Reproduce “Token Caching for Diffusion Transformer 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
Jinming Lou, Wenyang Luo, Yufan Liu, Bing Li, Xinmiao Ding, Weiming Hu, Yuming Li, Chenguang Ma
TokenCache accelerates Diffusion Transformer inference by implementing a dynamic, hierarchical caching framework across token, block, and timestep levels, achieving up to 3.01x speedup while maintaining high generation quality across image and video synthesis tasks.
Key insights
- Token updates exhibit varying degrees of redundancy across timesteps and among individual tokens, suggesting that a fine-grained, adaptive caching approach is superior to uniform or static methods.
- A learned predictor can effectively identify less critical tokens and blocks for caching, thereby optimizing computational savings without significant degradation in output quality.
- Dynamic timestep scheduling, based on token importance, is crucial for preventing error accumulation and refreshing cached updates at critical junctures of the denoising process.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.