Reproduce “sDMD: Sinkhorn Distribution Matching 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
Jaihoon Kim, Lingxiao Li, Soshi Shimada, Roshni Cooper, Minhyuk Sung, Yiwei Zhao
sDMD distills diffusion models into few-step generators by combining DMD’s local score guidance with a global Sinkhorn transport correction, applied when a discriminator indicates weak teacher support. In four-step Wan2.1-14B video generation, it scores 84.87 overall and 80.51 semantic on VBench, versus DMD2’s 84.69 and 76.75; its quality score is lower than DMD2’s.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.