Reproduce “Zigzag Diffusion Sampling: Diffusion Models Can Self-Improve via Self-Reflection” 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
Lichen Bai, Shitong Shao, Zikai Zhou, Zipeng Qi, Zhiqiang Xu, Haoyi Xiong, Zeke Xie
ZIGZAGDIFFUSIONSAMPLING introduces a training-free and plug-and-play method for diffusion models that improves image generation quality and prompt alignment by leveraging a "guidance gap" between denoising and inversion. The method achieved up to a 94% HPSv2 winning rate over original results on DrawBench and significantly enhanced handling of complex compositional attributes in generated images.
Key insights
- Latent representations inherently carry semantic information, and this information can be explicitly captured and injected into the sampling process.
- A "guidance gap"—using strong guidance during denoising and weak (or zero) guidance during inversion—effectively injects prompt-related semantic information into the latent space.
- A step-by-step "zigzag" self-reflection process, alternating denoising and inversion, more effectively accumulates semantic information and mitigates approximation errors compared to a single end-to-end inversion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.