Reproduce “Structural Self-Teaching for Compositional Generation in Unified Multimodal Models” 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

Jaeyoo Park, Bohyung Han

SST trains unified multimodal models to generate compositional images by using phrase-level spatial supports from their understanding pathway to guide generation with a training-only scaffold. Compared with LoRA fine-tuning on the same data, it improves AM and GM across four backbones; on BLIP-3o 4B, GM rises from 13.00 to 16.13. The scaffold is removed at inference, but high-atomicity generation remains difficult.

Summary by alphaXiv.

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