This work is done during my internship at Adobe Research. Big thanks to all my collaborators @pseeth.bsky.social, Bryan Russell, @urinieto.bsky.social, David Bourgin, @andrewowens.bsky.social, and @justinsalamon.bsky.social!

We jointly train our model on high-quality text-audio pairs as well as videos, enabling our model to generate full-bandwidth professional audio with fine-grained creative control and synchronization.

MultiFoley is a unified framework for video-guided audio generation leveraging text, audio, and video conditioning within a single model. As a result, we can do text-guided foley, audio-guided foley (e.g. sync your favorite sample with the video), and foley audio extension.

🎥 Introducing MultiFoley, a video-aware audio generation method with multimodal controls! 🔊
We can
⌨️Make a typewriter sound like a piano 🎹
🐱Make a cat meow like a lion roars! 🦁
⏱️Perfectly time existing SFX 💥 to a video.

arXiv: arxiv.org/abs/2411.17698
website: ificl.github.io/MultiFoley/