Float32 [0.0, 1.0]
|
| 134 |
+
"""
|
| 135 |
+
|
| 136 |
+
# Handle input shape
|
| 137 |
+
if isinstance(frames_np, list):
|
| 138 |
+
# Convert list of arrays to one big array for easier shape handling if needed,
|
| 139 |
+
# but here we just grab dims from first frame
|
| 140 |
+
T = len(frames_np)
|
| 141 |
+
H, W, C = frames_np[0].shape
|
| 142 |
+
else:
|
| 143 |
+
T, H, W, C = frames_np.shape
|
| 144 |
+
|
| 145 |
+
# 1. No Interpolation Case
|
| 146 |
+
if multiplier < 2:
|
| 147 |
+
# Just convert 4D array to list of 3D arrays
|
| 148 |
+
if isinstance(frames_np, np.ndarray):
|
| 149 |
+
return list(frames_np)
|
| 150 |
+
return frames_np
|
| 151 |
+
|
| 152 |
+
n_interp = multiplier - 1
|
| 153 |
+
|
| 154 |
+
# Pre-calc padding for RIFE (requires dimensions divisible by 32/scale)
|
| 155 |
+
tmp = max(128, int(128 / scale))
|
| 156 |
+
ph = ((H - 1) // tmp + 1) * tmp
|
| 157 |
+
pw = ((W - 1) // tmp + 1) * tmp
|
| 158 |
+
padding = (0, pw - W, 0, ph - H)
|
| 159 |
+
|
| 160 |
+
# Helper: Numpy (H, W, C) Float -> Tensor (1, C, H, W) Half
|
| 161 |
+
def to_tensor(frame_np):
|
| 162 |
+
# frame_np is float32 0-1
|
| 163 |
+
t = torch.from_numpy(frame_np).to(device)
|
| 164 |
+
# HWC -> CHW
|
| 165 |
+
t = t.permute(2, 0, 1).unsqueeze(0)
|
| 166 |
+
return F.pad(t, padding).half()
|
| 167 |
+
|
| 168 |
+
# Helper: Tensor (1, C, H, W) Half -> Numpy (H, W, C) Float
|
| 169 |
+
def from_tensor(tensor):
|
| 170 |
+
# Crop padding
|
| 171 |
+
t = tensor[0, :, :H, :W]
|
| 172 |
+
# CHW -> HWC
|
| 173 |
+
t = t.permute(1, 2, 0)
|
| 174 |
+
# Keep as float32, range 0-1
|
| 175 |
+
return t.float().cpu().numpy()
|
| 176 |
+
|
| 177 |
+
def make_inference(I0, I1, n):
|
| 178 |
+
if rife_model.version >= 3.9:
|
| 179 |
+
res = []
|
| 180 |
+
for i in range(n):
|
| 181 |
+
res.append(rife_model.inference(I0, I1, (i+1) * 1. / (n+1), scale))
|
| 182 |
+
return res
|
| 183 |
+
else:
|
| 184 |
+
middle = rife_model.inference(I0, I1, scale)
|
| 185 |
+
if n == 1:
|
| 186 |
+
return [middle]
|
| 187 |
+
first_half = make_inference(I0, middle, n=n//2)
|
| 188 |
+
second_half = make_inference(middle, I1, n=n//2)
|
| 189 |
+
if n % 2:
|
| 190 |
+
return [*first_half, middle, *second_half]
|
| 191 |
+
else:
|
| 192 |
+
return [*first_half, *second_half]
|
| 193 |
+
|
| 194 |
+
output_frames = []
|
| 195 |
+
|
| 196 |
+
# Process Frames
|
| 197 |
+
# Load first frame into GPU
|
| 198 |
+
I1 = to_tensor(frames_np[0])
|
| 199 |
+
|
| 200 |
+
total_steps = T - 1
|
| 201 |
+
|
| 202 |
+
with tqdm(total=total_steps, desc="Interpolating", unit="frame") as pbar:
|
| 203 |
+
|
| 204 |
+
for i in range(total_steps):
|
| 205 |
+
I0 = I1
|
| 206 |
+
# Add original frame to output
|
| 207 |
+
output_frames.append(from_tensor(I0))
|
| 208 |
+
|
| 209 |
+
# Load next frame
|
| 210 |
+
I1 = to_tensor(frames_np[i+1])
|
| 211 |
+
|
| 212 |
+
# Generate intermediate frames
|
| 213 |
+
mid_tensors = make_inference(I0, I1, n_interp)
|
| 214 |
+
|
| 215 |
+
# Append intermediate frames
|
| 216 |
+
for mid in mid_tensors:
|
| 217 |
+
output_frames.append(from_tensor(mid))
|
| 218 |
+
|
| 219 |
+
if (i + 1) % 50 == 0:
|
| 220 |
+
pbar.update(50)
|
| 221 |
+
pbar.update(total_steps % 50)
|
| 222 |
+
|
| 223 |
+
# Add the very last frame
|
| 224 |
+
output_frames.append(from_tensor(I1))
|
| 225 |
+
|
| 226 |
+
# Cleanup
|
| 227 |
+
del I0, I1, mid_tensors
|
| 228 |
+
torch.cuda.empty_cache()
|
| 229 |
+
|
| 230 |
+
return output_frames
|
| 231 |
+
|
| 232 |
+
|
| 233 |
+
# WAN
|
| 234 |
+
|
| 235 |
+
MODEL_ID = "thornmaze/WAMU_v3_WAN2.2_I2V_LIGHTNING"
|
| 236 |
+
|
| 237 |
+
LORA_MODELS = []
|
| 238 |
+
|
| 239 |
+
MAX_DIM = 832
|
| 240 |
+
MIN_DIM = 480
|
| 241 |
+
SQUARE_DIM = 640
|
| 242 |
+
MULTIPLE_OF = 16
|
| 243 |
+
MAX_SEED = np.iinfo(np.int32).max
|
| 244 |
+
|
| 245 |
+
FIXED_FPS = 16
|
| 246 |
+
MIN_FRAMES_MODEL = 8
|
| 247 |
+
MAX_FRAMES_MODEL = 321
|
| 248 |
+
|
| 249 |
+
MIN_DURATION = round(MIN_FRAMES_MODEL / FIXED_FPS, 1)
|
| 250 |
+
MAX_DURATION = round(MAX_FRAMES_MODEL / FIXED_FPS, 1)
|
| 251 |
+
|
| 252 |
+
SCHEDULER_MAP = {
|
| 253 |
+
"FlowMatchEulerDiscrete": FlowMatchEulerDiscreteScheduler,
|
| 254 |
+
"SASolver": SASolverScheduler,
|
| 255 |
+
"DEISMultistep": DEISMultistepScheduler,
|
| 256 |
+
"DPMSolverMultistepInverse": DPMSolverMultistepInverseScheduler,
|
| 257 |
+
"UniPCMultistep": UniPCMultistepScheduler,
|
| 258 |
+
"DPMSolverMultistep": DPMSolverMultistepScheduler,
|
| 259 |
+
"DPMSolverSinglestep": DPMSolverSinglestepScheduler,
|
| 260 |
+
}
|
| 261 |
+
|
| 262 |
+
pipe = WanImageToVideoPipeline.from_pretrained(
|
| 263 |
+
MODEL_ID,
|
| 264 |
+
torch_dtype=torch.bfloat16,
|
| 265 |
+
).to('cuda')
|
| 266 |
+
original_scheduler = copy.deepcopy(pipe.scheduler)
|
| 267 |
+
|
| 268 |
+
for i, lora in enumerate(LORA_MODELS):
|
| 269 |
+
name_high_tr = lora["high_tr"].split(".")[0].split("/")[-1] + "Hh"
|
| 270 |
+
name_low_tr = lora["low_tr"].split(".")[0].split("/")[-1] + "Ll"
|
| 271 |
+
|
| 272 |
+
try:
|
| 273 |
+
pipe.load_lora_weights(
|
| 274 |
+
lora["repo_id"],
|
| 275 |
+
weight_name=lora["high_tr"],
|
| 276 |
+
adapter_name=name_high_tr
|
| 277 |
+
)
|
| 278 |
+
|
| 279 |
+
kwargs_lora = {"load_into_transformer_2": True}
|
| 280 |
+
pipe.load_lora_weights(
|
| 281 |
+
lora["repo_id"],
|
| 282 |
+
weight_name=lora["low_tr"],
|
| 283 |
+
adapter_name=name_low_tr,
|
| 284 |
+
**kwargs_lora
|
| 285 |
+
)
|
| 286 |
+
|
| 287 |
+
pipe.set_adapters([name_high_tr, name_low_tr], adapter_weights=[1.0, 1.0])
|
| 288 |
+
|
| 289 |
+
pipe.fuse_lora(adapter_names=[name_high_tr], lora_scale=lora["high_scale"], components=["transformer"])
|
| 290 |
+
pipe.fuse_lora(adapter_names=[name_low_tr], lora_scale=lora["low_scale"], components=["transformer_2"])
|
| 291 |
+
|
| 292 |
+
pipe.unload_lora_weights()
|
| 293 |
+
|
| 294 |
+
print(f"Applied: {lora['high_tr']}, hs={lora['high_scale']}/ls={lora['low_scale']}, {i+1}/{len(LORA_MODELS)}")
|
| 295 |
+
except Exception as e:
|
| 296 |
+
print("Error:", str(e))
|
| 297 |
+
print("Failed LoRA:", name_high_tr)
|
| 298 |
+
pipe.unload_lora_weights()
|
| 299 |
+
|
| 300 |
+
# if os.path.exists(CACHE_DIR):
|
| 301 |
+
# shutil.rmtree(CACHE_DIR)
|
| 302 |
+
# print("Deleted Hugging Face cache.")
|
| 303 |
+
# else:
|
| 304 |
+
# print("No hub cache found.")
|
| 305 |
+
|
| 306 |
+
quantize_(pipe.text_encoder, Int8WeightOnlyConfig())
|
| 307 |
+
torch._dynamo.reset()
|
| 308 |
+
quantize_(pipe.transformer, Float8DynamicActivationFloat8WeightConfig())
|
| 309 |
+
torch._dynamo.reset()
|
| 310 |
+
quantize_(pipe.transformer_2, Float8DynamicActivationFloat8WeightConfig())
|
| 311 |
+
torch._dynamo.reset()
|
| 312 |
+
|
| 313 |
+
spaces.aoti_load(
|
| 314 |
+
module=pipe.transformer,
|
| 315 |
+
repo_id='thornmaze/WanTransformer3DModel-sm120-cu130-raa',
|
| 316 |
+
)
|
| 317 |
+
spaces.aoti_load(
|
| 318 |
+
module=pipe.transformer_2,
|
| 319 |
+
repo_id='thornmaze/WanTransformer3DModel-sm120-cu130-raa',
|
| 320 |
+
)
|
| 321 |
+
|
| 322 |
+
# pipe.vae.enable_slicing()
|
| 323 |
+
# pipe.vae.enable_tiling()
|
| 324 |
+
|
| 325 |
+
default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
|
| 326 |
+
default_negative_prompt = "色调艳丽, 过曝, 静态, 细节模糊不清, 字幕, 风格, 作品, 画作, 画面, 静止, 整体发灰, 最差质量, 低质量, JPEG压缩残留, 丑陋的, 残缺的, 多余的手指, 画得不好的手部, 画得不好的脸部, 畸形的, 毁容的, 形态畸形的肢体, 手指融合, 静止不动的画面, 杂乱的背景, 三条腿, 背景人很多, 倒着走"
|
| 327 |
+
|
| 328 |
+
|
| 329 |
+
def model_title():
|
| 330 |
+
return "## Wan 2.2 I2V 14B Lightning — NSFW"
|
| 331 |
+
|
| 332 |
+
|
| 333 |
+
def resize_image(image: Image.Image) -> Image.Image:
|
| 334 |
+
width, height = image.size
|
| 335 |
+
if width == height:
|
| 336 |
+
return image.resize((SQUARE_DIM, SQUARE_DIM), Image.LANCZOS)
|
| 337 |
+
|
| 338 |
+
aspect_ratio = width / height
|
| 339 |
+
MAX_ASPECT_RATIO = MAX_DIM / MIN_DIM
|
| 340 |
+
MIN_ASPECT_RATIO = MIN_DIM / MAX_DIM
|
| 341 |
+
|
| 342 |
+
image_to_resize = image
|
| 343 |
+
if aspect_ratio > MAX_ASPECT_RATIO:
|
| 344 |
+
target_w, target_h = MAX_DIM, MIN_DIM
|
| 345 |
+
crop_width = int(round(height * MAX_ASPECT_RATIO))
|
| 346 |
+
left = (width - crop_width) // 2
|
| 347 |
+
image_to_resize = image.crop((left, 0, left + crop_width, height))
|
| 348 |
+
elif aspect_ratio < MIN_ASPECT_RATIO:
|
| 349 |
+
target_w, target_h = MIN_DIM, MAX_DIM
|
| 350 |
+
crop_height = int(round(width / MIN_ASPECT_RATIO))
|
| 351 |
+
top = (height - crop_height) // 2
|
| 352 |
+
image_to_resize = image.crop((0, top, width, top + crop_height))
|
| 353 |
+
else:
|
| 354 |
+
if width > height:
|
| 355 |
+
target_w = MAX_DIM
|
| 356 |
+
target_h = int(round(target_w / aspect_ratio))
|
| 357 |
+
else:
|
| 358 |
+
target_h = MAX_DIM
|
| 359 |
+
target_w = int(round(target_h * aspect_ratio))
|
| 360 |
+
|
| 361 |
+
final_w = round(target_w / MULTIPLE_OF) * MULTIPLE_OF
|
| 362 |
+
final_h = round(target_h / MULTIPLE_OF) * MULTIPLE_OF
|
| 363 |
+
final_w = max(MIN_DIM, min(MAX_DIM, final_w))
|
| 364 |
+
final_h = max(MIN_DIM, min(MAX_DIM, final_h))
|
| 365 |
+
return image_to_resize.resize((final_w, final_h), Image.LANCZOS)
|
| 366 |
+
|
| 367 |
+
|
| 368 |
+
def resize_and_crop_to_match(target_image, reference_image):
|
| 369 |
+
ref_width, ref_height = reference_image.size
|
| 370 |
+
target_width, target_height = target_image.size
|
| 371 |
+
scale = max(ref_width / target_width, ref_height / target_height)
|
| 372 |
+
new_width, new_height = int(target_width * scale), int(target_height * scale)
|
| 373 |
+
resized = target_image.resize((new_width, new_height), Image.Resampling.LANCZOS)
|
| 374 |
+
left, top = (new_width - ref_width) // 2, (new_height - ref_height) // 2
|
| 375 |
+
return resized.crop((left, top, left + ref_width, top + ref_height))
|
| 376 |
+
|
| 377 |
+
|
| 378 |
+
def get_num_frames(duration_seconds: float):
|
| 379 |
+
raw = int(round(duration_seconds * FIXED_FPS))
|
| 380 |
+
raw = max(MIN_FRAMES_MODEL, min(MAX_FRAMES_MODEL, raw))
|
| 381 |
+
return ((raw - 1) // 4) * 4 + 1
|
| 382 |
+
|
| 383 |
+
|
| 384 |
+
def get_inference_duration(
|
| 385 |
+
resized_image,
|
| 386 |
+
processed_last_image,
|
| 387 |
+
prompt,
|
| 388 |
+
steps,
|
| 389 |
+
negative_prompt,
|
| 390 |
+
num_frames,
|
| 391 |
+
guidance_scale,
|
| 392 |
+
guidance_scale_2,
|
| 393 |
+
current_seed,
|
| 394 |
+
scheduler_name,
|
| 395 |
+
flow_shift,
|
| 396 |
+
frame_multiplier,
|
| 397 |
+
quality,
|
| 398 |
+
duration_seconds,
|
| 399 |
+
safe_mode,
|
| 400 |
+
lora_groups,
|
| 401 |
+
progress
|
| 402 |
+
):
|
| 403 |
+
BASE_FRAMES_HEIGHT_WIDTH = 81 * 832 * 624
|
| 404 |
+
BASE_STEP_DURATION = 5.
|
| 405 |
+
width, height = resized_image.size
|
| 406 |
+
factor = num_frames * width * height / BASE_FRAMES_HEIGHT_WIDTH
|
| 407 |
+
step_duration = BASE_STEP_DURATION * factor ** 1.5
|
| 408 |
+
gen_time = int(steps) * step_duration
|
| 409 |
+
|
| 410 |
+
if guidance_scale > 1:
|
| 411 |
+
gen_time = gen_time * 2.4
|
| 412 |
+
|
| 413 |
+
frame_factor = frame_multiplier // FIXED_FPS
|
| 414 |
+
if frame_factor > 1:
|
| 415 |
+
total_out_frames = (num_frames * frame_factor) - num_frames
|
| 416 |
+
inter_time = (total_out_frames * 0.02)
|
| 417 |
+
gen_time += inter_time
|
| 418 |
+
|
| 419 |
+
total_time = 15 + gen_time
|
| 420 |
+
if safe_mode:
|
| 421 |
+
total_time = total_time * 1.30
|
| 422 |
+
|
| 423 |
+
return total_time
|
| 424 |
+
|
| 425 |
+
|
| 426 |
+
@spaces.GPU(duration=get_inference_duration, size='xlarge')
|
| 427 |
+
def run_inference(
|
| 428 |
+
resized_image,
|
| 429 |
+
processed_last_image,
|
| 430 |
+
prompt,
|
| 431 |
+
steps,
|
| 432 |
+
negative_prompt,
|
| 433 |
+
num_frames,
|
| 434 |
+
guidance_scale,
|
| 435 |
+
guidance_scale_2,
|
| 436 |
+
current_seed,
|
| 437 |
+
scheduler_name,
|
| 438 |
+
flow_shift,
|
| 439 |
+
frame_multiplier,
|
| 440 |
+
quality,
|
| 441 |
+
duration_seconds,
|
| 442 |
+
safe_mode=False,
|
| 443 |
+
lora_groups=None,
|
| 444 |
+
progress=gr.Progress(track_tqdm=True),
|
| 445 |
+
):
|
| 446 |
+
scheduler_class = SCHEDULER_MAP.get(scheduler_name)
|
| 447 |
+
if scheduler_class.__name__ != pipe.scheduler.config._class_name or flow_shift != pipe.scheduler.config.get("flow_shift", "shift"):
|
| 448 |
+
config = copy.deepcopy(original_scheduler.config)
|
| 449 |
+
if scheduler_class == FlowMatchEulerDiscreteScheduler:
|
| 450 |
+
config['shift'] = flow_shift
|
| 451 |
+
else:
|
| 452 |
+
config['flow_shift'] = flow_shift
|
| 453 |
+
pipe.scheduler = scheduler_class.from_config(config)
|
| 454 |
+
|
| 455 |
+
clear_vram()
|
| 456 |
+
|
| 457 |
+
task_name = str(uuid.uuid4())[:8]
|
| 458 |
+
print(f"Generating {num_frames} frames, task: {task_name}, {duration_seconds}, {resized_image.size}, lora={lora_groups}")
|
| 459 |
+
start = time.time()
|
| 460 |
+
|
| 461 |
+
lora_loaded = False
|
| 462 |
+
lora_adapter_names = []
|
| 463 |
+
if lora_groups:
|
| 464 |
+
try:
|
| 465 |
+
for idx, name in enumerate(lora_groups):
|
| 466 |
+
if name and name != "(None)":
|
| 467 |
+
lora_adapter_names += lora_loader.load_lora_to_pipe(
|
| 468 |
+
pipe, name, adapter_name=f"lora_{idx}"
|
| 469 |
+
)
|
| 470 |
+
# Activate ALL loaded adapters together (HIGH in transformer, LOW in
|
| 471 |
+
# transformer_2); each load_lora_weights alone would leave only the
|
| 472 |
+
# last adapter active.
|
| 473 |
+
if lora_adapter_names:
|
| 474 |
+
pipe.set_adapters(lora_adapter_names, adapter_weights=[1.0] * len(lora_adapter_names))
|
| 475 |
+
lora_loaded = True
|
| 476 |
+
print(f"LoRA loaded: {lora_groups}")
|
| 477 |
+
except Exception as e:
|
| 478 |
+
print(f"LoRA warning: {e}")
|
| 479 |
+
lora_loader.unload_lora(pipe)
|
| 480 |
+
|
| 481 |
+
# LoRA prompt + user prompt combined — lora triggers first for activation
|
| 482 |
+
try:
|
| 483 |
+
additions = lora_loader.get_lora_prompt_additions(lora_groups)
|
| 484 |
+
except Exception:
|
| 485 |
+
additions = ""
|
| 486 |
+
if additions:
|
| 487 |
+
prompt = f"{additions} {prompt}"
|
| 488 |
+
print(f"Combined prompt: {prompt}")
|
| 489 |
+
|
| 490 |
+
result = pipe(
|
| 491 |
+
image=resized_image,
|
| 492 |
+
last_image=processed_last_image,
|
| 493 |
+
prompt=prompt,
|
| 494 |
+
negative_prompt=negative_prompt,
|
| 495 |
+
height=resized_image.height,
|
| 496 |
+
width=resized_image.width,
|
| 497 |
+
num_frames=num_frames,
|
| 498 |
+
guidance_scale=float(guidance_scale),
|
| 499 |
+
guidance_scale_2=float(guidance_scale_2),
|
| 500 |
+
num_inference_steps=int(steps),
|
| 501 |
+
generator=torch.Generator(device="cuda").manual_seed(current_seed),
|
| 502 |
+
output_type="np"
|
| 503 |
+
)
|
| 504 |
+
|
| 505 |
+
if lora_loaded:
|
| 506 |
+
lora_loader.unload_lora(pipe)
|
| 507 |
+
|
| 508 |
+
print("gen time passed:", time.time() - start)
|
| 509 |
+
|
| 510 |
+
raw_frames_np = result.frames[0] # Returns (T, H, W, C) float32
|
| 511 |
+
pipe.scheduler = original_scheduler
|
| 512 |
+
|
| 513 |
+
frame_factor = frame_multiplier // FIXED_FPS
|
| 514 |
+
if frame_factor > 1:
|
| 515 |
+
start = time.time()
|
| 516 |
+
print(f"Processing frames (RIFE Multiplier: {frame_factor}x)...")
|
| 517 |
+
rife_model.device()
|
| 518 |
+
rife_model.flownet = rife_model.flownet.half()
|
| 519 |
+
final_frames = interpolate_bits(raw_frames_np, multiplier=int(frame_factor))
|
| 520 |
+
print("Interpolation time passed:", time.time() - start)
|
| 521 |
+
else:
|
| 522 |
+
final_frames = list(raw_frames_np)
|
| 523 |
+
|
| 524 |
+
final_fps = FIXED_FPS * int(frame_factor)
|
| 525 |
+
|
| 526 |
+
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
|
| 527 |
+
video_path = tmpfile.name
|
| 528 |
+
|
| 529 |
+
start = time.time()
|
| 530 |
+
with tqdm(total=3, desc="Rendering Media", unit="clip") as pbar:
|
| 531 |
+
pbar.update(2)
|
| 532 |
+
export_to_video(final_frames, video_path, fps=final_fps, quality=quality)
|
| 533 |
+
pbar.update(1)
|
| 534 |
+
print(f"Export time passed, {final_fps} FPS:", time.time() - start)
|
| 535 |
+
|
| 536 |
+
return video_path, task_name
|
| 537 |
+
|
| 538 |
+
|
| 539 |
+
def generate_video(
|
| 540 |
+
input_image,
|
| 541 |
+
last_image,
|
| 542 |
+
prompt,
|
| 543 |
+
steps=4,
|
| 544 |
+
negative_prompt=default_negative_prompt,
|
| 545 |
+
duration_seconds=MAX_DURATION,
|
| 546 |
+
guidance_scale=1,
|
| 547 |
+
guidance_scale_2=1,
|
| 548 |
+
seed=42,
|
| 549 |
+
randomize_seed=False,
|
| 550 |
+
quality=5,
|
| 551 |
+
scheduler="UniPCMultistep",
|
| 552 |
+
flow_shift=6.0,
|
| 553 |
+
frame_multiplier=16,
|
| 554 |
+
safe_mode=False,
|
| 555 |
+
lora_groups=None,
|
| 556 |
+
auto_lora_enabled=True,
|
| 557 |
+
video_component=True,
|
| 558 |
+
progress=gr.Progress(track_tqdm=True),
|
| 559 |
+
):
|
| 560 |
+
"""
|
| 561 |
+
Generate a video from an input image using the Wan 2.2 14B I2V model with Lightning LoRA.
|
| 562 |
+
This function takes an input image and generates a video animation based on the provided
|
| 563 |
+
prompt and parameters. It uses an FP8 qunatized Wan 2.2 14B Image-to-Video model in with Lightning LoRA
|
| 564 |
+
for fast generation in 4-8 steps.
|
| 565 |
+
Args:
|
| 566 |
+
input_image (PIL.Image): The input image to animate. Will be resized to target dimensions.
|
| 567 |
+
last_image (PIL.Image, optional): The optional last image for the video.
|
| 568 |
+
prompt (str): Text prompt describing the desired animation or motion.
|
| 569 |
+
steps (int, optional): Number of inference steps. More steps = higher quality but slower.
|
| 570 |
+
Defaults to 4. Range: 1-30.
|
| 571 |
+
negative_prompt (str, optional): Negative prompt to avoid unwanted elements.
|
| 572 |
+
Defaults to default_negative_prompt (contains unwanted visual artifacts).
|
| 573 |
+
duration_seconds (float, optional): Duration of the generated video in seconds.
|
| 574 |
+
Defaults to 2. Clamped between MIN_FRAMES_MODEL/FIXED_FPS and MAX_FRAMES_MODEL/FIXED_FPS.
|
| 575 |
+
guidance_scale (float, optional): Controls adherence to the prompt. Higher values = more adherence.
|
| 576 |
+
Defaults to 1.0. Range: 0.0-20.0.
|
| 577 |
+
guidance_scale_2 (float, optional): Controls adherence to the prompt. Higher values = more adherence.
|
| 578 |
+
Defaults to 1.0. Range: 0.0-20.0.
|
| 579 |
+
seed (int, optional): Random seed for reproducible results. Defaults to 42.
|
| 580 |
+
Range: 0 to MAX_SEED (2147483647).
|
| 581 |
+
randomize_seed (bool, optional): Whether to use a random seed instead of the provided seed.
|
| 582 |
+
Defaults to False.
|
| 583 |
+
quality (float, optional): Video output quality. Default is 5. Uses variable bit rate.
|
| 584 |
+
Highest quality is 10, lowest is 1.
|
| 585 |
+
scheduler (str, optional): The name of the scheduler to use for inference. Defaults to "UniPCMultistep".
|
| 586 |
+
flow_shift (float, optional): The flow shift value for compatible schedulers. Defaults to 6.0.
|
| 587 |
+
frame_multiplier (int, optional): The int value for fps enhancer
|
| 588 |
+
video_component(bool, optional): Show video player in output.
|
| 589 |
+
Defaults to True.
|
| 590 |
+
progress (gr.Progress, optional): Gradio progress tracker. Defaults to gr.Progress(track_tqdm=True).
|
| 591 |
+
Returns:
|
| 592 |
+
tuple: A tuple containing:
|
| 593 |
+
- video_path (str): Path for the video component.
|
| 594 |
+
- video_path (str): Path for the file download component. Attempt to avoid reconversion in video component.
|
| 595 |
+
- current_seed (int): The seed used for generation.
|
| 596 |
+
Raises:
|
| 597 |
+
gr.Error: If input_image is None (no image uploaded).
|
| 598 |
+
Note:
|
| 599 |
+
- Frame count is calculated as duration_seconds * FIXED_FPS (24)
|
| 600 |
+
- Output dimensions are adjusted to be multiples of MOD_VALUE (32)
|
| 601 |
+
- The function uses GPU acceleration via the @spaces.GPU decorator
|
| 602 |
+
- Generation time varies based on steps and duration (see get_duration function)
|
| 603 |
+
"""
|
| 604 |
+
|
| 605 |
+
if input_image is None:
|
| 606 |
+
raise gr.Error("Please upload an input image.")
|
| 607 |
+
|
| 608 |
+
num_frames = get_num_frames(duration_seconds)
|
| 609 |
+
current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
|
| 610 |
+
resized_image = resize_image(input_image)
|
| 611 |
+
|
| 612 |
+
processed_last_image = None
|
| 613 |
+
if last_image:
|
| 614 |
+
processed_last_image = resize_and_crop_to_match(last_image, resized_image)
|
| 615 |
+
|
| 616 |
+
# Auto-activate LoRAs from the prompt when enabled, merging with manual picks
|
| 617 |
+
if auto_lora_enabled and prompt:
|
| 618 |
+
try:
|
| 619 |
+
auto_loras = lora_loader.auto_activate_loras(prompt)
|
| 620 |
+
except Exception as e:
|
| 621 |
+
auto_loras = []
|
| 622 |
+
print(f"Auto-LoRA detection error: {e}")
|
| 623 |
+
if auto_loras:
|
| 624 |
+
print(f"Auto-activated LoRAs: {auto_loras}")
|
| 625 |
+
manual = list(lora_groups or [])
|
| 626 |
+
merged = list(auto_loras)
|
| 627 |
+
for m in manual:
|
| 628 |
+
if m not in merged:
|
| 629 |
+
merged.append(m)
|
| 630 |
+
lora_groups = merged
|
| 631 |
+
|
| 632 |
+
video_path, task_n = run_inference(
|
| 633 |
+
resized_image,
|
| 634 |
+
processed_last_image,
|
| 635 |
+
prompt,
|
| 636 |
+
steps,
|
| 637 |
+
negative_prompt,
|
| 638 |
+
num_frames,
|
| 639 |
+
guidance_scale,
|
| 640 |
+
guidance_scale_2,
|
| 641 |
+
current_seed,
|
| 642 |
+
scheduler,
|
| 643 |
+
flow_shift,
|
| 644 |
+
frame_multiplier,
|
| 645 |
+
quality,
|
| 646 |
+
duration_seconds,
|
| 647 |
+
safe_mode,
|
| 648 |
+
lora_groups,
|
| 649 |
+
progress,
|
| 650 |
+
)
|
| 651 |
+
print(f"GPU complete: {task_n}")
|
| 652 |
+
|
| 653 |
+
return (video_path if video_component else None), video_path, current_seed
|
| 654 |
+
|
| 655 |
+
|
| 656 |
+
def _update_auto_lora_preview(prompt, auto_lora_enabled):
|
| 657 |
+
if not auto_lora_enabled or not prompt:
|
| 658 |
+
return ""
|
| 659 |
+
try:
|
| 660 |
+
details = lora_loader.auto_activation_details(prompt)
|
| 661 |
+
except Exception as e:
|
| 662 |
+
return f"<span style='color:#888'>auto-detection error: {e}</span>"
|
| 663 |
+
if not details:
|
| 664 |
+
return "<span style='color:#888'>No LoRAs matched.</span>"
|
| 665 |
+
items = []
|
| 666 |
+
for d in details:
|
| 667 |
+
phrases = ", ".join(d.get("matched", [])) or "—"
|
| 668 |
+
items.append(f"<li><b>{d['group']}</b> <span style='color:#888'>(score {d['score']})</span><br><small>{phrases}</small></li>")
|
| 669 |
+
return "<b>Auto-activated:</b><ul style='margin:2px 0;padding-left:18px'>" + "".join(items) + "</ul>"
|
| 670 |
+
|
| 671 |
+
|
| 672 |
+
CSS = """
|
| 673 |
+
/* ── Hidden timestamp (preserved) ── */
|
| 674 |
+
#hidden-timestamp {
|
| 675 |
+
opacity: 0;
|
| 676 |
+
height: 0px;
|
| 677 |
+
width: 0px;
|
| 678 |
+
margin: 0px;
|
| 679 |
+
padding: 0px;
|
| 680 |
+
overflow: hidden;
|
| 681 |
+
position: absolute;
|
| 682 |
+
pointer-events: none;
|
| 683 |
+
}
|
| 684 |
+
|
| 685 |
+
/* ── GLOBAL BACKGROUND — vivid blue→green→purple gradient ── */
|
| 686 |
+
.gradio-container {
|
| 687 |
+
background: linear-gradient(135deg,
|
| 688 |
+
#0a0e2a 0%,
|
| 689 |
+
#0d1b3e 12%,
|
| 690 |
+
#0f2847 25%,
|
| 691 |
+
#0a3d2e 40%,
|
| 692 |
+
#0e5135 55%,
|
| 693 |
+
#1a3a5c 65%,
|
| 694 |
+
#2d1b69 80%,
|
| 695 |
+
#1e1145 100%) !important;
|
| 696 |
+
background-attachment: fixed !important;
|
| 697 |
+
min-height: 100vh !important;
|
| 698 |
+
color: #e8ecf4 !important;
|
| 699 |
+
font-family: 'Segoe UI', system-ui, -apple-system, sans-serif !important;
|
| 700 |
+
}
|
| 701 |
+
|
| 702 |
+
/* Subtle animated grain overlay for atmosphere */
|
| 703 |
+
.gradio-container::before {
|
| 704 |
+
content: '' !important;
|
| 705 |
+
position: fixed !important;
|
| 706 |
+
inset: 0 !important;
|
| 707 |
+
background-image:
|
| 708 |
+
radial-gradient(ellipse at 20% 20%, rgba(0, 200, 180, 0.08) 0%, transparent 60%),
|
| 709 |
+
radial-gradient(ellipse at 80% 30%, rgba(100, 60, 220, 0.1) 0%, transparent 55%),
|
| 710 |
+
radial-gradient(ellipse at 50% 80%, rgba(30, 120, 200, 0.07) 0%, transparent 50%) !important;
|
| 711 |
+
pointer-events: none !important;
|
| 712 |
+
z-index: 0 !important;
|
| 713 |
+
}
|
| 714 |
+
|
| 715 |
+
.gradio-container > * {
|
| 716 |
+
position: relative !important;
|
| 717 |
+
z-index: 1 !important;
|
| 718 |
+
}
|
| 719 |
+
|
| 720 |
+
/* ── HEADER / TITLE ── */
|
| 721 |
+
.gradio-container h1,
|
| 722 |
+
.gradio-container .gr-title,
|
| 723 |
+
.gradio-container .prose {
|
| 724 |
+
background: linear-gradient(90deg, #00e5c8, #6c5ce7, #a855f7) !important;
|
| 725 |
+
-webkit-background-clip: text !important;
|
| 726 |
+
-webkit-text-fill-color: transparent !important;
|
| 727 |
+
background-clip: text !important;
|
| 728 |
+
font-weight: 800 !important;
|
| 729 |
+
letter-spacing: -0.02em !important;
|
| 730 |
+
}
|
| 731 |
+
|
| 732 |
+
/* Subtitle text */
|
| 733 |
+
.gradio-container .prose p,
|
| 734 |
+
.gradio-container > div > div > .prose:nth-child(2) {
|
| 735 |
+
color: #b0c4de !important;
|
| 736 |
+
}
|
| 737 |
+
|
| 738 |
+
/* ── PANELS / CARDS — frosted glass ── */
|
| 739 |
+
.gradio-container .panel,
|
| 740 |
+
.gradio-container .gr-group,
|
| 741 |
+
.gradio-container .gr-block {
|
| 742 |
+
background: rgba(255, 255, 255, 0.04) !important;
|
| 743 |
+
backdrop-filter: blur(16px) saturate(1.4) !important;
|
| 744 |
+
-webkit-backdrop-filter: blur(16px) saturate(1.4) !important;
|
| 745 |
+
border: 1px solid rgba(255, 255, 255, 0.08) !important;
|
| 746 |
+
border-radius: 14px !important;
|
| 747 |
+
box-shadow:
|
| 748 |
+
0 4px 30px rgba(0, 0, 0, 0.3),
|
| 749 |
+
inset 0 1px 0 rgba(255, 255, 255, 0.06) !important;
|
| 750 |
+
}
|
| 751 |
+
|
| 752 |
+
/* Image/video upload areas */
|
| 753 |
+
.gradio-container .upload-container,
|
| 754 |
+
.gradio-container .file-upload,
|
| 755 |
+
.gradio-container [data-testid="image"] {
|
| 756 |
+
background: rgba(255, 255, 255, 0.03) !important;
|
| 757 |
+
border: 2px dashed rgba(100, 200, 255, 0.2) !important;
|
| 758 |
+
border-radius: 12px !important;
|
| 759 |
+
transition: border-color 0.3s ease, box-shadow 0.3s ease !important;
|
| 760 |
+
}
|
| 761 |
+
.gradio-container .upload-container:hover,
|
| 762 |
+
.gradio-container .file-upload:hover,
|
| 763 |
+
.gradio-container [data-testid="image"]:hover {
|
| 764 |
+
border-color: rgba(0, 229, 200, 0.5) !important;
|
| 765 |
+
box-shadow: 0 0 20px rgba(0, 229, 200, 0.1) !important;
|
| 766 |
+
}
|
| 767 |
+
|
| 768 |
+
/* ── LABELS — crisp white with subtle glow ── */
|
| 769 |
+
.gradio-container label,
|
| 770 |
+
.gradio-container .gr-label,
|
| 771 |
+
.gradio-container .gr-block-label {
|
| 772 |
+
color: #d6e4f0 !important;
|
| 773 |
+
font-weight: 600 !important;
|
| 774 |
+
text-shadow: 0 0 8px rgba(100, 180, 255, 0.15) !important;
|
| 775 |
+
}
|
| 776 |
+
|
| 777 |
+
/* ── TEXT INPUTS & TEXTBOX — glass borders ── */
|
| 778 |
+
.gradio-container input[type="text"],
|
| 779 |
+
.gradio-container textarea,
|
| 780 |
+
.gradio-container .gr-input,
|
| 781 |
+
.gradio-container .gr-textbox,
|
| 782 |
+
.gradio-container .gr-textarea {
|
| 783 |
+
background: rgba(10, 15, 40, 0.6) !important;
|
| 784 |
+
border: 1px solid rgba(100, 180, 255, 0.15) !important;
|
| 785 |
+
border-radius: 10px !important;
|
| 786 |
+
color: #e8ecf4 !important;
|
| 787 |
+
box-shadow: inset 0 2px 6px rgba(0, 0, 0, 0.2) !important;
|
| 788 |
+
transition: border-color 0.3s ease, box-shadow 0.3s ease !important;
|
| 789 |
+
}
|
| 790 |
+
.gradio-container input[type="text"]:focus,
|
| 791 |
+
.gradio-container textarea:focus,
|
| 792 |
+
.gradio-container .gr-input:focus,
|
| 793 |
+
.gradio-container .gr-textbox:focus {
|
| 794 |
+
border-color: rgba(108, 92, 231, 0.6) !important;
|
| 795 |
+
box-shadow:
|
| 796 |
+
inset 0 2px 6px rgba(0, 0, 0, 0.2),
|
| 797 |
+
0 0 12px rgba(108, 92, 231, 0.2) !important;
|
| 798 |
+
outline: none !important;
|
| 799 |
+
}
|
| 800 |
+
|
| 801 |
+
/* ── SLIDERS — glowing track ── */
|
| 802 |
+
.gradio-container input[type="range"],
|
| 803 |
+
.gradio-container .range-slider,
|
| 804 |
+
.gradio-container .gr-slider {
|
| 805 |
+
accent-color: #00d4aa !important;
|
| 806 |
+
}
|
| 807 |
+
|
| 808 |
+
.gradio-container input[type="range"]::-webkit-slider-runnable-track {
|
| 809 |
+
background: linear-gradient(90deg, #00d4aa, #6c5ce7) !important;
|
| 810 |
+
height: 6px !important;
|
| 811 |
+
border-radius: 3px !important;
|
| 812 |
+
}
|
| 813 |
+
.gradio-container input[type="range"]::-webkit-slider-thumb {
|
| 814 |
+
-webkit-appearance: none !important;
|
| 815 |
+
width: 18px !important;
|
| 816 |
+
height: 18px !important;
|
| 817 |
+
border-radius: 50% !important;
|
| 818 |
+
background: linear-gradient(135deg, #00e5c8, #6c5ce7) !important;
|
| 819 |
+
box-shadow: 0 0 10px rgba(0, 229, 200, 0.5), 0 2px 6px rgba(0, 0, 0, 0.3) !important;
|
| 820 |
+
cursor: pointer !important;
|
| 821 |
+
border: 2px solid rgba(255, 255, 255, 0.3) !important;
|
| 822 |
+
margin-top: -6px !important;
|
| 823 |
+
transition: transform 0.2s ease, box-shadow 0.2s ease !important;
|
| 824 |
+
}
|
| 825 |
+
.gradio-container input[type="range"]::-webkit-slider-thumb:hover {
|
| 826 |
+
transform: scale(1.15) !important;
|
| 827 |
+
box-shadow: 0 0 18px rgba(0, 229, 200, 0.7), 0 2px 8px rgba(0, 0, 0, 0.4) !important;
|
| 828 |
+
}
|
| 829 |
+
|
| 830 |
+
/* Number inputs */
|
| 831 |
+
.gradio-container input[type="number"],
|
| 832 |
+
.gradio-container .gr-number {
|
| 833 |
+
background: rgba(10, 15, 40, 0.6) !important;
|
| 834 |
+
border: 1px solid rgba(100, 180, 255, 0.15) !important;
|
| 835 |
+
border-radius: 8px !important;
|
| 836 |
+
color: #e8ecf4 !important;
|
| 837 |
+
box-shadow: inset 0 2px 6px rgba(0, 0, 0, 0.2) !important;
|
| 838 |
+
}
|
| 839 |
+
|
| 840 |
+
/* ── DROPDOWN / SELECT ── */
|
| 841 |
+
.gradio-container select,
|
| 842 |
+
.gradio-container .gr-dropdown,
|
| 843 |
+
.gradio-container .multiselect {
|
| 844 |
+
background: rgba(10, 15, 40, 0.7) !important;
|
| 845 |
+
border: 1px solid rgba(100, 180, 255, 0.15) !important;
|
| 846 |
+
border-radius: 10px !important;
|
| 847 |
+
color: #e8ecf4 !important;
|
| 848 |
+
box-shadow: inset 0 2px 6px rgba(0, 0, 0, 0.2) !important;
|
| 849 |
+
transition: border-color 0.3s ease, box-shadow 0.3s ease !important;
|
| 850 |
+
}
|
| 851 |
+
.gradio-container select:focus,
|
| 852 |
+
.gradio-container .gr-dropdown:focus {
|
| 853 |
+
border-color: rgba(108, 92, 231, 0.6) !important;
|
| 854 |
+
box-shadow: 0 0 12px rgba(108, 92, 231, 0.2) !important;
|
| 855 |
+
}
|
| 856 |
+
|
| 857 |
+
/* ── CHECKBOXES ── */
|
| 858 |
+
.gradio-container input[type="checkbox"] {
|
| 859 |
+
accent-color: #00d4aa !important;
|
| 860 |
+
width: 18px !important;
|
| 861 |
+
height: 18px !important;
|
| 862 |
+
cursor: pointer !important;
|
| 863 |
+
}
|
| 864 |
+
.gradio-container input[type="checkbox"]:checked {
|
| 865 |
+
accent-color: #00e5c8 !important;
|
| 866 |
+
}
|
| 867 |
+
|
| 868 |
+
/* ── ACCORDION ── */
|
| 869 |
+
.gradio-container .accordion,
|
| 870 |
+
.gradio-container details {
|
| 871 |
+
background: rgba(255, 255, 255, 0.03) !important;
|
| 872 |
+
border: 1px solid rgba(100, 180, 255, 0.1) !important;
|
| 873 |
+
border-radius: 12px !important;
|
| 874 |
+
overflow: hidden !important;
|
| 875 |
+
}
|
| 876 |
+
.gradio-container .accordion summary,
|
| 877 |
+
.gradio-container details summary {
|
| 878 |
+
color: #c8d8e8 !important;
|
| 879 |
+
font-weight: 600 !important;
|
| 880 |
+
padding: 12px 16px !important;
|
| 881 |
+
cursor: pointer !important;
|
| 882 |
+
transition: color 0.2s ease !important;
|
| 883 |
+
}
|
| 884 |
+
.gradio-container .accordion summary:hover,
|
| 885 |
+
.gradio-container details summary:hover {
|
| 886 |
+
color: #00e5c8 !important;
|
| 887 |
+
}
|
| 888 |
+
|
| 889 |
+
/* ── GENERATE BUTTON — vibrant gradient with glow ── */
|
| 890 |
+
.gradio-container button.primary,
|
| 891 |
+
.gradio-container button.lg.primary,
|
| 892 |
+
.gradio-container .gr-button-primary,
|
| 893 |
+
.gradio-container button[data-variant="primary"] {
|
| 894 |
+
background: linear-gradient(135deg, #00b894, #0984e3, #6c5ce7) !important;
|
| 895 |
+
background-size: 200% 200% !important;
|
| 896 |
+
animation: btn-shimmer 4s ease-in-out infinite !important;
|
| 897 |
+
color: #ffffff !important;
|
| 898 |
+
font-weight: 700 !important;
|
| 899 |
+
font-size: 1.05em !important;
|
| 900 |
+
letter-spacing: 0.03em !important;
|
| 901 |
+
border: none !important;
|
| 902 |
+
border-radius: 12px !important;
|
| 903 |
+
padding: 14px 32px !important;
|
| 904 |
+
box-shadow:
|
| 905 |
+
0 4px 20px rgba(0, 184, 148, 0.35),
|
| 906 |
+
0 0 40px rgba(108, 92, 231, 0.15),
|
| 907 |
+
inset 0 1px 0 rgba(255, 255, 255, 0.15) !important;
|
| 908 |
+
transition: transform 0.25s ease, box-shadow 0.3s ease !important;
|
| 909 |
+
text-shadow: 0 1px 3px rgba(0, 0, 0, 0.3) !important;
|
| 910 |
+
cursor: pointer !important;
|
| 911 |
+
}
|
| 912 |
+
.gradio-container button.primary:hover,
|
| 913 |
+
.gradio-container button.lg.primary:hover,
|
| 914 |
+
.gradio-container .gr-button-primary:hover,
|
| 915 |
+
.gradio-container button[data-variant="primary"]:hover {
|
| 916 |
+
transform: translateY(-2px) !important;
|
| 917 |
+
box-shadow:
|
| 918 |
+
0 8px 32px rgba(0, 184, 148, 0.5),
|
| 919 |
+
0 0 60px rgba(108, 92, 231, 0.25),
|
| 920 |
+
inset 0 1px 0 rgba(255, 255, 255, 0.2) !important;
|
| 921 |
+
}
|
| 922 |
+
.gradio-container button.primary:active,
|
| 923 |
+
.gradio-container button.lg.primary:active {
|
| 924 |
+
transform: translateY(0px) !important;
|
| 925 |
+
box-shadow:
|
| 926 |
+
0 2px 12px rgba(0, 184, 148, 0.3),
|
| 927 |
+
inset 0 2px 4px rgba(0, 0, 0, 0.2) !important;
|
| 928 |
+
}
|
| 929 |
+
|
| 930 |
+
@keyframes btn-shimmer {
|
| 931 |
+
0%, 100% { background-position: 0% 50%; }
|
| 932 |
+
50% { background-position: 100% 50%; }
|
| 933 |
+
}
|
| 934 |
+
|
| 935 |
+
/* ── SECONDARY BUTTONS ── */
|
| 936 |
+
.gradio-container button.secondary,
|
| 937 |
+
.gradio-container button.lg.secondary,
|
| 938 |
+
.gradio-container .gr-button-secondary,
|
| 939 |
+
.gradio-container button[data-variant="secondary"] {
|
| 940 |
+
background: rgba(255, 255, 255, 0.06) !important;
|
| 941 |
+
backdrop-filter: blur(8px) !important;
|
| 942 |
+
border: 1px solid rgba(100, 180, 255, 0.2) !important;
|
| 943 |
+
color: #b8c8e0 !important;
|
| 944 |
+
border-radius: 10px !important;
|
| 945 |
+
font-weight: 600 !important;
|
| 946 |
+
transition: all 0.25s ease !important;
|
| 947 |
+
}
|
| 948 |
+
.gradio-container button.secondary:hover,
|
| 949 |
+
.gradio-container button.lg.secondary:hover,
|
| 950 |
+
.gradio-container .gr-button-secondary:hover,
|
| 951 |
+
.gradio-container button[data-variant="secondary"]:hover {
|
| 952 |
+
background: rgba(108, 92, 231, 0.15) !important;
|
| 953 |
+
border-color: rgba(108, 92, 231, 0.4) !important;
|
| 954 |
+
color: #e0e8f8 !important;
|
| 955 |
+
box-shadow: 0 0 20px rgba(108, 92, 231, 0.15) !important;
|
| 956 |
+
}
|
| 957 |
+
|
| 958 |
+
/* ── FILE OUTPUT ── */
|
| 959 |
+
.gradio-container .file-preview,
|
| 960 |
+
.gradio-container [data-testid="file"] {
|
| 961 |
+
background: rgba(10, 15, 40, 0.5) !important;
|
| 962 |
+
border: 1px solid rgba(100, 180, 255, 0.1) !important;
|
| 963 |
+
border-radius: 10px !important;
|
| 964 |
+
}
|
| 965 |
+
|
| 966 |
+
/* ── INFO TEXT ── */
|
| 967 |
+
.gradio-container .prose small,
|
| 968 |
+
.gradio-container small,
|
| 969 |
+
.gradio-container .gr-info {
|
| 970 |
+
color: #7a8fa8 !important;
|
| 971 |
+
}
|
| 972 |
+
|
| 973 |
+
/* ── SCROLLBAR — themed ── */
|
| 974 |
+
.gradio-container ::-webkit-scrollbar {
|
| 975 |
+
width: 8px !important;
|
| 976 |
+
height: 8px !important;
|
| 977 |
+
}
|
| 978 |
+
.gradio-container ::-webkit-scrollbar-track {
|
| 979 |
+
background: rgba(10, 15, 40, 0.4) !important;
|
| 980 |
+
border-radius: 4px !important;
|
| 981 |
+
}
|
| 982 |
+
.gradio-container ::-webkit-scrollbar-thumb {
|
| 983 |
+
background: linear-gradient(180deg, #00b894, #6c5ce7) !important;
|
| 984 |
+
border-radius: 4px !important;
|
| 985 |
+
}
|
| 986 |
+
.gradio-container ::-webkit-scrollbar-thumb:hover {
|
| 987 |
+
background: linear-gradient(180deg, #00d4aa, #a855f7) !important;
|
| 988 |
+
}
|
| 989 |
+
|
| 990 |
+
/* ── AUTO-DETECTED LORA PREVIEW ── */
|
| 991 |
+
#auto-lora-preview {
|
| 992 |
+
color: #b8d4e8 !important;
|
| 993 |
+
}
|
| 994 |
+
#auto-lora-preview b {
|
| 995 |
+
color: #00e5c8 !important;
|
| 996 |
+
}
|
| 997 |
+
|
| 998 |
+
/* ── COLUMN SEPARATOR — subtle glow line ── */
|
| 999 |
+
.gradio-container .gr-row > .gr-col,
|
| 1000 |
+
.gradio-container .gr-row > div {
|
| 1001 |
+
border-left: 1px solid rgba(108, 92, 231, 0.08) !important;
|
| 1002 |
+
}
|
| 1003 |
+
"""
|
| 1004 |
+
|
| 1005 |
+
|
| 1006 |
+
with gr.Blocks(delete_cache=(3600, 10800)) as demo:
|
| 1007 |
+
gr.Markdown(model_title())
|
| 1008 |
+
gr.Markdown("Run Wan 2.2 in just 4-8 steps, fp8 quantization & AoT compilation - compatible with 🧨 diffusers and ZeroGPU")
|
| 1009 |
+
|
| 1010 |
+
with gr.Row():
|
| 1011 |
+
with gr.Column():
|
| 1012 |
+
input_image_component = gr.Image(type="pil", label="Input Image", sources=["upload", "clipboard"])
|
| 1013 |
+
prompt_input = gr.Textbox(label="Prompt", value=default_prompt_i2v)
|
| 1014 |
+
auto_lora_checkbox = gr.Checkbox(
|
| 1015 |
+
label="Auto-activate LoRAs from prompt",
|
| 1016 |
+
value=True,
|
| 1017 |
+
interactive=True,
|
| 1018 |
+
info="Automatically detect matching LoRAs from your prompt text (deduplicated by concept).",
|
| 1019 |
+
)
|
| 1020 |
+
auto_lora_preview = gr.HTML(
|
| 1021 |
+
label="Auto-detected LoRAs",
|
| 1022 |
+
value="",
|
| 1023 |
+
elem_id="auto-lora-preview",
|
| 1024 |
+
)
|
| 1025 |
+
duration_seconds_input = gr.Slider(minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.1, value=3.5, label="Duration (seconds)", info=f"Clamped to model's {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps.")
|
| 1026 |
+
frame_multi = gr.Dropdown(
|
| 1027 |
+
choices=[FIXED_FPS, FIXED_FPS*2, FIXED_FPS*4, FIXED_FPS*8],
|
| 1028 |
+
value=FIXED_FPS,
|
| 1029 |
+
label="Video Fluidity (Frames per Second)",
|
| 1030 |
+
info="Extra frames will be generated using flow estimation, which estimates motion between frames to make the video smoother."
|
| 1031 |
+
)
|
| 1032 |
+
safe_mode_checkbox = gr.Checkbox(
|
| 1033 |
+
label="🛠️ Safe Mode",
|
| 1034 |
+
value=True,
|
| 1035 |
+
info="Requests 30% extra processing time to try to prevent unfinished tasks when the server is busy."
|
| 1036 |
+
)
|
| 1037 |
+
with gr.Accordion("Advanced Settings", open=False):
|
| 1038 |
+
last_image_component = gr.Image(type="pil", label="Last Image (Optional)", sources=["upload", "clipboard"])
|
| 1039 |
+
negative_prompt_input = gr.Textbox(label="Negative Prompt", value=default_negative_prompt, info="Used if any Guidance Scale > 1.", lines=3)
|
| 1040 |
+
quality_slider = gr.Slider(minimum=1, maximum=10, step=1, value=6, label="Video Quality", info="If set to 10, the generated video may be too large and won't play in the Gradio preview.")
|
| 1041 |
+
seed_input = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42, interactive=True)
|
| 1042 |
+
randomize_seed_checkbox = gr.Checkbox(label="Randomize seed", value=True, interactive=True)
|
| 1043 |
+
steps_slider = gr.Slider(minimum=1, maximum=30, step=1, value=6, label="Inference Steps")
|
| 1044 |
+
guidance_scale_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=1, label="Guidance Scale - high noise stage", info="Values above 1 increase GPU usage and may take longer to process.")
|
| 1045 |
+
guidance_scale_2_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=1, label="Guidance Scale 2 - low noise stage")
|
| 1046 |
+
scheduler_dropdown = gr.Dropdown(
|
| 1047 |
+
label="Scheduler",
|
| 1048 |
+
choices=list(SCHEDULER_MAP.keys()),
|
| 1049 |
+
value="UniPCMultistep",
|
| 1050 |
+
info="Select a custom scheduler."
|
| 1051 |
+
)
|
| 1052 |
+
flow_shift_slider = gr.Slider(minimum=0.5, maximum=15.0, step=0.1, value=3.0, label="Flow Shift")
|
| 1053 |
+
lora_dropdown = gr.Dropdown(choices=lora_loader.get_lora_choices(), label="LoRA (NSFW)", multiselect=True, info="Select scenario LoRAs")
|
| 1054 |
+
play_result_video = gr.Checkbox(label="Display result", value=True, interactive=True)
|
| 1055 |
+
|
| 1056 |
+
generate_button = gr.Button("Generate Video", variant="primary")
|
| 1057 |
+
|
| 1058 |
+
with gr.Column():
|
| 1059 |
+
# ASSIGNED elem_id="generated-video" so JS can find it
|
| 1060 |
+
video_output = gr.Video(label="Generated Video", autoplay=True, sources=["upload"], buttons=["download", "share"], interactive=True, elem_id="generated-video")
|
| 1061 |
+
|
| 1062 |
+
# --- Frame Grabbing UI ---
|
| 1063 |
+
with gr.Row():
|
| 1064 |
+
grab_frame_btn = gr.Button("📸 Use Current Frame as Input", variant="secondary")
|
| 1065 |
+
timestamp_box = gr.Number(value=0, label="Timestamp", visible=True, elem_id="hidden-timestamp")
|
| 1066 |
+
# -------------------------
|
| 1067 |
+
|
| 1068 |
+
file_output = gr.File(label="Download Video")
|
| 1069 |
+
|
| 1070 |
+
ui_inputs = [
|
| 1071 |
+
input_image_component, last_image_component, prompt_input, steps_slider,
|
| 1072 |
+
negative_prompt_input, duration_seconds_input,
|
| 1073 |
+
guidance_scale_input, guidance_scale_2_input, seed_input, randomize_seed_checkbox,
|
| 1074 |
+
quality_slider, scheduler_dropdown, flow_shift_slider, frame_multi,
|
| 1075 |
+
safe_mode_checkbox,
|
| 1076 |
+
lora_dropdown,
|
| 1077 |
+
auto_lora_checkbox,
|
| 1078 |
+
play_result_video
|
| 1079 |
+
]
|
| 1080 |
+
|
| 1081 |
+
generate_button.click(
|
| 1082 |
+
fn=generate_video,
|
| 1083 |
+
inputs=ui_inputs,
|
| 1084 |
+
outputs=[video_output, file_output, seed_input]
|
| 1085 |
+
)
|
| 1086 |
+
|
| 1087 |
+
# --- Auto-LoRA Preview Events ---
|
| 1088 |
+
prompt_input.change(
|
| 1089 |
+
fn=_update_auto_lora_preview,
|
| 1090 |
+
inputs=[prompt_input, auto_lora_checkbox],
|
| 1091 |
+
outputs=[auto_lora_preview],
|
| 1092 |
+
)
|
| 1093 |
+
auto_lora_checkbox.change(
|
| 1094 |
+
fn=_update_auto_lora_preview,
|
| 1095 |
+
inputs=[prompt_input, auto_lora_checkbox],
|
| 1096 |
+
outputs=[auto_lora_preview],
|
| 1097 |
+
)
|
| 1098 |
+
|
| 1099 |
+
# --- Frame Grabbing Events ---
|
| 1100 |
+
# 1. Click button -> JS runs -> puts time in hidden number box
|
| 1101 |
+
grab_frame_btn.click(
|
| 1102 |
+
fn=None,
|
| 1103 |
+
inputs=None,
|
| 1104 |
+
outputs=[timestamp_box],
|
| 1105 |
+
js=get_timestamp_js
|
| 1106 |
+
)
|
| 1107 |
+
|
| 1108 |
+
# 2. Hidden number box changes -> Python runs -> puts frame in Input Image
|
| 1109 |
+
timestamp_box.change(
|
| 1110 |
+
fn=extract_frame,
|
| 1111 |
+
inputs=[video_output, timestamp_box],
|
| 1112 |
+
outputs=[input_image_component]
|
| 1113 |
+
)
|
| 1114 |
+
|
| 1115 |
+
if __name__ == "__main__":
|
| 1116 |
+
demo.queue().launch(
|
| 1117 |
+
mcp_server=True,
|
| 1118 |
+
css=CSS,
|
| 1119 |
+
show_error=True,
|
| 1120 |
)
|