README.md

metadata

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.

Code | Paper

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:

@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}
}