Reproduce “Speculative Forcing for Efficient Autoregressive Video 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
Hyunwook Choi, Seungryong Kim, Dongyoon Han, Sangdoo Yun
Speculative Forcing reduces denoising work for video and interactive world generation by predicting when latent video tokens are ready to stop, while continuing to refine others. In 5-second LongLive generation, it averages 1.68 evaluations per token versus 4.00 and measures 41.74 versus 26.40 frames per second, with VBench Total scores of 83.21 versus 83.08. These measurements are specific to this evaluation.
Key insights
- Denoising depth can be allocated per token rather than uniformly across a chunk, since token estimates converge at different rates.
- Token-level commitment performs better than applying the same learned decisions at chunk level in the reported ablation, where settled and unsettled regions need different treatment.
- Scene dynamics affect how much computation can be saved: the world-model setting with changing camera views uses more denoising steps per token than LongLive.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.