README.md

---
license: cc-by-nc-4.0
pipeline_tag: other
tags:
- video-to-audio
- audio-generation
---

# AC-Foley: Reference-Audio-Guided Video-to-Audio Synthesis with Acoustic Transfer

This repository contains the model described in the paper [AC-Foley: Reference-Audio-Guided Video-to-Audio Synthesis with Acoustic Transfer](https://huggingface.co/papers/2603.15597).

[**Code**](https://github.com/ff2416/AC-Foley) | [**Paper**](https://huggingface.co/papers/2603.15597)

## Model Description

AC-Foley is an audio-conditioned video-to-audio (V2A) model designed to achieve precise and fine-grained control over generated sounds. Unlike traditional V2A methods that rely heavily on text prompts, AC-Foley directly leverages reference audio to bypass the semantic ambiguities of text descriptions. This approach enables several key features:

- **Fine-grained sound synthesis**: Precise manipulation of acoustic attributes.
- **Timbre transfer**: Applying the characteristics of a reference audio to the video context.
- **Zero-shot generation**: Synthesizing high-quality audio for unseen categories.
- **Improved audio quality**: Achieving state-of-the-art performance for Foley generation.

## Citation

If you find this work useful for your research, please cite:

```bibtex
@article{fang2026acfoley,
  title={AC-Foley: Reference-Audio-Guided Video-to-Audio Synthesis with Acoustic Transfer},
  author={Fang, Pengjun and He, Yingqing and Xing, Yazhou and Chen, Qifeng and Lim, Ser-Nam and Yang, Harry},
  journal={arXiv preprint arXiv:2603.15597},
  year={2026}
}
```