Reproduce “RouteREPA: Routed Representation Alignment for Just Image 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
Yi Tan, Beier Zhu, Sirui Chen, Cheng Chen, Xingyu Zhu, Yuan Zhou, Zhenshan Tan, Tianrun Chen and 1 more
RouteREPA helps JiT generate images directly from RGB pixels by softly routing intermediate features into separate generation and vision-encoder alignment paths. This addresses demands of alignment and pixel reconstruction. On class-conditional ImageNet-1K at 256×256, JiT-H/16 reaches FID 4.96 after 200 epochs, versus PixelREPA’s 5.52 after 600 epochs, without classifier-free guidance; this shows an epoch advantage, not lower training time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.