Reflected Diffusion Models: Principled Data Constraints in Generative Modeling

Explore reflected diffusion models in this 55-minute talk by Aaron Lou from Valence Labs. Learn how these models address numerical error issues in complex tasks by reversing a reflected stochastic differential equation evolving on the data support. Discover the generalized score matching loss approach for learning perturbed score functions and extensions to key diffusion model components. Examine the theoretical connection between reflected SDEs and thresholding schemes. Review competitive results on image benchmarks and the benefits for classifier-free guidance, including fast exact sampling with ODEs and improved sample fidelity under high guidance weight. Dive into topics like diffusion model recaps, divergent sampling, forward and reverse reflected SDEs, CIFAR-10 image generation, and maximum likelihood training through detailed chapters and a Q&A session.

- Intro & Diffusion Models Recap
- Divergent Sampling & Thresholding
- Forward + Reverse Reflected SDEs
- CIFAR-10 Image Generation
- Maximum Likelihood Training & Results
- Q+A