---
license: cc-by-nc-4.0
tags:
- mmaudio
- audio-generation
- video-to-audio
- fp16
---

# MMAudio Models (FP16 Safetensors)

FP16 `.safetensors` conversions of [MMAudio](https://github.com/hkchengrex/MMAudio) model checkpoints and dependencies for use with [ComfyUI-FFMPEGA](https://github.com/AEmotionStudio/ComfyUI-FFMPEGA).

## Models

### MMAudio Core

| File | Description | FP16 Size |
|------|-------------|-----------|
| `mmaudio_large_44k_v2.safetensors` | MMAudio large model (44kHz, v2) | 1,966 MB |
| `v1-44.safetensors` | VAE decoder (44kHz) | 583 MB |
| `synchformer_state_dict.safetensors` | Synchformer temporal encoder | 453 MB |

### Dependencies

| File | Description | Size |
|------|-------------|------|
| `apple_DFN5B-CLIP-ViT-H-14-384_fp16.safetensors` | CLIP vision encoder (fp16) | 1,882 MB |
| `bigvgan_v2_44khz_128band_512x/` | BigVGAN v2 vocoder (44kHz) | ~467 MB |

All models converted from original FP32 sources → **FP16** `.safetensors` (50% size reduction).

## Sources

All models were downloaded from their **original sources** and converted by us:

- **MMAudio**: [hkchengrex/MMAudio](https://huggingface.co/hkchengrex/MMAudio)
- **CLIP**: [apple/DFN5B-CLIP-ViT-H-14-384](https://huggingface.co/apple/DFN5B-CLIP-ViT-H-14-384)
- **BigVGAN**: [nvidia/bigvgan_v2_44khz_128band_512x](https://huggingface.co/nvidia/bigvgan_v2_44khz_128band_512x)

## Usage

These models are automatically downloaded by the `generate_audio` skill in ComfyUI-FFMPEGA. No manual setup needed.

**Manual installation:**
```
ComfyUI/models/mmaudio/
├── mmaudio_large_44k_v2.safetensors
├── v1-44.safetensors
├── synchformer_state_dict.safetensors
├── apple_DFN5B-CLIP-ViT-H-14-384_fp16.safetensors
└── bigvgan_v2_44khz_128band_512x/
    ├── config.json
    └── bigvgan_generator.pt
```

## License

> ⚠️ **CC-BY-NC 4.0** — MMAudio model checkpoints are licensed under [Creative Commons Attribution-NonCommercial 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
> Commercial use of the models is restricted. The code that loads/runs them is MIT/GPL-3.0.
>
> BigVGAN is licensed under MIT. CLIP (DFN5B) is licensed under Apple's research license.

## Paper

*Taming Multimodal Joint Training for High-Quality Video-to-Audio Synthesis* (CVPR 2025)