Reproduce “V2M-Zero: Zero-Pair Time-Aligned Video-to-Music Generation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yan-Bo Lin, Jonah Casebeer, Long Mai, Aniruddha Mahapatra, Gedas Bertasius, Nicholas J. Bryan

V2M-ZERO generates music for video without paired video–music training data. It fine-tunes a pretrained text-to-music model on music and text, using music-derived event curves, then substitutes curves extracted from video at inference. Text guides musical character, while curves indicate timing. On OES-Pub, its scene-cut hit score is 0.61, versus 0.40 for VidMuse, measuring local beat-to-cut coincidence.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.