构建一个使用Gradio的本地深度视频转换器

# depth_video_converter.py import os import sys import tempfile import subprocess import numpy as np import cv2 import torch import gradio as gr from PIL import Image from transformers import pipeline from pathlib import Path import shutil import json # ========== Configuration ========== MODEL_SIZES = { "small": "depth-anything/Depth-Anything-V2-Small", "base": "depth-anything/Depth-Anything-V2-Base", "large": "depth-anything/Depth-Anything-V2-Large", "giant": "depth-anything/Depth-Anything-V2-Giant" } DEFAULT_OUTPUT_RES = 512 TEMP_DIR = tempfile.mkdtemp(prefix="depth_video_") # ========== Device Detection ========== def get_device(): """Detect best available device (CUDA > MPS > CPU)""" if torch.cuda.is_available(): return "cuda" elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available(): return "mps" else: return "cpu" DEVICE = get_device() print(f"Using device: {DEVICE}") # ========== Depth Model ========== def load_depth_model(model_size): """Load Depth Anything V2 model via HuggingFace pipeline""" model_id = MODEL_SIZES.get(model_size, MODEL_SIZES["small"]) print(f"Loading model: {model_id}") pipe = pipeline( task="depth-estimation", model=model_id, device=DEVICE if DEVICE != "mps" else -1 # MPS fallback to CPU for pipeline ) return pipe # ========== Video Processing ========== def extract_frames(video_path, target_fps=None): """Extract frames from video using OpenCV""" cap = cv2.VideoCapture(video_path) if not cap.isOpened(): raise ValueError(f"Cannot open video: {video_path}") fps = cap.get(cv2.CAP_PROP_FPS) if target_fps and target_fps < fps: frame_interval = int(fps / target_fps) else: frame_interval = 1 frames = [] frame_count = 0 while True: ret, frame = cap.read() if not ret: break if frame_count % frame_interval == 0: frames.append(frame) frame_count += 1 cap.release() return frames, fps def process_frames_with_depth(frames, model, output_res, invert, temporal_smoothing): """Convert frames to depth maps with optional temporal smoothing""" depth_frames = [] prev_depth = None for i, frame in enumerate(frames): # Convert BGR to RGB for model rgb_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) pil_image = Image.fromarray(rgb_frame) # Get depth prediction result = model(pil_image) depth = result["depth"] # Convert to numpy array if isinstance(depth, Image.Image): depth_np = np.array(depth) else: depth_np = np.array(depth) # Resize to output resolution depth_np = cv2.resize(depth_np, (output_res, output_res), interpolation=cv2.INTER_LINEAR) # Normalize to 0-255 depth_np = cv2.normalize(depth_np, None, 0, 255, cv2.NORM_MINMAX).astype(np.uint8) # Invert if requested if invert: depth_np = 255 - depth_np # Temporal smoothing (exponential moving average) if temporal_smoothing > 0 and prev_depth is not None: alpha = temporal_smoothing depth_np = (alpha * depth_np + (1 - alpha) * prev_depth).astype(np.uint8) prev_depth = depth_np depth_frames.append(depth_np) # Progress update if (i + 1) % 10 == 0: print(f"Processed {i + 1}/{len(frames)} frames") return depth_frames def create_video_from_frames(frames, output_path, fps, audio_path=None): """Create video from frames, optionally adding audio""" height, width = frames[0].shape[:2] # Write video fourcc = cv2.VideoWriter_fourcc(*'mp4v') out = cv2.VideoWriter(output_path, fourcc, fps, (width, height), isColor=False) for frame in frames: out.write(frame) out.release() # Add audio if provided if audio_path and os.path.exists(audio_path): temp_video = output_path + ".temp.mp4" os.rename(output_path, temp_video) cmd = [ "ffmpeg", "-y", "-i", temp_video, "-i", audio_path, "-c:v", "libx264", "-c:a", "aac", "-shortest", output_path ] subprocess.run(cmd, capture_output=True) os.remove(temp_video) # ========== Main Processing Function ========== def convert_video( input_video, model_size="small", output_res=DEFAULT_OUTPUT_RES, invert=False, temporal_smoothing=0.5, preserve_audio=True, progress=gr.Progress() ): """Main conversion function for Gradio UI""" try: progress(0, desc="Initializing...") # Validate input if not input_video: return None, "Please upload a video file." # Load model progress(0.1, desc="Loading depth model...") model = load_depth_model(model_size) # Extract frames progress(0.2, desc="Extracting frames...") frames, original_fps = extract_frames(input_video) if len(frames) == 0: return None, "No frames extracted from video." # Process frames progress(0.3, desc="Processing depth maps...") depth_frames = process_frames_with_depth( frames, model, output_res, invert, temporal_smoothing ) # Create output video progress(0.9, desc="Creating output video...") output_path = os.path.join(TEMP_DIR, "depth_output.mp4") # Extract audio if needed audio_path = None if preserve_audio: audio_path = os.path.join(TEMP_DIR, "temp_audio.aac") cmd = [ "ffmpeg", "-y", "-i", input_video, "-vn", "-acodec", "copy", audio_path ] result = subprocess.run(cmd, capture_output=True) if result.returncode != 0: audio_path = None # Audio extraction failed, continue without create_video_from_frames(depth_frames, output_path, original_fps, audio_path) progress(1.0, desc="Done!") return output_path, "Conversion successful!" except Exception as e: return None, f"Error: {str(e)}" # ========== Gradio UI ========== def create_ui(): """Create Gradio interface""" with gr.Blocks(title="Depth Video Converter") as demo: gr.Markdown(""" # Depth Video Converter Convert regular videos to grayscale depth videos using Depth Anything V2. """) with gr.Row(): with gr.Column(): input_video = gr.Video( label="Input Video (MP4/MOV)", format="mp4" ) model_size = gr.Dropdown( choices=list(MODEL_SIZES.keys()), value="small", label="Model Size" ) output_res = gr.Slider( minimum=256, maximum=1024, value=DEFAULT_OUTPUT_RES, step=64, label="Output Resolution" ) invert = gr.Checkbox( label="Invert Depth (black/white)", value=False ) temporal_smoothing = gr.Slider( minimum=0.0, maximum=0.9, value=0.5, step=0.1, label="Temporal Smoothing (reduce flicker)" ) preserve_audio = gr.Checkbox( label="Preserve Original Audio", value=True ) convert_btn = gr.Button("Convert to Depth Video", variant="primary") with gr.Column(): output_video = gr.Video(label="Output Depth Video") status = gr.Textbox(label="Status", interactive=False) convert_btn.click( fn=convert_video, inputs=[ input_video, model_size, output_res, invert, temporal_smoothing, preserve_audio ], outputs=[output_video, status] ) return demo # ========== Main Entry ========== if __name__ == "__main__": print("=" * 50) print("Depth Video Converter") print(f"Device: {DEVICE}") print("=" * 50) demo = create_ui() demo.launch( server_name="127.0.0.1", server_port=7860, share=False ) ``` ```text # requirements.txt gradio>=4.0.0 opencv-python>=4.8.0 torch>=2.0.0 transformers>=4.35.0 Pillow>=10.0.0 numpy>=1.24.0 ffmpeg-python>=0.2.0 ``` ```markdown # Depth Video Converter Convert regular videos to grayscale depth videos using Depth Anything V2, running locally on Windows and macOS. ## Features - **Web UI** via Gradio (simple browser interface) - **MP4/MOV** video upload support - **Depth Anything V2** model for depth estimation - **Hardware acceleration**: - NVIDIA CUDA on Windows (when available) - Apple Silicon (MPS) on macOS (when available) - CPU fallback - **Model size selection**: small, base, large, giant - **Output resolution** control (256-1024px) - **Black/white inversion** option - **Temporal smoothing** to reduce flicker - **Audio preservation** with ffmpeg - **MP4 export** compatible with most players ## Installation ### Prerequisites - Python 3.8 or higher - ffmpeg installed on your system ### Windows ```bash # 1. Install Python from python.org (check "Add to PATH") # 2. Install ffmpeg # Download from https://ffmpeg.org/download.html # Add ffmpeg.exe to your PATH # 3. Create and activate virtual environment python -m venv venv venv\Scripts\activate # 4. Install dependencies pip install -r requirements.txt # 5. (Optional) Install CUDA support for NVIDIA GPUs pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118 ``` ### macOS ```bash # 1. Install Homebrew if not installed /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)" # 2. Install Python and ffmpeg brew install python ffmpeg # 3. Create and activate virtual environment python3 -m venv venv source venv/bin/activate # 4. Install dependencies pip install -r requirements.txt # 5. (Optional) Install PyTorch with MPS support (Apple Silicon) pip install torch torchvision ``` ## Usage ### Launch the application ```bash # Windows python depth_video_converter.py # macOS python3 depth_video_converter.py ``` ### Using the Web UI 1. Open your browser to `http://127.0.0.1:7860` 2. Upload an MP4 or MOV video file 3. Configure settings: - **Model Size**: Choose between small (fast) and giant (accurate) - **Output Resolution**: Higher = better quality but slower - **Invert Depth**: Swap black/white for different visual styles - **Temporal Smoothing**: Higher values reduce flicker but may blur motion - **Preserve Audio**: Keep original audio track 4. Click "Convert to Depth Video" 5. Download the result ## Performance Tips - **Small model + 256px resolution**: Fastest, good for previews - **Large model + 512px resolution**: Balanced quality/speed - **Giant model + 1024px resolution**: Best quality, requires powerful GPU - Use CUDA on Windows for 2-5x speedup - Use MPS on Apple Silicon for 1.5-3x speedup ## Troubleshooting ### "No module named 'torch'" Run: `pip install torch torchvision` ### ffmpeg not found - Windows: Download and add to PATH - macOS: `brew install ffmpeg` ### CUDA not detected - Ensure NVIDIA drivers are installed - Install CUDA version of PyTorch (see installation) ### MPS not detected on Mac - Requires Apple Silicon (M1/M2/M3) - Update macOS and PyTorch ## License This project uses Depth Anything V2 (Apache 2.0) and is for educational/research purposes. ``` ```text # Installation and Launch Commands ## Windows (PowerShell or CMD) ```bash # One-time setup python -m venv venv venv\Scripts\activate pip install -r requirements.txt # Optional: CUDA support for NVIDIA GPUs pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118 # Launch python depth_video_converter.py ``` ## macOS (Terminal) ```bash # One-time setup python3 -m venv venv source venv/bin/activate pip install -r requirements.txt # Optional: MPS support for Apple Silicon pip install torch torchvision # Launch python3 depth_video_converter.py ``` ## Quick Start (both platforms) ```bash # After installation, just run: python depth_video_converter.py # or python3 depth_video_converter.py ``` Then open http://127.0.0.1:7860 in your browser.