sDMD adds adaptive Sinkhorn correction to diffusion distillation
Original title: sDMD: Sinkhorn Distribution Matching Distillation
With 2–4 steps on FLUX.1-dev, Qwen-Image-20B, and Wan2.1-14B, authors report gains over DMD2 and other baselines.
The paper presents sDMD, a variant of distribution matching distillation that adds an optimal-transport term estimated by Sinkhorn divergence. The correction is applied adaptively from discriminator outputs so it mainly moves outlier samples, where reverse-KL guidance can be unreliable or mode-seeking. On FLUX.1-dev, Qwen-Image-20B, and Wan2.1-14B with 2–4 sampling steps, the authors say sDMD beats distillation baselines on GenEval, DPG, WISE, and VBench and improves diversity over DMD2.
Unconfirmed: Benchmark wins and diversity gains are reported by the authors only.
- sDMD augments integral KL with a batch-estimated Sinkhorn term.
- The Sinkhorn correction is gated by discriminator outputs.
- Evaluation uses 2–4 steps on FLUX.1-dev, Qwen-Image-20B, and Wan2.1-14B.
- Reported benchmarks are GenEval, DPG, WISE, and VBench.
- Authors say sample diversity improves over DMD2.
From the original article
1 min read at arxiv.orgDistribution Matching Distillation (DMD) distills multi-step diffusion models into few-step generators by minimizing reverse KL divergence.
By Jaihoon Kim · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.