Reproduce “UniCtrl: Improving the Spatiotemporal Consistency of Text-to-Video Diffusion Models via Training-Free Unified Attention Control” 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
Tian Xia, Xuweiyi Chen, Sihan Xu
UniCtrl stabilizes text-to-video generation while retaining motion by reusing first-frame attention and injecting motion queries from a separate branch. It synchronizes the branches during inference without additional training or learned parameters. On AnimateDiff, at motion-injection degree 1 with 3 FreeInit iterations, it scores 96.38 DINO and 23.29 RAFT versus FreeInit’s 96.04 and 11.62; these measure similarity and flow magnitude, not natural motion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.