Reproduce “HiUni: Harmonizing Understanding and Generation in Shared Visual Space” 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
Jiahao Hu, Jingdong Du, Yichen Zhang, Yi Song, Zhenyu Zhang, Jiaxin Niu, Ruihang Li, Xiangzhao Hao and 5 more
HiUni tests whether one vision-language model can understand and generate images using shared visual and dense Transformer backbones. It adds a generation-only Pixel Diffusion Head and a HiGate route that supplies raw pixel features near the output. In a controlled comparison, Dense-2B nearly matches MoT-2B on understanding (58.55 vs. 59.23) but scores lower on OneIG (0.291 vs. 0.345); this supports coexistence, not consistent cross-task synergy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.