Reproduce “Restoring Semantic Fidelity for Training-Free High-Resolution Image Generation in Diffusion 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

Junjie Li, Zengbin Wang, Gaoxiang Li, Muyi Sun, Yixin Zhang, Man Zhang

The method addresses semantic degradation in high-resolution text-to-image generation by keeping text attention global, restricting image attention to a training-scale neighborhood, and correcting the denoising trajectory at each step. Its guidance combines lower-resolution layout with high-resolution texture and evolving detail. On FLUX.1-dev at 4096×4096, it reports patch-level FID of 14.047 versus 18.468 for HiFlow, within tests at three resolutions.

Summary by alphaXiv.

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