Reproduce “Rethinking Cross-Layer Information Routing in Diffusion Transformers” 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
Chao Xu, Maohua Li, Qirui Li, Yixuan Xu, Yanke Zhou, Yunhe Li, Cuifeng Shen, Hanlin Tang and 4 more
Diffusion Attention Routing (DAR) replaces fixed residual addition in Diffusion Transformers with learned weighting over earlier layer outputs, optionally conditioned on denoising time. Chunking limits the stored source set. On ImageNet 256×256, SiT-XL/2 with static, chunk-4 DAR reaches FID 7.56 after 600K iterations, versus 9.67 after 1.75M for baseline; this is an empirical result on the tested setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.