---
title: C-MET CPU
emoji: 🎭
colorFrom: green
colorTo: red
sdk: gradio
sdk_version: 6.12.0
python_version: "3.10"
app_file: app.py
pinned: false
license: mit
tags:
  - talking-face
  - emotion-transfer
  - cpu
  - onnx
  - mcp-server
short_description: C-MET Emotion Transfer - CPU (ONNX + JIT)
models:
  - coldhyuk/C-MET
---

# C-MET: Cross-Modal Emotion Transfer (CPU)

Talking-face emotion editing from text, audio, or video prompts. CPU-optimized port of [C-MET](https://github.com/ChanHyeok-Choi/C-MET).

## Real benchmarks on HF cpu-basic (2 vCPU, 7s video, 174 frames)

| Stage | Time | Method |
|---|---|---|
| audio2lip | 21.5s | ONNX FP32 |
| compute_alpha_D | 137.5s | PyTorch (StyleGAN QR op blocks ONNX) |
| connector | 0.4s | ONNX FP32 |
| render (174 frames) | 438.7s (2.5s/frame) | PyTorch eager |
| video_prep + encode | 4.5s | ffmpeg |
| **Total** | **~10 min** | |

ONNX artifacts are built on first startup (~1 min), cached in `./artifacts/`.

## API -- Python Client

```python
from gradio_client import Client, handle_file

client = Client("WeReCooking/C-MET-CPU")

result = client.predict(
    source_image=handle_file("face.png"),
    audio=handle_file("voice.wav"),
    pose=handle_file("pose.mp4"),
    emotion="happy",        # angry contempt disgusted fear happy sad surprised
                            # charismatic desirous empathetic envious romantic sarcastic
    intensity=3,
    seed=42,
    api_name="/infer",
)
output_path, timings_str = result
print(timings_str)
```

## MCP

```json
{
  "mcpServers": {
    "cmet": {"url": "https://werecooking-c-met-cpu.hf.space/gradio_api/mcp/"}
  }
}
```

## CLI

```bash
python app.py infer \
  -s asset/identity/ChatGPT_man3_crop.png \
  -a asset/audio/W009_038.wav \
  -v asset/video/W009_038.mp4 \
  -e happy \
  -o output.mp4
```

All 13 emotions: `angry contempt disgusted fear happy sad surprised charismatic desirous empathetic envious romantic sarcastic`

## Web UI

1. Upload source face (256x256 crop)
2. Upload driving audio (.wav)
3. Upload pose driving video (25 fps, face crop)
4. Pick target emotion
5. Click Generate -> wait -> get video with timing breakdown

## Credits

Based on [C-MET](https://github.com/ChanHyeok-Choi/C-MET) (CVPR 2026). Original ZeroGPU demo: [coldhyuk/C-MET](https://huggingface.co/spaces/coldhyuk/C-MET).