Reproduce “Spectral Amplitude Purification in Distribution Matching for Diffusion Distillation” 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
Zhenyu Zhou, Can Wang, Chun Chen, Zeyu Zheng, Defang Chen
The authors propose SAP-DMD to improve few-step text-to-image diffusion distillation by purifying the DMD update in the Fourier domain. It adaptively clips unusually large spectral amplitudes while preserving phase, rather than changing the model or sampling steps. On SD3.5 Medium, 2-step HPSv2.1 averaged 32.67 versus 31.27 for DMD2, with faster spectral recovery, but complex semantic detail remains challenging.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.