Reproduce “Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative 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
Jiongfan Yang, Weichen Fan, Haiwen Diao, Zhongang Cai, Ziwei Liu
SpecSlot assigns global communication to low-frequency image carriers while keeping finer output slots local. Each anchor produces multiple slots with disjoint spectral roles, but only one joins dense global attention. On ImageNet-512 it reaches 9.14 FID versus 11.98 for a FLOPs-matched JiT-CM control, with 3.9% fewer conditional forward FLOPs; it remains 0.69 FID worse than JiT-B/16.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.