Reproduce “Formulating Discrete Probability Flow Through Optimal Transport” 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
Pengze Zhang, Hubery Yin, Chen Li, Xiaohua Xie
Researchers from Sun Yat-sen University and WeChat, Tencent Inc. established a discrete probability flow for discrete diffusion models by applying optimal transport theory, which significantly reduces sampling uncertainty by over 80% on both synthetic and image datasets while maintaining generation quality.
Key insights
- Optimal transport theory provides a suitable mathematical framework for conceptualizing and defining deterministic-like probability flow in both continuous and discrete diffusion processes.
- By designing a discrete diffusion generator that enforces unidirectional probability flow towards higher density states, a discrete optimal transport plan can be established, bridging a theoretical gap with continuous models.
- Leveraging this discrete probability flow in sampling enables a significant reduction in generational uncertainty, allowing for more controlled and predictable outputs from discrete diffusion models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.