Reproduce “Heuristics are Not Enough: Calibrated and Corrected Caching for Video Diffusion Acceleration” 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
Jiyue Jiang, Zhiyu Huo, Yuhan Chen, Yanruisheng Shao, Gong Zhang, Lei Deng
AlignCache accelerates text-to-video diffusion by reusing predicted intermediate residuals and adjusting cache thresholds from the discrepancy between exact and drafted classifier-free-guided outputs at selected denoising steps. On Wan2.1-1.3B, it takes 57.5 seconds and reaches 2.29× speedup, versus EasyCache’s 61.1 seconds and 2.19×, with higher PSNR and lower LPIPS; these measure fidelity to the unaccelerated output, not ground-truth video quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.