Reproduce “UniPC: A Unified Predictor-Corrector Framework for Fast Sampling of Diffusion Models” 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
Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, Jiwen Lu
Tsinghua University researchers developed UniPC, a unified predictor-corrector framework for Diffusion Probabilistic Models that achieves high-quality image generation with extremely few function evaluations. UniPC consistently outperforms strong baselines like DPM-Solver++ across various datasets and sampling scenarios, significantly reducing the required steps while maintaining computational efficiency.
Key insights
- The Unified Corrector (UniC) can increase the order of accuracy of any existing p-order DPM sampler by one, without incurring extra function evaluations in the current step by reusing model outputs for the next step.
- A unified analytical form for both predictors (UniP) and correctors (UniC) enables the derivation of arbitrary-order solvers, simplifying implementation and analysis compared to previous ad-hoc derivations.
- The choice of the B(h) function (e.g., B1(h)=h vs. B2(h)=e^h-1) and the specific order schedule significantly impact performance, particularly at very low NFE.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.