Reproduce “Rethinking the design space of Vanilla Pixel 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

Soumik Mukhopadhyay, Long Mai, Abhinav Shrivastava, Aniruddha Mahapatra

MFLOW adapts pixel-diffusion prediction targets across data directions and adds a convolutional refiner for fine detail. It learns continuous gains between clean-image and velocity prediction, then trains against recovered velocity; the refiner is trained separately with the diffusion backbone frozen. On ImageNet-1K at 256×256, MFLOW-B/16 reaches FID 3.33 with LPIPS, but does not lead across settings. The refiner narrows, but does not eliminate, the high-frequency deficit.

Summary by alphaXiv.

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