Zigzag Diffusion Sampling: Diffusion Models Can Self-Improve via Self-Reflection

Authors: Lichen Bai, Shitong Shao, zikai zhou, Zipeng Qi, Zhiqiang Xu, Haoyi Xiong, Zeke Xie

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Third, our extensive experiments demonstrate that Z-Sampling can generally and significantly enhance generation quality across various benchmark datasets, diffusion models, and performance evaluation metrics. For example, Dream Shaper with Z-Sampling can self-improve with the HPSv2 winning rate up to 94% over the original results. Moreover, Z-Sampling can further enhance existing diffusion models combined with other orthogonal methods, including Diffusion-DPO. The code is publicly available at github.com/xie-lab-ml/Zigzag-Diffusion-Sampling.
Researcher Affiliation Collaboration Lichen Bai1 Shitong Shao1 Zikai Zhou1 Zipeng Qi2 Zhiqiang Xu3 Haoyi Xiong4 Zeke Xie1 1x Lea F Lab, The Hong Kong University of Science and Technology (Guangzhou) 2Beihang University 3Mohamed bin Zayed University of Artificial Intelligence 4Baidu Inc
Pseudocode Yes Algorithm 1 Z-Sampling
Open Source Code Yes The code is publicly available at github.com/xie-lab-ml/Zigzag-Diffusion-Sampling.
Open Datasets Yes Datasets Pick-a-Pic (Kirstain et al., 2023), Draw Bench dataset (Saharia et al., 2022), and Gen Eval (Ghosh et al., 2024). We leave more details in Appendix A.1.
Dataset Splits Yes Here we use only the first 100 prompts as the test set, which is sufficient to reflect the model s capabilities.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU models, CPU types, memory amounts) used for running its experiments. It mentions computational costs but not the hardware specifications.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch versions) needed to replicate the experiment.
Experiment Setup Yes For SD2.1 (Rombach et al., 2022), SDXL (Podell et al., 2023), and Hunyuan-Di T (Li et al., 2024), we perform 50 denoising steps. For Dream Shaper-xl-v2-turbo, which achieves efficient and high-quality generation by fine-tuning SDXL Turbo (Sauer et al., 2023), we set denoising step T only to 4. And we set γ1 = 5.5 in SDXL/SD2.1, γ1 = 6.0 in Hunyuan-Di T, and γ1 = 3.5 in Dream Shaper-xl-v2-turbo, all to the recommended default values. We set the zigzag operation to be executed throughout the entire path (λ = T 1) and inversion guidance scale γ2 as zero, unless we specify them otherwisely.