Reproduce “No Cache Left Idle: Accelerating diffusion model via Extreme-slimming Caching” 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

Tingyan Wen, Haoyu Li, Yihuang Chen, Xing Zhou, Lifei Zhu, Xueqian Wang

X-Slim, developed by Tsinghua University and Huawei, introduces a unified, training-free caching framework that exploits redundancy across temporal, structural, and spatial dimensions to significantly accelerate diffusion model inference. The method achieves up to 4.97x speedup for text-to-image generation and 3.52x for text-to-video, while maintaining or improving generative quality compared to previous state-of-the-art approaches.

Key insights

  • Exploiting cacheable redundancy simultaneously across temporal, structural, and spatial dimensions is crucial for achieving extreme acceleration in diffusion models.
  • A "push-then-polish" adaptive control mechanism, using dual error thresholds, simplifies the policy search space and effectively balances aggressive reuse with quality preservation.
  • Understanding and leveraging the distinct evolution patterns of features at different levels (U-shaped for steps, depth-wise for blocks, content-dependent for tokens) allows for intelligent, fine-grained computational control.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.