A new arXiv preprint asks whether autoregressive video diffusion can be accelerated further when only four denoising steps remain. The authors note that existing feature caching approaches struggle in this regime, because large feature changes between denoising timesteps undermine the assumptions that make caching effective.
The proposed method, called TGuide, uses historical trajectory guidance to address this problem. While the abstract does not spell out the full mechanism, the core idea is to guide the diffusion process using information from earlier trajectories rather than relying on stale cached features.
The paper is presented as a new announcement on arXiv, and the abstract focuses on the failure mode of feature caching in the low-step regime. No experimental results or comparisons are included in the abstract itself.