Reproduce “Few-Step Diffusion via Score identity 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

Mingyuan Zhou, Yi Gu, Zhendong Wang

Researchers from the University of Texas at Austin introduce a framework to accelerate high-resolution Text-to-Image (T2I) diffusion models into efficient few-step generators. The method extends Score identity Distillation (SiD) with a Uniform-Step Matching strategy and incorporates Diffusion GAN-based adversarial learning and novel guidance techniques (Zero-CFG, Anti-CFG), achieving state-of-the-art FID and CLIP scores on SD1.5 and SDXL while significantly reducing generation steps.

Key insights

  • Uniform-Step Matching, with theoretical justification (Lemma 1), enables efficient multi-step distillation by allowing a single fake score network to learn from a uniform mixture of intermediate outputs across all generation steps.
  • Integrating a Diffusion GAN adversarial loss, even with limited real data, significantly boosts the diversity (FID) of distilled models, complementing SiD's data-free capabilities.
  • Zero-CFG and Anti-CFG provide flexible and memory-efficient alternatives to conventional CFG, effectively mitigating the alignment-diversity trade-off and offering pathways to optimize specific performance metrics.

Summary by alphaXiv.

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