Reproduce “Pixel-space video generation with U-JiT” 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

Théophane Vallaeys, Jakob Verbeek, Matthieu Cord

U-JiT generates videos directly from pixels while learning progressive compression and expansion inside a transformer, rather than using a separate pretrained video autoencoder. It uses fine patches at the input and output around a compact inner representation. On Kinetics-600 frames-to-video at 128×128, U-JiT-XL achieves FVD 0.61 versus WaiT-XL’s 0.84, a reported 27% improvement in that benchmark setting.

Summary by alphaXiv.

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