import os os.environ["TOKENIZERS_PARALLELISM"] = "false" # ── spaces MUST be imported before any CUDA-related package ─────────────────── import spaces # ── Auto-upgrade torchao if version is too old (requires >= 0.16.0) ──────────── import subprocess, sys, re, pathlib, site as _site_mod # ── Patch 1: Fix diffusers torchao_quantizer.py — 'logger' not defined bug ─── try: _site_pkgs = pathlib.Path(_site_mod.getsitepackages()[0]) _quantizer_path = _site_pkgs / "diffusers/quantizers/torchao/torchao_quantizer.py" if _quantizer_path.exists(): _src = _quantizer_path.read_text(encoding="utf-8") if "logger = logging.getLogger" not in _src and "logger.warning" in _src: if "import logging" in _src: _src = _src.replace( "import logging\n", "import logging\nlogger = logging.getLogger(__name__) # patch: auto-injected\n", 1, ) else: _src = "import logging\nlogger = logging.getLogger(__name__) # patch: auto-injected\n" + _src _quantizer_path.write_text(_src, encoding="utf-8") print("[Patch] ✅ diffusers torchao_quantizer.py — logger injected successfully") else: print("[Patch] ✓ torchao_quantizer.py already has logger — no patch needed") else: print("[Patch] ⚠️ torchao_quantizer.py not found — skipping patch") except Exception as _pe: print(f"[Patch] ⚠️ Could not patch torchao_quantizer.py: {_pe}") # ── Patch 2: Ensure torchao >= 0.16.0 (required by diffusers Float8 quantizer) ─ try: import torchao as _tao from packaging.version import Version if Version(_tao.__version__) < Version("0.16.0"): print(f"[torchao] Found v{_tao.__version__} < 0.16.0 — upgrading in background…") subprocess.check_call([ sys.executable, "-m", "pip", "install", "-q", "--upgrade", "torchao>=0.16.0" ]) print("[torchao] ✅ Upgrade complete — will take effect on next restart") else: print(f"[torchao] ✅ Version OK: {_tao.__version__}") except Exception as _e: print(f"[torchao] version check failed: {_e}") import torch from diffusers.pipelines.wan.pipeline_wan_i2v import WanImageToVideoPipeline from diffusers.models.transformers.transformer_wan import WanTransformer3DModel from diffusers.utils.export_utils import export_to_video import gradio as gr import tempfile import numpy as np from PIL import Image import random import gc import json from datetime import datetime from torchao.quantization import quantize_ from torchao.quantization import Float8DynamicActivationFloat8WeightConfig from torchao.quantization import Int8WeightOnlyConfig import aoti # ── Anthropic for Prompt Analysis (Colorization / Music) ───────────────────── try: import anthropic _anthropic_key = os.environ.get("ANTHROPIC_API_KEY", "") ANTHROPIC_AVAILABLE = bool(_anthropic_key) if not ANTHROPIC_AVAILABLE: print("[Anthropic] ⚠️ ANTHROPIC_API_KEY not found in Secrets — Claude disabled") else: print("[Anthropic] ✅ API key found") except ImportError: ANTHROPIC_AVAILABLE = False print("[Anthropic] ⚠️ Library not installed") # ── MoviePy for audio mixing ─────────────────────────────────────────────────── try: from moviepy.editor import VideoFileClip, AudioFileClip MOVIEPY_AVAILABLE = True except ImportError: MOVIEPY_AVAILABLE = False print("[Music] moviepy not installed — audio features disabled") # ── MusicGen (Meta) ──────────────────────────────────────────────────────────── MUSICGEN_AVAILABLE = False _musicgen_pipe = None MUSICGEN_MODEL = "facebook/musicgen-small" MUSICGEN_SAMPLE_RATE = 32000 try: from transformers import pipeline as hf_pipeline import scipy.io.wavfile as _wavfile MUSICGEN_AVAILABLE = True print("[MusicGen] transformers available ✅") except ImportError: print("[MusicGen] transformers not installed — music generation disabled") def _load_musicgen(): global _musicgen_pipe if _musicgen_pipe is not None: return True if not MUSICGEN_AVAILABLE: return False try: print("[MusicGen] Loading model…") _musicgen_pipe = hf_pipeline( "text-to-audio", model=MUSICGEN_MODEL, device="cpu", torch_dtype=torch.float32, ) print("[MusicGen] ✅ Model ready") return True except Exception as e: print(f"[MusicGen] Load failed: {e}") return False def _prompt_to_music_description(video_prompt: str) -> str: if ANTHROPIC_AVAILABLE: try: client = anthropic.Anthropic() msg = client.messages.create( model="claude-opus-4-6", max_tokens=80, messages=[{ "role": "user", "content": ( "You are a professional music composer for films.\n" "Convert this video scene description into a short music generation prompt " "(max 20 words). Focus on: genre, instruments, tempo, mood.\n" "Examples:\n" "- 'cinematic orchestral, slow tempo, dramatic strings, epic brass'\n" "- 'ambient electronic, soft pads, 80 bpm, peaceful and dreamy'\n" "- 'upbeat jazz, piano and drums, 120 bpm, energetic and fun'\n\n" f"Video description: {video_prompt}\n\n" "Return ONLY the music prompt, nothing else." ), }], ) result = msg.content[0].text.strip() print(f"[MusicGen] Music description from Claude: {result}") return result except Exception as e: print(f"[MusicGen] Claude API error: {e}") prompt_lower = video_prompt.lower() if any(w in prompt_lower for w in ["action", "fast", "fight", "run", "dynamic"]): return "fast paced action music, energetic drums, 140 bpm, intense" if any(w in prompt_lower for w in ["nature", "forest", "ocean", "calm", "peaceful"]): return "peaceful ambient nature sounds, soft flute, 70 bpm, serene" if any(w in prompt_lower for w in ["magic", "fantasy", "dream", "ethereal", "sparkle"]): return "magical fantasy music, harp and strings, 90 bpm, whimsical" if any(w in prompt_lower for w in ["city", "urban", "night", "street"]): return "urban electronic music, synthesizer, 120 bpm, modern and cool" if any(w in prompt_lower for w in ["sad", "emotion", "cry", "nostalgic"]): return "emotional piano solo, slow tempo, 60 bpm, melancholic and tender" return "cinematic orchestral music, dramatic strings, 100 bpm, epic and powerful" @spaces.GPU(duration=60) def generate_music_for_video( video_path: str, video_prompt: str, music_duration: float, fade_in: float, fade_out: float, volume: float, progress=gr.Progress(track_tqdm=True), ) -> tuple: if video_path is None: return None, "", "⚠️ No video found — generate a video first." if not MOVIEPY_AVAILABLE: return video_path, "", "⚠️ moviepy is not installed." progress(0.05, desc="🤖 Claude is writing the music description…") music_prompt = _prompt_to_music_description(video_prompt) progress(0.15, desc="🎵 Loading MusicGen…") if not _load_musicgen(): return video_path, music_prompt, "❌ Failed to load MusicGen." try: tokens_per_sec = 50 max_tokens = max(128, min(1500, int(music_duration * tokens_per_sec))) progress(0.25, desc=f"🎼 Generating music: «{music_prompt[:50]}»…") audio_out = _musicgen_pipe( music_prompt, forward_params={ "do_sample": True, "max_new_tokens": max_tokens, }, ) audio_array = audio_out["audio"][0] sample_rate = audio_out["sampling_rate"] if audio_array.ndim == 2: audio_array = audio_array.mean(axis=0) audio_array = (audio_array / (np.abs(audio_array).max() + 1e-8) * 32767).astype(np.int16) with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_wav: wav_path = tmp_wav.name _wavfile.write(wav_path, sample_rate, audio_array) progress(0.75, desc="🎬 Merging music with video…") video_clip = VideoFileClip(video_path) audio_clip = AudioFileClip(wav_path) audio_clip = audio_clip.subclip(0, min(audio_clip.duration, video_clip.duration)) if fade_in > 0: audio_clip = audio_clip.audio_fadein(fade_in) if fade_out > 0: audio_clip = audio_clip.audio_fadeout(fade_out) audio_clip = audio_clip.volumex(volume) final_video = video_clip.set_audio(audio_clip) with tempfile.NamedTemporaryFile(suffix="_musicgen.mp4", delete=False) as tmp_out: output_path = tmp_out.name final_video.write_videofile( output_path, codec="libx264", audio_codec="aac", temp_audiofile=output_path + "_tmp.m4a", remove_temp=True, verbose=False, logger=None, ) video_clip.close() audio_clip.close() final_video.close() os.unlink(wav_path) progress(1.0, desc="✅ Done!") status = ( f"✅ **AI music generated and merged successfully!**\n\n" f"🎼 **Music description:** {music_prompt}\n\n" f"⏱ Duration: {music_duration:.1f}s | 🔊 Volume: {int(volume*100)}% | " f"Fade In: {fade_in}s | Fade Out: {fade_out}s" ) return output_path, music_prompt, status except Exception as e: print(f"[MusicGen] Error: {e}") import traceback; traceback.print_exc() return video_path, music_prompt, f"❌ MusicGen error: {str(e)[:120]}" # ── Constants & Core Settings ───────────────────────────────────────────────── MODEL_ID = "Wan-AI/Wan2.2-I2V-A14B-Diffusers" SQUARE_DIM = 480 MAX_DIM = 832 MIN_DIM = 480 MULTIPLE_OF = 16 FIXED_FPS = 16 MIN_FRAMES_MODEL = 1 MAX_FRAMES_MODEL = 128 MIN_DURATION = 1.0 MAX_DURATION = 8.0 MAX_SEED = 2**32 - 1 HISTORY_FILE = "generation_history.json" default_prompt_i2v = "make this image come alive with cinematic motion" default_negative_prompt = ( "static, blurry, low quality, watermark, text, deformed, ugly" ) # ── 🎵 Built-in Music Library ───────────────────────────────────────────────── MUSIC_LIBRARY = { "🎬 Cinematic Epic": { "url": "https://www.soundhelix.com/examples/mp3/SoundHelix-Song-1.mp3", "mood": "dramatic, powerful, orchestral", "bpm": 120, }, "🌊 Ambient Flow": { "url": "https://www.soundhelix.com/examples/mp3/SoundHelix-Song-2.mp3", "mood": "calm, flowing, atmospheric", "bpm": 80, }, "⚡ Action Drive": { "url": "https://www.soundhelix.com/examples/mp3/SoundHelix-Song-3.mp3", "mood": "energetic, fast, intense", "bpm": 140, }, "🌿 Nature Serenity": { "url": "https://www.soundhelix.com/examples/mp3/SoundHelix-Song-4.mp3", "mood": "peaceful, soft, organic", "bpm": 70, }, "✨ Magical Wonder": { "url": "https://www.soundhelix.com/examples/mp3/SoundHelix-Song-5.mp3", "mood": "dreamy, magical, ethereal", "bpm": 95, }, "🎭 Dramatic Tension": { "url": "https://www.soundhelix.com/examples/mp3/SoundHelix-Song-6.mp3", "mood": "tense, suspenseful, dark", "bpm": 110, }, "🌅 Sunrise Journey": { "url": "https://www.soundhelix.com/examples/mp3/SoundHelix-Song-7.mp3", "mood": "uplifting, hopeful, warm", "bpm": 100, }, "🎸 Indie Vibes": { "url": "https://www.soundhelix.com/examples/mp3/SoundHelix-Song-8.mp3", "mood": "cool, modern, casual", "bpm": 125, }, "🎹 Piano Emotion": { "url": "https://www.soundhelix.com/examples/mp3/SoundHelix-Song-9.mp3", "mood": "emotional, intimate, classical", "bpm": 85, }, "🌃 Night City": { "url": "https://www.soundhelix.com/examples/mp3/SoundHelix-Song-10.mp3", "mood": "urban, electronic, night", "bpm": 130, }, "🚀 Futuristic": { "url": "https://www.soundhelix.com/examples/mp3/SoundHelix-Song-11.mp3", "mood": "sci-fi, electronic, futuristic", "bpm": 135, }, "🏔️ Epic Adventure": { "url": "https://www.soundhelix.com/examples/mp3/SoundHelix-Song-12.mp3", "mood": "heroic, adventurous, grand", "bpm": 115, }, } PRESET_TO_MUSIC = { "🌊 Flowing": "🌊 Ambient Flow", "🎥 Cinematic": "🎬 Cinematic Epic", "💨 Dynamic": "⚡ Action Drive", "🌿 Nature": "🌿 Nature Serenity", "✨ Magical": "✨ Magical Wonder", "🏃 Action": "⚡ Action Drive", "🌅 Timelapse": "🌅 Sunrise Journey", "🎭 Dramatic": "🎭 Dramatic Tension", } MOTION_PRESETS = { "🌊 Flowing": "gentle flowing motion, smooth camera drift", "🎥 Cinematic": "cinematic dolly shot, dramatic lighting, film grain", "💨 Dynamic": "fast dynamic motion, energetic movement", "🌿 Nature": "gentle breeze, leaves rustling, natural ambient motion", "✨ Magical": "magical sparkles, ethereal glow, fantastical transformation", "🏃 Action": "fast paced action, rapid movement, high energy", "🌅 Timelapse": "time lapse clouds, light changing, atmospheric mood", "🎭 Dramatic": "dramatic reveal, slow zoom in, tense atmosphere", } # ── Colorization support ────────────────────────────────────────────────────── import cv2, urllib.request as _urlreq OPENCV_COLOR_AVAILABLE = False _opencv_net = None _COLOR_CACHE = os.path.join(os.path.expanduser("~"), ".cache", "cv_colorizer") _MUSIC_CACHE = os.path.join(os.path.expanduser("~"), ".cache", "music_library") os.makedirs(_COLOR_CACHE, exist_ok=True) os.makedirs(_MUSIC_CACHE, exist_ok=True) _PROTOTXT_CONTENT = """ name: "colorization" layer { name: "data_l" type: "Input" top: "data_l" input_param { shape { dim: 1 dim: 1 dim: 224 dim: 224 } } } layer { name: "conv1_1" type: "Convolution" bottom: "data_l" top: "conv1_1" convolution_param { num_output: 64 kernel_size: 3 stride: 1 pad: 1 } } layer { name: "relu1_1" type: "ReLU" bottom: "conv1_1" top: "conv1_1" } layer { name: "conv1_2" type: "Convolution" bottom: "conv1_1" top: "conv1_2" convolution_param { num_output: 64 kernel_size: 3 stride: 1 pad: 1 } } layer { name: "relu1_2" type: "ReLU" bottom: "conv1_2" top: "conv1_2" } layer { name: "conv1_2norm" type: "BatchNorm" bottom: "conv1_2" top: "conv1_2norm" } layer { name: "pool1" type: "Pooling" bottom: "conv1_2norm" top: "pool1" pooling_param { pool: MAX kernel_size: 2 stride: 2 } } layer { name: "conv2_1" type: "Convolution" bottom: "pool1" top: "conv2_1" convolution_param { num_output: 128 kernel_size: 3 stride: 1 pad: 1 } } layer { name: "relu2_1" type: "ReLU" bottom: "conv2_1" top: "conv2_1" } layer { name: "conv2_2" type: "Convolution" bottom: "conv2_1" top: "conv2_2" convolution_param { num_output: 128 kernel_size: 3 stride: 1 pad: 1 } } layer { name: "relu2_2" type: "ReLU" bottom: "conv2_2" top: "conv2_2" } layer { name: "conv2_2norm" type: "BatchNorm" bottom: "conv2_2" top: "conv2_2norm" } layer { name: "pool2" type: "Pooling" bottom: "conv2_2norm" top: "pool2" pooling_param { pool: MAX kernel_size: 2 stride: 2 } } layer { name: "conv3_1" type: "Convolution" bottom: "pool2" top: "conv3_1" convolution_param { num_output: 256 kernel_size: 3 stride: 1 pad: 1 } } layer { name: "relu3_1" type: "ReLU" bottom: "conv3_1" top: "conv3_1" } layer { name: "conv3_2" type: "Convolution" bottom: "conv3_1" top: "conv3_2" convolution_param { num_output: 256 kernel_size: 3 stride: 1 pad: 1 } } layer { name: "relu3_2" type: "ReLU" bottom: "conv3_2" top: "conv3_2" } layer { name: "conv3_3" type: "Convolution" bottom: "conv3_2" top: "conv3_3" convolution_param { num_output: 256 kernel_size: 3 stride: 1 pad: 1 } } layer { name: "relu3_3" type: "ReLU" bottom: "conv3_3" top: "conv3_3" } layer { name: "conv3_3norm" type: "BatchNorm" bottom: "conv3_3" top: "conv3_3norm" } layer { name: "pool3" type: "Pooling" bottom: "conv3_3norm" top: "pool3" pooling_param { pool: MAX kernel_size: 2 stride: 2 } } layer { name: "conv4_1" type: "Convolution" bottom: "pool3" top: "conv4_1" convolution_param { num_output: 512 kernel_size: 3 stride: 1 pad: 1 dilation: 1 } } layer { name: "relu4_1" type: "ReLU" bottom: "conv4_1" top: "conv4_1" } layer { name: "conv4_2" type: "Convolution" bottom: "conv4_1" top: "conv4_2" convolution_param { num_output: 512 kernel_size: 3 stride: 1 pad: 1 dilation: 1 } } layer { name: "relu4_2" type: "ReLU" bottom: "conv4_2" top: "conv4_2" } layer { name: "conv4_3" type: "Convolution" bottom: "conv4_1" top: "conv4_3" convolution_param { num_output: 512 kernel_size: 3 stride: 1 pad: 1 dilation: 1 } } layer { name: "relu4_3" type: "ReLU" bottom: "conv4_3" top: "conv4_3" } layer { name: "conv4_3norm" type: "BatchNorm" bottom: "conv4_3" top: "conv4_3norm" } layer { name: "conv5_1" type: "Convolution" bottom: "conv4_3norm" top: "conv5_1" convolution_param { num_output: 512 kernel_size: 3 stride: 1 pad: 2 dilation: 2 } } layer { name: "relu5_1" type: "ReLU" bottom: "conv5_1" top: "conv5_1" } layer { name: "conv5_2" type: "Convolution" bottom: "conv5_1" top: "conv5_2" convolution_param { num_output: 512 kernel_size: 3 stride: 1 pad: 2 dilation: 2 } } layer { name: "relu5_2" type: "ReLU" bottom: "conv5_2" top: "conv5_2" } layer { name: "conv5_3" type: "Convolution" bottom: "conv5_2" top: "conv5_3" convolution_param { num_output: 512 kernel_size: 3 stride: 1 pad: 2 dilation: 2 } } layer { name: "relu5_3" type: "ReLU" bottom: "conv5_3" top: "conv5_3" } layer { name: "conv5_3norm" type: "BatchNorm" bottom: "conv5_3" top: "conv5_3norm" } layer { name: "conv6_1" type: "Convolution" bottom: "conv5_3norm" top: "conv6_1" convolution_param { num_output: 512 kernel_size: 3 stride: 1 pad: 2 dilation: 2 } } layer { name: "relu6_1" type: "ReLU" bottom: "conv6_1" top: "conv6_1" } layer { name: "conv6_2" type: "Convolution" bottom: "conv6_1" top: "conv6_2" convolution_param { num_output: 512 kernel_size: 3 stride: 1 pad: 2 dilation: 2 } } layer { name: "relu6_2" type: "ReLU" bottom: "conv6_2" top: "conv6_2" } layer { name: "conv6_3" type: "Convolution" bottom: "conv6_2" top: "conv6_3" convolution_param { num_output: 512 kernel_size: 3 stride: 1 pad: 2 dilation: 2 } } layer { name: "relu6_3" type: "ReLU" bottom: "conv6_3" top: "conv6_3" } layer { name: "conv6_3norm" type: "BatchNorm" bottom: "conv6_3" top: "conv6_3norm" } layer { name: "conv7_1" type: "Convolution" bottom: "conv6_3norm" top: "conv7_1" convolution_param { num_output: 256 kernel_size: 3 stride: 1 pad: 1 dilation: 1 } } layer { name: "relu7_1" type: "ReLU" bottom: "conv7_1" top: "conv7_1" } layer { name: "conv7_2" type: "Convolution" bottom: "conv7_1" top: "conv7_2" convolution_param { num_output: 256 kernel_size: 3 stride: 1 pad: 1 dilation: 1 } } layer { name: "relu7_2" type: "ReLU" bottom: "conv7_2" top: "conv7_2" } layer { name: "conv7_3" type: "Convolution" bottom: "conv7_2" top: "conv7_3" convolution_param { num_output: 256 kernel_size: 3 stride: 1 pad: 1 dilation: 1 } } layer { name: "relu7_3" type: "ReLU" bottom: "conv7_3" top: "conv7_3" } layer { name: "conv7_3norm" type: "BatchNorm" bottom: "conv7_3" top: "conv7_3norm" } layer { name: "conv8_1" type: "Deconvolution" bottom: "conv7_3norm" top: "conv8_1" convolution_param { num_output: 128 kernel_size: 4 stride: 2 pad: 1 } } layer { name: "relu8_1" type: "ReLU" bottom: "conv8_1" top: "conv8_1" } layer { name: "conv8_2" type: "Convolution" bottom: "conv8_1" top: "conv8_2" convolution_param { num_output: 128 kernel_size: 3 stride: 1 pad: 1 } } layer { name: "relu8_2" type: "ReLU" bottom: "conv8_2" top: "conv8_2" } layer { name: "conv8_3" type: "Convolution" bottom: "conv8_2" top: "conv8_3" convolution_param { num_output: 128 kernel_size: 3 stride: 1 pad: 1 } } layer { name: "relu8_3" type: "ReLU" bottom: "conv8_3" top: "conv8_3" } layer { name: "conv8_3norm" type: "BatchNorm" bottom: "conv8_3" top: "conv8_3norm" } layer { name: "conv8_313" type: "Convolution" bottom: "conv8_3norm" top: "conv8_313" convolution_param { num_output: 313 kernel_size: 1 stride: 1 pad: 0 } } layer { name: "class8_ab" type: "Convolution" bottom: "conv8_313" top: "class8_313_rh" convolution_param { num_output: 2 kernel_size: 1 bias_term: false } } layer { name: "conv8_313_rh" type: "Convolution" bottom: "class8_313_rh" top: "class8_ab" convolution_param { num_output: 2 kernel_size: 1 bias_term: false } } """ def _ensure_opencv_model() -> bool: global _opencv_net, OPENCV_COLOR_AVAILABLE if _opencv_net is not None: return True try: model_path = os.path.join(_COLOR_CACHE, "colorization_release_v2.caffemodel") pts_path = os.path.join(_COLOR_CACHE, "pts_in_hull.npy") proto_path = os.path.join(_COLOR_CACHE, "colorization_deploy_v2.prototxt") if not os.path.exists(proto_path): with open(proto_path, "w") as f: f.write(_PROTOTXT_CONTENT) if not os.path.exists(model_path): print("[Colorizer] Downloading caffemodel (~130 MB)…") _urlreq.urlretrieve( "https://eecs.berkeley.edu/~rich.zhang/projects/2016_colorization/files/demo_v2/colorization_release_v2.caffemodel", model_path, ) if not os.path.exists(pts_path): _urlreq.urlretrieve( "https://github.com/richzhang/colorization/raw/caffe/colorization/resources/pts_in_hull.npy", pts_path, ) net = cv2.dnn.readNetFromCaffe(proto_path, model_path) pts_arr = np.load(pts_path).transpose().reshape(2, 313, 1, 1).astype(np.float32) net.getLayer(net.getLayerId("class8_ab")).blobs = [pts_arr] net.getLayer(net.getLayerId("conv8_313_rh")).blobs = [ np.full([1, 313], 2.606, dtype=np.float32) ] _opencv_net = net OPENCV_COLOR_AVAILABLE = True return True except Exception as exc: print(f"[Colorizer] OpenCV DNN init failed: {exc}") return False def _opencv_colorize(img_pil: Image.Image) -> Image.Image: img_rgb = np.array(img_pil.convert("RGB")).astype(np.float32) / 255.0 img_lab = cv2.cvtColor(img_rgb, cv2.COLOR_RGB2Lab) l_chan = img_lab[:, :, 0] l_rs = cv2.resize(l_chan, (224, 224)) - 50.0 _opencv_net.setInput(cv2.dnn.blobFromImage(l_rs)) ab_dec = _opencv_net.forward()[0, :, :, :].transpose((1, 2, 0)) h, w = img_rgb.shape[:2] ab_up = cv2.resize(ab_dec, (w, h)) lab_out = np.concatenate([l_chan[:, :, np.newaxis], ab_up], axis=2) rgb_out = np.clip(cv2.cvtColor(lab_out.astype(np.float32), cv2.COLOR_Lab2RGB), 0, 1) return Image.fromarray((rgb_out * 255).astype(np.uint8)) import base64, io as _io, json as _json def _pil_to_b64(img: Image.Image, max_px: int = 512) -> str: img = img.convert("RGB") w, h = img.size if max(w, h) > max_px: ratio = max_px / max(w, h) img = img.resize((max(1, int(w*ratio)), max(1, int(h*ratio))), Image.LANCZOS) buf = _io.BytesIO() img.save(buf, format="JPEG", quality=82) return base64.b64encode(buf.getvalue()).decode() def _apply_palette(img_pil: Image.Image, palette: dict) -> Image.Image: img_rgb = np.array(img_pil.convert("RGB")) gray = cv2.cvtColor(img_rgb, cv2.COLOR_RGB2GRAY) h, w = gray.shape gf = gray.astype(np.float32) / 255.0 def gblur(m, k): k = max(3, int(k) | 1) return cv2.GaussianBlur(m.astype(np.float32), (k, k), 0) gray3 = cv2.cvtColor(gray, cv2.COLOR_GRAY2RGB) lab = cv2.cvtColor(gray3.astype(np.float32) / 255.0, cv2.COLOR_RGB2Lab) L = lab[:, :, 0].copy() face_mask = np.zeros((h, w), np.float32) try: det = cv2.CascadeClassifier(cv2.data.haarcascades + "haarcascade_frontalface_default.xml") faces = det.detectMultiScale(gray, scaleFactor=1.1, minNeighbors=4, minSize=(50, 50)) if len(faces) > 0: for (fx, fy, fw, fh) in faces: pad = int(min(fw, fh) * 0.18) face_mask[max(0,fy-pad):min(h,fy+fh+pad), max(0,fx-pad):min(w,fx+fw+pad)] = 1.0 face_mask = gblur(face_mask, min(w, h) // 7) else: cy, cx = h//2, w//2 Y, X = np.ogrid[:h, :w] ell = np.clip(1.0 - np.sqrt(((X-cx)/(w*0.38))**2 + ((Y-cy)/(h*0.48))**2), 0, 1) face_mask = gblur(ell**2, min(w,h)//5) except Exception: pass no_face = np.clip(1.0 - face_mask * 1.4, 0, 1) blur2 = gblur(gf, 5) texture = np.abs(gf - blur2) skin_raw = gblur(((gf > 0.22) & (gf < 0.92)).astype(np.float32) * np.clip(1.0 - texture*10.0, 0, 1), 23) skin_mask = gblur(np.clip(skin_raw * (face_mask*0.75 + 0.25), 0, 1), 29) hair_raw = gblur(np.clip((gf < 0.33).astype(np.float32) + (gf > 0.76).astype(np.float32)*0.55, 0, 1) * no_face, 17) dark_r = gblur((gf < 0.45).astype(np.float32), 9) lap = cv2.Laplacian(gray, cv2.CV_32F) bg_mask = gblur(np.clip(1.0 - np.abs(lap)/10.0, 0, 1) * no_face, 39) * 0.55 lz = np.zeros((h, w), np.float32) lz[int(h*0.60):int(h*0.76), int(w*0.30):int(w*0.70)] = 1.0 lip_mask = gblur(((gf > 0.26) & (gf < 0.70)).astype(np.float32) * lz * face_mask, 9) * 0.45 p = palette A_off = (bg_mask * p["bg_a"] + hair_raw * (dark_r * p["hd_a"] + (1-dark_r) * p["hl_a"]) + skin_mask * p["sk_a"] + lip_mask * p["lp_a"]) B_off = (bg_mask * p["bg_b"] + hair_raw * (dark_r * p["hd_b"] + (1-dark_r) * p["hl_b"]) + skin_mask * p["sk_b"] + lip_mask * p["lp_b"]) lab_out = np.stack([L, np.clip(A_off,-50,50), np.clip(B_off,-50,50)], axis=2).astype(np.float32) rgb_out = np.clip(cv2.cvtColor(lab_out, cv2.COLOR_Lab2RGB), 0, 1) hsv = cv2.cvtColor(rgb_out, cv2.COLOR_RGB2HSV) hsv[:,:,1] = np.clip(hsv[:,:,1] * 1.18, 0, 1) return Image.fromarray((np.clip(cv2.cvtColor(hsv, cv2.COLOR_HSV2RGB),0,1)*255).astype(np.uint8)) def _smart_colorize(img_pil: Image.Image, style: str = "natural") -> Image.Image: STYLES = { "natural": ( 6.0, 13.0, -1.0, -2.0, 3.0, 5.0, 4.0, 15.0, 11.0, 7.0), "cinematic": ( 5.0, 15.0, -3.0, -9.0, 2.0, 4.0, 4.0, 18.0, 13.0, 5.0), "warm": ( 8.0, 17.0, 2.0, 5.0, 4.0, 9.0, 6.0, 20.0, 14.0, 9.0), } sk_a,sk_b, bg_a,bg_b, hd_a,hd_b, hl_a,hl_b, lp_a,lp_b = STYLES.get(style, STYLES["natural"]) return _apply_palette(img_pil, dict(sk_a=sk_a,sk_b=sk_b,bg_a=bg_a,bg_b=bg_b, hd_a=hd_a,hd_b=hd_b,hl_a=hl_a,hl_b=hl_b, lp_a=lp_a,lp_b=lp_b)) def _claude_colorize(img_pil: Image.Image, style: str = "natural") -> tuple: style_desc = { "natural": "realistic true-to-life natural colors", "cinematic": "cinematic film look, warm skin, teal shadows, golden highlights", "warm": "warm golden-hour, amber tones, sun-kissed skin", }.get(style, "realistic natural colors") prompt = ( "You are a professional photo colorization AI.\n" "Analyze this black-and-white image carefully and return ONLY a valid JSON " "object with LAB color offset values for each region.\n" "Style requested: " + style_desc + "\n\n" "Return this exact JSON (numbers only, no comments, no markdown):\n" "{\n" ' "skin_a": ,\n' ' "skin_b": ,\n' ' "hair_a": ,\n' ' "hair_b": ,\n' ' "bg_a": ,\n' ' "bg_b": ,\n' ' "lip_a": ,\n' ' "lip_b": ,\n' ' "description": ""\n' "}\n\n" "LAB reference: A negative=green, A positive=red. B negative=blue, B positive=yellow.\n" "Return ONLY the JSON object." ) try: client = anthropic.Anthropic() b64 = _pil_to_b64(img_pil, max_px=480) msg = client.messages.create( model = "claude-opus-4-6", max_tokens = 300, messages = [{ "role": "user", "content": [ {"type": "image", "source": { "type": "base64", "media_type": "image/jpeg", "data": b64}}, {"type": "text", "text": prompt}, ], }], ) raw = msg.content[0].text.strip().replace("```json","").replace("```","").strip() vals = _json.loads(raw) desc = vals.pop("description", "") def gv(k, d): return float(vals.get(k, d)) palette = dict( sk_a=gv("skin_a",7), sk_b=gv("skin_b",14), bg_a=gv("bg_a",-2), bg_b=gv("bg_b",-4), hd_a=gv("hair_a",3), hd_b=gv("hair_b",6), hl_a=gv("hair_a",4), hl_b=gv("hair_b",15), lp_a=gv("lip_a",12), lp_b=gv("lip_b",8), ) result = _apply_palette(img_pil, palette) note = (" — " + desc) if desc else "" return result, "🤖 **Claude AI Colorized** — " + style_desc + note except Exception as exc: print("[Claude colorize] error:", exc) return None, str(exc) def _numpy_colorize(img_pil: Image.Image) -> Image.Image: return _smart_colorize(img_pil, style="natural") def colorize_image(input_img, style="natural", render_factor=35): if input_img is None: return None, "⚠️ Please upload an image first." if ANTHROPIC_AVAILABLE: result, status = _claude_colorize(input_img, style) if result is not None: return result, status if _ensure_opencv_model(): try: result = _opencv_colorize(input_img) return result, "✅ Colorized with **OpenCV DNN** — Zhang et al. (ECCV 2016)" except Exception as exc: print("[Colorizer] OpenCV DNN failed:", exc) result = _smart_colorize(input_img, style) labels = {"natural": "🌿 Natural", "cinematic": "🎬 Cinematic", "warm": "🌅 Warm"} return result, "✅ Colorized with **Smart Semantic Engine** — " + labels.get(style, style) # ── 🎵 Music Functions ───────────────────────────────────────────────────────── def _get_music_cache_path(track_name: str) -> str: safe_name = ( track_name .replace(" ", "_").replace("/", "_") .replace("🎬","").replace("🌊","").replace("⚡","").replace("🌿","") .replace("✨","").replace("🎭","").replace("🌅","").replace("🎸","") .replace("🎹","").replace("🌃","").replace("🚀","").replace("🏔️","") .strip() ) return os.path.join(_MUSIC_CACHE, f"{safe_name}.mp3") def _download_music(track_name: str) -> str | None: if track_name not in MUSIC_LIBRARY: return None cache_path = _get_music_cache_path(track_name) if os.path.exists(cache_path) and os.path.getsize(cache_path) > 10_000: return cache_path try: url = MUSIC_LIBRARY[track_name]["url"] print(f"[Music] Downloading: {track_name}") _urlreq.urlretrieve(url, cache_path) return cache_path except Exception as e: print(f"[Music] Download failed (network may be disabled): {e}") if os.path.exists(cache_path): os.unlink(cache_path) return None def add_music_to_video( video_path: str, music_source: str, selected_track: str, custom_audio_path: str, volume: float, fade_in: float, fade_out: float, loop_music: bool, ) -> tuple: if not MOVIEPY_AVAILABLE: return video_path, "⚠️ moviepy is not installed — video has no music." if music_source == "none" or video_path is None: return video_path, "🎬 Video without music." audio_file = None if music_source == "library": if not selected_track or selected_track not in MUSIC_LIBRARY: return video_path, "⚠️ Please select a track from the library first." audio_file = _download_music(selected_track) if audio_file is None: return video_path, ( "❌ Failed to download music.\n" "💡 **Solution:** Upload an audio file directly instead of using the library, " "or use the **🤖 MusicGen AI** tab to generate music without internet." ) elif music_source == "upload": if not custom_audio_path: return video_path, "⚠️ Please upload an audio file first." audio_file = custom_audio_path try: video_clip = VideoFileClip(video_path) video_dur = video_clip.duration audio_clip = AudioFileClip(audio_file) if loop_music and audio_clip.duration < video_dur: from moviepy.editor import concatenate_audioclips n_loops = int(np.ceil(video_dur / audio_clip.duration)) audio_clip = concatenate_audioclips([audio_clip] * n_loops) audio_clip = audio_clip.subclip(0, min(audio_clip.duration, video_dur)) if fade_in > 0: audio_clip = audio_clip.audio_fadein(fade_in) if fade_out > 0: audio_clip = audio_clip.audio_fadeout(fade_out) audio_clip = audio_clip.volumex(volume) final_video = video_clip.set_audio(audio_clip) with tempfile.NamedTemporaryFile(suffix="_music.mp4", delete=False) as tmp: output_path = tmp.name final_video.write_videofile( output_path, codec="libx264", audio_codec="aac", temp_audiofile=output_path + "_temp_audio.m4a", remove_temp=True, verbose=False, logger=None, ) video_clip.close() audio_clip.close() final_video.close() track_info = "" if music_source == "library" and selected_track in MUSIC_LIBRARY: mood = MUSIC_LIBRARY[selected_track]["mood"] track_info = f" | {selected_track} ({mood})" return output_path, ( f"✅ **Music added successfully!**{track_info}\n" f"🔊 Volume: {int(volume*100)}% | Fade In: {fade_in}s | Fade Out: {fade_out}s" ) except Exception as e: print(f"[Music] Error mixing audio: {e}") return video_path, f"❌ Error merging audio: {str(e)[:100]}" def suggest_music_for_preset(preset_name: str) -> str: return PRESET_TO_MUSIC.get(preset_name, "🎬 Cinematic Epic") def suggest_music_with_claude(prompt: str) -> tuple: if not ANTHROPIC_AVAILABLE: return "🎬 Cinematic Epic", "💡 Default selection (Claude not available)" track_list = "\n".join([ f"- {name}: {info['mood']} (BPM: {info['bpm']})" for name, info in MUSIC_LIBRARY.items() ]) try: client = anthropic.Anthropic() msg = client.messages.create( model="claude-opus-4-6", max_tokens=100, messages=[{ "role": "user", "content": ( "You are a music supervisor for video content.\n" f"Based on this video prompt: '{prompt}'\n\n" f"Choose the BEST matching music track from this list:\n{track_list}\n\n" "Reply with ONLY the exact track name (including emoji), nothing else." ), }], ) suggested = msg.content[0].text.strip() if suggested in MUSIC_LIBRARY: mood = MUSIC_LIBRARY[suggested]["mood"] return suggested, f"🤖 **Claude selected:** {suggested} — {mood}" return "🎬 Cinematic Epic", "💡 Default selection" except Exception as e: return "🎬 Cinematic Epic", f"💡 Default selection ({str(e)[:50]})" # ── Load pipeline — memory-efficient sequential loading ────────────────────── _BF16_REPO = 'cbensimon/Wan2.2-I2V-A14B-bf16-Diffusers' def _gc(): gc.collect() torch.cuda.empty_cache() torch.cuda.synchronize() print("[Load] Step 1/5 — loading transformer …") _transformer = WanTransformer3DModel.from_pretrained( _BF16_REPO, subfolder='transformer', torch_dtype=torch.bfloat16, device_map='cuda', low_cpu_mem_usage=True, ) _gc() print("[Load] Step 2/5 — loading transformer_2 …") _transformer_2 = WanTransformer3DModel.from_pretrained( _BF16_REPO, subfolder='transformer_2', torch_dtype=torch.bfloat16, device_map='cuda', low_cpu_mem_usage=True, ) _gc() print("[Load] Step 3/5 — loading full pipeline (text-encoder + VAE on CPU) …") pipe = WanImageToVideoPipeline.from_pretrained( MODEL_ID, transformer=_transformer, transformer_2=_transformer_2, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, ) del _transformer, _transformer_2 _gc() pipe.text_encoder.to('cpu') pipe.vae.to('cuda') pipe.transformer.to('cuda') pipe.transformer_2.to('cuda') # ── ARCHITECTURE FIX: Split CPU/GPU Routing Patch ───────────────────────────── WanImageToVideoPipeline._execution_device = property(lambda self: torch.device("cuda")) WanImageToVideoPipeline.device = property(lambda self: torch.device("cuda")) orig_te_forward = pipe.text_encoder.forward def patched_te_forward(*args, **kwargs): new_args = tuple(a.to("cpu") if isinstance(a, torch.Tensor) else a for a in args) new_kwargs = {k: v.to("cpu") if isinstance(v, torch.Tensor) else v for k, v in (kwargs or {}).items()} return orig_te_forward(*new_args, **new_kwargs) pipe.text_encoder.forward = patched_te_forward # ───────────────────────────────────────────────────────────────────────────── _gc() print("[Load] Step 4/5 — fusing LoRA weights …") pipe.load_lora_weights( "Kijai/WanVideo_comfy", weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors", adapter_name="lightx2v", ) pipe.load_lora_weights( "Kijai/WanVideo_comfy", weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors", adapter_name="lightx2v_2", load_into_transformer_2=True, ) pipe.set_adapters(["lightx2v", "lightx2v_2"], adapter_weights=[1.0, 1.0]) pipe.fuse_lora(adapter_names=["lightx2v"], lora_scale=3.0, components=["transformer"]) pipe.fuse_lora(adapter_names=["lightx2v_2"], lora_scale=1.0, components=["transformer_2"]) pipe.unload_lora_weights() _gc() print("[Load] Step 5/5 — quantising components …") quantize_(pipe.text_encoder, Int8WeightOnlyConfig()) _gc() quantize_(pipe.transformer, Float8DynamicActivationFloat8WeightConfig()) _gc() quantize_(pipe.transformer_2, Float8DynamicActivationFloat8WeightConfig()) _gc() aoti.aoti_blocks_load(pipe.transformer, 'zerogpu-aoti/Wan2', variant='fp8da') aoti.aoti_blocks_load(pipe.transformer_2, 'zerogpu-aoti/Wan2', variant='fp8da') _gc() print("[Load] ✅ Pipeline ready") # ── Helpers ──────────────────────────────────────────────────────────────────── 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 = max(MIN_DIM, min(MAX_DIM, round(target_w / MULTIPLE_OF) * MULTIPLE_OF)) final_h = max(MIN_DIM, min(MAX_DIM, round(target_h / MULTIPLE_OF) * MULTIPLE_OF)) return image_to_resize.resize((final_w, final_h), Image.LANCZOS) def get_num_frames(duration_seconds: float) -> int: raw = int(np.clip( int(round(duration_seconds * FIXED_FPS)), MIN_FRAMES_MODEL, MAX_FRAMES_MODEL, )) total = 1 + raw remainder = (total - 1) % 4 if remainder != 0: total += (4 - remainder) if total > MAX_FRAMES_MODEL: total -= 4 return total def get_duration(input_image, prompt, steps, negative_prompt, duration_seconds, guidance_scale, guidance_scale_2, seed, randomize_seed, progress): if input_image is None: return 60 BASE_FRAMES_HEIGHT_WIDTH = 81 * 832 * 624 BASE_STEP_DURATION = 15 if isinstance(input_image, str): input_image = Image.open(input_image).convert("RGB") elif isinstance(input_image, __import__("numpy").ndarray): input_image = Image.fromarray(input_image).convert("RGB") width, height = resize_image(input_image).size frames = get_num_frames(duration_seconds) factor = frames * width * height / BASE_FRAMES_HEIGHT_WIDTH step_duration = BASE_STEP_DURATION * factor ** 1.5 return 10 + int(steps) * step_duration def save_to_history(prompt: str, seed: int, duration: float): history = [] if os.path.exists(HISTORY_FILE): try: with open(HISTORY_FILE, 'r') as f: history = json.load(f) except Exception: history = [] history.append({ "timestamp": datetime.now().strftime("%Y-%m-%d %H:%M"), "prompt": prompt[:120], "seed": seed, "duration": duration, }) with open(HISTORY_FILE, 'w') as f: json.dump(history[-100:], f, indent=2) def load_history_display(): if not os.path.exists(HISTORY_FILE): return "No generations yet." try: with open(HISTORY_FILE, 'r') as f: history = json.load(f) rows = [] for h in reversed(history[-20:]): rows.append( f"**{h['timestamp']}** | 🌱 {h['seed']} | ⏱ {h['duration']}s\n" f"> {h['prompt']}\n" ) return "\n---\n".join(rows) if rows else "No generations yet." except Exception: return "Could not load history." def apply_preset(preset_name: str) -> tuple: prompt = default_prompt_i2v suggested_music = "🎬 Cinematic Epic" if preset_name and preset_name in MOTION_PRESETS: prompt = f"make this image come alive, {MOTION_PRESETS[preset_name]}" suggested_music = PRESET_TO_MUSIC.get(preset_name, "🎬 Cinematic Epic") return prompt, suggested_music # ── Core generation ──────────────────────────────────────────────────────────── @spaces.GPU(duration=get_duration) def generate_video( input_image, prompt, steps = 6, negative_prompt = default_negative_prompt, duration_seconds = 3.5, guidance_scale = 1.0, guidance_scale_2 = 1.0, seed = 42, randomize_seed = True, progress = gr.Progress(track_tqdm=True), ): if input_image is None: raise gr.Error("⚠️ Please upload an input image first!") if isinstance(input_image, str): input_image = Image.open(input_image).convert("RGB") elif isinstance(input_image, __import__("numpy").ndarray): input_image = Image.fromarray(input_image).convert("RGB") final_prompt = prompt num_frames = get_num_frames(duration_seconds) current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed) resized = resize_image(input_image) torch.cuda.empty_cache() gc.collect() progress(0, desc="🎬 Generating video…") try: output_frames = pipe( image = resized, prompt = final_prompt, negative_prompt = negative_prompt, height = resized.height, width = resized.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), ).frames[0] except torch.cuda.OutOfMemoryError: torch.cuda.empty_cache() gc.collect() raise gr.Error("🔴 GPU out of memory. Try reducing duration or resolution and retry.") with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp: video_path = tmp.name export_to_video(output_frames, video_path, fps=FIXED_FPS) save_to_history(final_prompt, current_seed, duration_seconds) torch.cuda.empty_cache() gc.collect() return video_path, current_seed # ── CSS ──────────────────────────────────────────────────────────────────────── custom_css = """ @import url('https://fonts.googleapis.com/css2?family=Inter:wght@300;400;500;600;700&display=swap'); body, .gradio-container { font-family: 'Inter', sans-serif !important; background: #0f0f13 !important; color: #e2e2e8 !important; } .gradio-container .contain { max-width: 1280px !important; margin: 0 auto !important; padding: 0 16px !important; } .hero-header { text-align: center; padding: 32px 16px 16px; background: linear-gradient(135deg, #1a1a2e 0%, #16213e 50%, #0f3460 100%); border-radius: 16px; margin-bottom: 20px; border: 1px solid rgba(99,102,241,0.3); } .hero-header h1 { font-size: 2.2em !important; font-weight: 700 !important; background: linear-gradient(135deg, #818cf8, #c084fc, #f472b6) !important; -webkit-background-clip: text !important; -webkit-text-fill-color: transparent !important; margin-bottom: 8px !important; } .hero-header p { color: #94a3b8 !important; font-size: 1em !important; } .badge { display: inline-block; background: rgba(99,102,241,0.2); border: 1px solid rgba(99,102,241,0.5); border-radius: 20px; padding: 3px 12px; font-size: 0.78em; color: #818cf8; margin: 4px 3px; } .badge-music { display: inline-block; background: rgba(234,179,8,0.15); border: 1px solid rgba(234,179,8,0.4); border-radius: 20px; padding: 3px 12px; font-size: 0.78em; color: #fbbf24; margin: 4px 3px; } .tab-nav button { background: #1e1e2e !important; color: #94a3b8 !important; border: 1px solid #2d2d3f !important; border-radius: 8px 8px 0 0 !important; font-weight: 500 !important; font-size: 0.9em !important; padding: 10px 18px !important; margin-right: 4px !important; transition: all 0.2s !important; } .tab-nav button.selected { background: linear-gradient(135deg, #4f46e5, #7c3aed) !important; color: #ffffff !important; border-color: transparent !important; } .panel-card { background: #1a1a2e !important; border: 1px solid #2d2d3f !important; border-radius: 12px !important; padding: 16px !important; } .music-panel { background: linear-gradient(135deg, rgba(234,179,8,0.06), rgba(251,146,60,0.06)); border: 1px solid rgba(234,179,8,0.25); border-radius: 12px; padding: 16px; margin-top: 8px; } .music-info { background: linear-gradient(135deg, rgba(234,179,8,0.08), rgba(251,146,60,0.08)); border: 1px solid rgba(234,179,8,0.3); border-radius: 10px; padding: 12px 16px; font-size: 0.85em; color: #fde68a; margin-bottom: 12px; } .music-info b { color: #fbbf24; } .colorize-info { background: linear-gradient(135deg, rgba(16,185,129,0.08), rgba(6,182,212,0.08)); border: 1px solid rgba(16,185,129,0.25); border-radius: 10px; padding: 12px 16px; font-size: 0.85em; color: #6ee7b7; margin-bottom: 12px; } .colorize-info b { color: #34d399; } textarea, input[type="text"], input[type="number"] { background: #0f0f18 !important; border: 1px solid #3d3d5c !important; border-radius: 8px !important; color: #e2e2e8 !important; font-family: 'Inter', sans-serif !important; } textarea:focus, input:focus { border-color: #6366f1 !important; box-shadow: 0 0 0 3px rgba(99,102,241,0.15) !important; } .gr-slider input[type=range] { accent-color: #6366f1 !important; } .btn-generate { background: linear-gradient(135deg, #4f46e5 0%, #7c3aed 50%, #a21caf 100%) !important; border: none !important; border-radius: 10px !important; color: #ffffff !important; font-size: 1.05em !important; font-weight: 600 !important; padding: 14px 28px !important; width: 100% !important; cursor: pointer !important; transition: opacity 0.2s !important; box-shadow: 0 4px 20px rgba(99,102,241,0.4) !important; } .btn-generate:hover { opacity: 0.88 !important; } .btn-music { background: linear-gradient(135deg, #d97706 0%, #ea580c 100%) !important; border: none !important; border-radius: 10px !important; color: #ffffff !important; font-size: 1.05em !important; font-weight: 600 !important; padding: 14px 28px !important; width: 100% !important; cursor: pointer !important; transition: opacity 0.2s !important; box-shadow: 0 4px 20px rgba(217,119,6,0.4) !important; } .btn-music:hover { opacity: 0.88 !important; } .btn-colorize { background: linear-gradient(135deg, #059669 0%, #0891b2 100%) !important; border: none !important; border-radius: 10px !important; color: #ffffff !important; font-size: 1.05em !important; font-weight: 600 !important; padding: 14px 28px !important; width: 100% !important; cursor: pointer !important; transition: opacity 0.2s !important; box-shadow: 0 4px 20px rgba(16,185,129,0.35) !important; } .btn-colorize:hover { opacity: 0.88 !important; } .btn-send-to-generate { background: rgba(99,102,241,0.15) !important; border: 1px solid rgba(99,102,241,0.5) !important; border-radius: 8px !important; color: #a5b4fc !important; font-size: 0.9em !important; font-weight: 500 !important; padding: 9px 18px !important; cursor: pointer !important; transition: all 0.2s !important; width: 100% !important; } .btn-send-to-generate:hover { background: rgba(99,102,241,0.28) !important; color: #ffffff !important; } .btn-ai-music { background: rgba(234,179,8,0.15) !important; border: 1px solid rgba(234,179,8,0.5) !important; border-radius: 8px !important; color: #fbbf24 !important; font-size: 0.88em !important; font-weight: 500 !important; padding: 8px 16px !important; cursor: pointer !important; transition: all 0.2s !important; width: 100% !important; } .btn-ai-music:hover { background: rgba(234,179,8,0.25) !important; } .musicgen-panel { background: linear-gradient(135deg, rgba(139,92,246,0.08), rgba(236,72,153,0.06)); border: 1px solid rgba(139,92,246,0.3); border-radius: 12px; padding: 16px; margin-top: 8px; } .musicgen-info { background: linear-gradient(135deg, rgba(139,92,246,0.1), rgba(236,72,153,0.08)); border: 1px solid rgba(139,92,246,0.35); border-radius: 10px; padding: 14px 16px; font-size: 0.85em; color: #c4b5fd; margin-bottom: 14px; } .musicgen-info b { color: #a78bfa; } .btn-musicgen { background: linear-gradient(135deg, #7c3aed 0%, #db2777 100%) !important; border: none !important; border-radius: 10px !important; color: #ffffff !important; font-size: 1.05em !important; font-weight: 600 !important; padding: 14px 28px !important; width: 100% !important; cursor: pointer !important; transition: opacity 0.2s !important; box-shadow: 0 4px 20px rgba(124,58,237,0.45) !important; } .btn-musicgen:hover { opacity: 0.88 !important; } .music-prompt-box { background: rgba(139,92,246,0.06) !important; border: 1px solid rgba(139,92,246,0.25) !important; border-radius: 8px !important; color: #c4b5fd !important; font-size: 0.88em !important; } .preset-radio .wrap { gap: 8px !important; flex-wrap: wrap !important; } .preset-radio label { background: #1e1e2e !important; border: 1px solid #3d3d5c !important; border-radius: 20px !important; color: #a5b4fc !important; font-size: 0.85em !important; padding: 7px 16px !important; cursor: pointer !important; transition: all 0.2s !important; margin: 0 !important; } .preset-radio label:hover { background: rgba(99,102,241,0.2) !important; border-color: #6366f1 !important; color: #ffffff !important; } .preset-radio input[type="radio"]:checked + span, .preset-radio label.selected { background: linear-gradient(135deg,#4f46e5,#7c3aed) !important; border-color: transparent !important; color: #ffffff !important; } .preset-radio .gr-radio-row { display: none !important; } .preset-radio > .wrap > label > input { display: none !important; } video { border-radius: 12px !important; border: 2px solid #2d2d3f !important; } label > span { color: #a5b4fc !important; font-size: 0.85em !important; font-weight: 500 !important; text-transform: uppercase !important; letter-spacing: 0.04em !important; } .stats-bar { background: #12121c; border: 1px solid #2d2d3f; border-radius: 8px; padding: 8px 16px; font-size: 0.82em; color: #64748b; text-align: center; margin-top: 8px; } .history-box { background: #0f0f18 !important; border: 1px solid #2d2d3f !important; border-radius: 10px !important; padding: 12px !important; font-size: 0.84em !important; max-height: 400px !important; overflow-y: auto !important; } .gr-accordion { border: 1px solid #2d2d3f !important; border-radius: 10px !important; } footer { display: none !important; } .runpod-banner { background: linear-gradient(135deg, #1e1b4b 0%, #312e81 50%, #4f46e5 100%); border: 1px solid rgba(99, 102, 241, 0.4); border-radius: 12px; padding: 18px 24px; margin: 0 auto 20px auto; text-align: center; max-width: 900px; box-shadow: 0 4px 20px rgba(79, 70, 229, 0.25); transition: all 0.3s ease; text-decoration: none !important; display: block; } .runpod-banner:hover { transform: translateY(-3px); box-shadow: 0 8px 25px rgba(79, 70, 229, 0.4); border-color: rgba(99, 102, 241, 0.8); } .runpod-title { color: #ffffff; font-size: 1.15em; font-weight: 600; margin-bottom: 6px; } .runpod-subtitle { color: #c7d2fe; font-size: 0.95em; margin-bottom: 12px; } .runpod-highlight { color: #fbbf24; font-weight: 800; text-shadow: 0 0 10px rgba(251, 191, 36, 0.3); } .runpod-button { display: inline-block; background: #ffffff; color: #4f46e5 !important; padding: 8px 22px; border-radius: 8px; font-weight: 700; font-size: 0.9em; transition: background 0.2s; } .runpod-banner:hover .runpod-button { background: #fbbf24; color: #78350f !important; } """ # ── UI ───────────────────────────────────────────────────────────────────────── with gr.Blocks(theme=gr.themes.Base(), css=custom_css, title="Wan 2.2 I2V ⚡") as demo: gr.HTML("""

⚡ Wan 2.2 I2V — Lightning Video

Transform any image into a cinematic video in just 4–8 steps


🔥 14B Parameters ⚡ Lightning LoRA 🎯 FP8 Quantized 🚀 AoT Compiled 🎨 B&W Colorization 🎵 Music Composer 🤖 MusicGen AI
🚀 Skip the Queue & Run This Model Instantly on Your Own GPU!
Deploy your private instance on RunPod for lightning-fast generations and get a Bonus Credit up to $500!
*Bonus is applied automatically after your first $10 account fund.
Claim Your Bonus & Deploy Now ⚡
 """) with gr.Tabs(): # ── Tab 1: Generate ──────────────────────────────────────────────────── with gr.TabItem("🎬 Generate"): with gr.Row(equal_height=False): with gr.Column(scale=1, min_width=380): input_image_component = gr.Image( type="numpy", label="📸 Upload Image", height=280, elem_classes=["panel-card"], ) with gr.Group(elem_classes=["panel-card"]): prompt_input = gr.Textbox( label="✍️ Prompt", value=default_prompt_i2v, lines=3, placeholder="Describe the motion, camera, atmosphere…", ) gr.Markdown("**⚡ Motion Presets**", elem_classes=["panel-card"]) preset_radio = gr.Radio( choices=list(MOTION_PRESETS.keys()), value=None, label="", interactive=True, elem_classes=["preset-radio"], ) duration_seconds_input = gr.Slider( minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.5, value=3.5, label="⏱️ Duration (seconds)", info=f"Model range: {MIN_FRAMES_MODEL}–{MAX_FRAMES_MODEL} frames @ {FIXED_FPS} fps", ) with gr.Accordion("⚙️ Advanced Settings", open=False): negative_prompt_input = gr.Textbox( label="🚫 Negative Prompt", value=default_negative_prompt, lines=3, ) with gr.Row(): seed_input = gr.Slider( label="🌱 Seed", minimum=0, maximum=MAX_SEED, step=1, value=42, ) randomize_seed_checkbox = gr.Checkbox( label="🎲 Randomize", value=True, ) steps_slider = gr.Slider( minimum=1, maximum=30, step=1, value=6, label="🔁 Inference Steps", info="4–8 steps recommended with Lightning LoRA", ) guidance_scale_input = gr.Slider( minimum=0.0, maximum=10.0, step=0.5, value=1.0, label="🎯 Guidance Scale (high-noise stage)", ) guidance_scale_2_input = gr.Slider( minimum=0.0, maximum=10.0, step=0.5, value=1.0, label="🎯 Guidance Scale 2 (low-noise stage)", ) generate_button = gr.Button( "🚀 Generate Video", variant="primary", size="lg", elem_classes=["btn-generate"], ) with gr.Column(scale=1, min_width=400): video_output = gr.Video( label="🎬 Generated Video", autoplay=True, interactive=False, height=360, ) with gr.Row(): seed_display = gr.Number( label="🌱 Used Seed", interactive=False, precision=0, ) # ── 🎵 Music Panel (inside Generate tab) ────────────────── with gr.Accordion("🎵 Add Music to Video", open=False, elem_classes=["music-panel"]): gr.HTML("""
🎵 Music Composer — Add cinematic music to your video! Choose from 12 ready-to-use royalty-free tracks or upload your own. Claude AI can automatically pick the best music for your prompt!
""") music_source_radio = gr.Radio( choices=["🎵 From Library", "📁 Upload File", "🔇 No Music"], value="🎵 From Library", label="Music Source", ) with gr.Group(visible=True) as music_library_group: music_track_dropdown = gr.Dropdown( choices=list(MUSIC_LIBRARY.keys()), value="🎬 Cinematic Epic", label="🎼 Select Track", info="12 Royalty-Free tracks ready to use", ) ai_music_suggest_btn = gr.Button( "🤖 Let Claude Pick the Best Music", elem_classes=["btn-ai-music"], size="sm", ) music_ai_status = gr.Markdown("", label="") with gr.Group(visible=False) as music_upload_group: custom_audio_upload = gr.Audio( label="📁 Upload Audio File (MP3/WAV/OGG)", type="filepath", sources=["upload"], ) with gr.Row(): music_volume = gr.Slider( minimum=0.0, maximum=1.0, step=0.05, value=0.6, label="🔊 Volume", ) with gr.Row(): music_fade_in = gr.Slider( minimum=0.0, maximum=3.0, step=0.5, value=1.0, label="🌅 Fade In (seconds)", ) music_fade_out = gr.Slider( minimum=0.0, maximum=3.0, step=0.5, value=1.5, label="🌇 Fade Out (seconds)", ) music_loop_checkbox = gr.Checkbox( label="🔁 Loop music if shorter than video", value=True, ) add_music_btn = gr.Button( "🎵 Merge Music with Video", elem_classes=["btn-music"], size="lg", ) music_status = gr.Markdown("", label="") video_with_music = gr.Video( label="🎬🎵 Video with Music", autoplay=True, interactive=False, height=300, visible=True, ) gr.HTML("""
💡 Tip: Use 4–6 steps for fastest results  |  Randomize seed for variety  |  🎵 Add Music for cinematic feel  |  🎨 Colorize B&W images first!
""") # ── Tab 2: 🎵 Music Studio ───────────────────────────────────────────── with gr.TabItem("🎵 Music Studio"): gr.HTML("""
🎵 Music Studio — Add music to any existing video!
Upload a video and choose music from our library or from your own files. Claude AI automatically picks the most fitting music for you!
""") with gr.Row(equal_height=False): with gr.Column(scale=1, min_width=360): studio_video_input = gr.Video( label="📹 Upload Video", height=280, ) studio_music_source = gr.Radio( choices=["🎵 From Library", "📁 Upload File", "🔇 No Music"], value="🎵 From Library", label="Music Source", ) with gr.Group(visible=True) as studio_library_group: studio_track = gr.Dropdown( choices=list(MUSIC_LIBRARY.keys()), value="🎬 Cinematic Epic", label="🎼 Track", ) track_info_display = gr.Markdown( value="**Mood:** cinematic, dramatic | **BPM:** 120", label="", ) with gr.Group(visible=False) as studio_upload_group: studio_custom_audio = gr.Audio( label="📁 Upload Audio File", type="filepath", sources=["upload"], ) studio_volume = gr.Slider( minimum=0.0, maximum=1.0, step=0.05, value=0.65, label="🔊 Volume", ) with gr.Row(): studio_fade_in = gr.Slider( minimum=0.0, maximum=3.0, step=0.5, value=1.0, label="🌅 Fade In", ) studio_fade_out = gr.Slider( minimum=0.0, maximum=3.0, step=0.5, value=2.0, label="🌇 Fade Out", ) studio_loop = gr.Checkbox(label="🔁 Loop Music", value=True) studio_add_btn = gr.Button( "🎵 Merge Music", variant="primary", size="lg", elem_classes=["btn-music"], ) studio_status = gr.Markdown("") with gr.Column(scale=1, min_width=400): studio_output = gr.Video( label="🎬🎵 Result", autoplay=True, height=400, ) gr.Markdown("### 🎼 Music Library") music_grid_html = "
" for name, info in MUSIC_LIBRARY.items(): music_grid_html += f"""
{name}
{info['mood']}
BPM: {info['bpm']}
""" music_grid_html += "
" gr.HTML(music_grid_html) # ── Tab 3: 🤖 MusicGen AI ───────────────────────────────────────────── with gr.TabItem("🤖 MusicGen AI"): gr.HTML("""
🤖 MusicGen AI (Meta) — Generates 100% original music custom-made for your video!

Workflow: Claude analyzes the prompt → writes a professional music description → MusicGen generates the music → automatically merged with your video.

⚠️ Note: Generation takes 20–40 seconds on CPU — the quality is worth the wait!
""") with gr.Row(equal_height=False): with gr.Column(scale=1, min_width=360): musicgen_video_input = gr.Video( label="📹 Video to Add Music To", height=260, ) musicgen_prompt_input = gr.Textbox( label="✍️ Video Scene Description (for AI)", placeholder="e.g. A cat skateboarding on the beach in a playful and cheerful way…", lines=3, info="The more precise the description, the better the music fit", ) gr.HTML("""
💡 Tip: You can copy the prompt directly from the Generate tab
""") with gr.Row(): musicgen_duration = gr.Slider( minimum=2.0, maximum=30.0, step=1.0, value=5.0, label="⏱️ Music Duration (seconds)", info="Will be automatically trimmed to video length", ) with gr.Row(): musicgen_volume = gr.Slider( minimum=0.1, maximum=1.0, step=0.05, value=0.7, label="🔊 Volume", ) with gr.Row(): musicgen_fade_in = gr.Slider( minimum=0.0, maximum=2.0, step=0.25, value=0.5, label="🌅 Fade In (seconds)", ) musicgen_fade_out = gr.Slider( minimum=0.0, maximum=2.0, step=0.25, value=1.0, label="🌇 Fade Out (seconds)", ) musicgen_generate_btn = gr.Button( "🤖 Generate AI Music & Merge", variant="primary", size="lg", elem_classes=["btn-musicgen"], ) copy_from_generate_btn = gr.Button( "📋 Transfer Video from Generate Tab", size="sm", elem_classes=["btn-send-to-generate"], ) with gr.Column(scale=1, min_width=400): musicgen_output_video = gr.Video( label="🎬🎵 Video with AI Music", autoplay=True, height=320, ) musicgen_music_prompt_display = gr.Textbox( label="🎼 Music Description Used by MusicGen", interactive=False, lines=2, elem_classes=["music-prompt-box"], placeholder="The music description will appear here after generation…", ) musicgen_status = gr.Markdown("", label="") gr.HTML("""
🤖 MusicGen Small (~300MB) loads automatically on first use  |  🎼 Claude writes the music description  |  ✂️ Music is precisely trimmed to video length
""") with gr.Accordion("💡 Music Description Examples", open=False): gr.HTML("""
🌊 Nature Scene
peaceful ambient, soft piano, nature sounds, 70 bpm
⚡ Action Scene
action rock, electric guitar, fast drums, 140 bpm
🎬 Dramatic Scene
cinematic orchestra, dramatic strings, 100 bpm, epic
✨ Magical Scene
magical fantasy, harp and bells, 85 bpm, whimsical
""") # ── Tab 4: Colorize B&W ──────────────────────────────────────────────── with gr.TabItem("🎨 Colorize B&W"): gr.HTML("""
🎨 Deep-Learning Colorization — Upload a black-and-white (or faded) image and get a vivid, AI-colorized version. Then send it directly to Generate to create a cinematic video from your newly colorized photo!

Auto-selects best available engine: 🥇 Claude AI Vision  →  🥈 OpenCV DNN (auto-download)  →  🥉 Smart Semantic (always works)
""") with gr.Row(equal_height=False): with gr.Column(scale=1, min_width=360): colorize_input = gr.Image( type="pil", label="🖼️ Upload B&W Image", height=300, elem_classes=["panel-card"], ) colorize_style = gr.Radio( choices=["natural", "cinematic", "warm"], value="natural", label="🎨 Color Style", ) render_factor_slider = gr.Slider( minimum=10, maximum=45, step=5, value=35, label="🔍 Render Factor", ) colorize_btn = gr.Button( "🎨 Colorize Image", variant="primary", size="lg", elem_classes=["btn-colorize"], ) colorize_status = gr.Markdown("", label="") with gr.Column(scale=1, min_width=400): colorize_output = gr.Image( type="pil", label="🌈 Colorized Result", height=380, interactive=False, elem_classes=["panel-card"], ) send_to_generate_btn = gr.Button( "🚀 Send to Generate Tab", size="lg", elem_classes=["btn-send-to-generate"], ) with gr.Accordion("⚖️ Before / After Comparison", open=False): with gr.Row(): compare_original = gr.Image(type="pil", label="Original (B&W)", interactive=False, height=220) compare_colorized = gr.Image(type="pil", label="Colorized", interactive=False, height=220) def _run_colorize(img, style, factor): colorized, status = colorize_image(img, style, factor) return colorized, status, img, colorized colorize_btn.click( fn=_run_colorize, inputs=[colorize_input, colorize_style, render_factor_slider], outputs=[colorize_output, colorize_status, compare_original, compare_colorized], ) send_to_generate_btn.click( fn=lambda img: img, inputs=[colorize_output], outputs=[input_image_component], ) # ── Tab 5: Examples ──────────────────────────────────────────────────── with gr.TabItem("📚 Examples"): gr.Markdown("### Ready-to-run examples") gr.Examples( examples=[ ["wan_i2v_input.JPG", "POV selfie video, white cat with sunglasses standing on surfboard, relaxed smile, tropical beach behind.", 4], ["wan22_input_2.jpg", "A sleek lunar vehicle glides, astronauts hop aboard. Aurora borealis ribbons dance across the star-filled sky.", 4], ["kill_bill.jpeg", "Uma Thurman's blade slowly warps and droops, molten silver flowing downward. Her expression shifts to bewilderment.", 6], ], inputs=[input_image_component, prompt_input, steps_slider], outputs=[video_output, seed_display], fn=generate_video, cache_examples=True, cache_mode="lazy", ) # ── Tab 6: History ───────────────────────────────────────────────────── with gr.TabItem("📊 History"): gr.Markdown("### Your last 20 generations") refresh_history_btn = gr.Button("🔄 Refresh", size="sm") history_display = gr.Markdown(value=load_history_display(), elem_classes=["history-box"]) refresh_history_btn.click(fn=load_history_display, inputs=[], outputs=[history_display]) # ── Tab 7: About ─────────────────────────────────────────────────────── with gr.TabItem("ℹ️ About"): gr.Markdown(""" ## About this Space | Feature | Detail | |---|---| | **Model** | Wan-AI/Wan2.2-I2V-A14B — 14 billion parameters | | **LoRA** | Lightning LoRA — 4–8 step generation | | **Quantisation** | FP8 dynamic + INT8 text encoder | | **🎨 Colorization** | Claude Vision + OpenCV DNN + Smart Semantic | | **🎵 Music** | 12 Royalty-Free tracks + custom upload + AI selection | ### 🎵 Music Composer - **12 ready-to-use tracks** Royalty-Free for all tastes - **Upload your own file** MP3/WAV/OGG - **Claude AI** selects the best music based on your prompt - **Fade In/Out** automatic control + volume adjustment - **Loop** music to cover the full video duration --- ### 🚀 Host It Yourself on RunPod Tired of waiting in the public queue? You can run this exact model and UI on your own private GPU using **RunPod**. * **Instant Access:** Get your own dedicated GPU in seconds and bypass all limits. * **Special Offer:** Use my [referral link](https://console.runpod.io/deploy?template=adf0boho9x&ref=ev68fdmc) to sign up. When you add your first $10 to fund your account, RunPod will instantly reward you with a **random bonus credit ranging from $5 to $500**! """) # ── Event Wiring ─────────────────────────────────────────────────────────── copy_from_generate_btn.click( fn=lambda v: v, inputs=[video_output], outputs=[musicgen_video_input], ) musicgen_generate_btn.click( fn=generate_music_for_video, inputs=[ musicgen_video_input, musicgen_prompt_input, musicgen_duration, musicgen_fade_in, musicgen_fade_out, musicgen_volume, ], outputs=[musicgen_output_video, musicgen_music_prompt_display, musicgen_status], ) ui_inputs = [ input_image_component, prompt_input, steps_slider, negative_prompt_input, duration_seconds_input, guidance_scale_input, guidance_scale_2_input, seed_input, randomize_seed_checkbox, ] generate_button.click( fn=generate_video, inputs=ui_inputs, outputs=[video_output, seed_display], ) preset_radio.change( fn=apply_preset, inputs=[preset_radio], outputs=[prompt_input, music_track_dropdown], ) def _toggle_music_source(source): is_lib = source == "🎵 From Library" is_upload = source == "📁 Upload File" return gr.update(visible=is_lib), gr.update(visible=is_upload) music_source_radio.change( fn=_toggle_music_source, inputs=[music_source_radio], outputs=[music_library_group, music_upload_group], ) studio_music_source.change( fn=_toggle_music_source, inputs=[studio_music_source], outputs=[studio_library_group, studio_upload_group], ) def _update_track_info(track_name): if track_name and track_name in MUSIC_LIBRARY: info = MUSIC_LIBRARY[track_name] return f"**Mood:** {info['mood']} | **BPM:** {info['bpm']}" return "" studio_track.change( fn=_update_track_info, inputs=[studio_track], outputs=[track_info_display], ) def _ai_suggest(prompt_text): track, status = suggest_music_with_claude(prompt_text) return track, status ai_music_suggest_btn.click( fn=_ai_suggest, inputs=[prompt_input], outputs=[music_track_dropdown, music_ai_status], ) def _add_music_generate( video_path, source, track, custom_audio, volume, fade_in, fade_out, loop ): src_map = { "🎵 From Library": "library", "📁 Upload File": "upload", "🔇 No Music": "none", } result_path, status = add_music_to_video( video_path, src_map.get(source, "none"), track, custom_audio, volume, fade_in, fade_out, loop, ) return result_path, status add_music_btn.click( fn=_add_music_generate, inputs=[ video_output, music_source_radio, music_track_dropdown, custom_audio_upload, music_volume, music_fade_in, music_fade_out, music_loop_checkbox, ], outputs=[video_with_music, music_status], ) def _studio_add_music(video, source, track, custom_audio, volume, fade_in, fade_out, loop): if video is None: return None, "⚠️ Please upload a video first." src_map = { "🎵 From Library": "library", "📁 Upload File": "upload", "🔇 No Music": "none", } result_path, status = add_music_to_video( video, src_map.get(source, "none"), track, custom_audio, volume, fade_in, fade_out, loop, ) return result_path, status studio_add_btn.click( fn=_studio_add_music, inputs=[ studio_video_input, studio_music_source, studio_track, studio_custom_audio, studio_volume, studio_fade_in, studio_fade_out, studio_loop, ], outputs=[studio_output, studio_status], ) # ── Launch ───────────────────────────────────────────────────────────────────── if __name__ == "__main__": demo.queue(max_size=10).launch(mcp_server=True)