Reproduce “MoDDiT: Simple Token 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

Jarek Liesen, Matthew T. Jackson, Benjamin Ellis, Shimon Whiteson, Jakob N. Foerster

MoDDiT lets diffusion models allocate updates selectively across tokens. It learns separate routes for self-attention and MLP blocks, selecting which patch representations receive each update. On ImageNet-L/2 at 25% capacity and 512 denoising steps, it reports 21.7% fewer denoiser FLOPs than dense DiT-L/2, with better FID and Inception Score but worse sFID; these arithmetic results do not establish faster elapsed generation.

Summary by alphaXiv.

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