Reproduce “Hierarchical Diffusion Transformer for End-to-End Image Generation” 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

Xiaoyu Wu, Yifei Wang, Weihang Guo, Chen Wei, Lydia E Kavraki, Niloofar Mireshghallah, Steven Wu

HiT generates images directly from pixels by grouping image patches into learned, content-dependent summaries, processing them centrally, and restoring features to image positions. Skip connections retain detail. On ImageNet after 600 epochs, HiT-XL/16 reaches FID 1.48, versus JiT-H/16’s 1.86, at similar parameter counts. This supports learned hierarchy in this comparison, but does not establish hierarchy as the sole cause or wall-clock speed.

Summary by alphaXiv.

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