Reproduce “Accelerating Diffusion Transformer via Increment-Calibrated Caching with Channel-Aware Singular Value Decomposition” 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
Zhiyuan Chen, Keyi Li, Yifan Jia, Le Ye, Yufei Ma
ICC reduces the cost of class-conditional ImageNet generation by reusing cached layer outputs and correcting them with a low-rank estimate of changed inputs, without retraining DiT weights. On DiT-XL/2, 30-step ICC uses 4.55T MACs versus 8.30T for 35-step DDIM, with higher IS and a reported FID increase below 0.06; this operation-count comparison does not measure latency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.