Reproduce “Recast the Past: Stable Long Video Generation with Full Context Memory” 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
Yukuo Ma, Cong Liu, Junke Wang, Junqi Liu, Haibin Huang, Zuxuan Wu, Xuelong Li
ReCam addresses long-form autoregressive video generation by grouping historical latents into spatiotemporal blocks and making their compressed representations readable through the model’s attention, while noisy recent context limits propagated errors. On 30-second, 480p VBench-Long generation, its 14B model scores 80.1 versus FramePack’s 79.5. ReCam leads all four full-set off-screen consistency metrics, with face-metric exceptions across subject-presence groups.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.