README.md

metadata

title: Wav2Lip CPU
emoji: 👄
sdk: gradio
sdk_version: 6.3.0
colorFrom: red
colorTo: pink
app_file: app.py
pinned: false
license: mit

Wav2Lip - CPU Lip Sync

Based on saifhassan/Wav2Lip-HD

CPU-only conversion for HuggingFace Spaces free tier (no GPU required).

Features

• ONNX Wav2Lip model (145MB) - runs on CPU
• OpenCV face detection - no GPU needed
• Mouth-paste approach - preserves original face quality
• Temporal smoothing - reduces face bbox flickering

How It Works

1. Detect face using OpenCV Haar Cascade
2. Crop and resize to 96x96
3. Run Wav2Lip ONNX inference
4. Extract mouth region from output
5. Scale and paste onto original face
6. Feather blend at seam

Limitations

• Max 500 frames (~20 seconds at 25fps)
• ~1-2 sec/frame on CPU
• Best with frontal faces

Credits

• Wav2Lip - Original model
• Wav2Lip-HD - HD approach
• bluefoxcreation/Wav2lip-Onnx - ONNX models