Scientific generative models: scoring samples under your own model matters most

MarkNeumannnn · x · 2026-09-17

In a discussion with @DdelAlamo, researcher Mark Neumann argues that being able to score samples under your own model is more important in scientific domains, so some points from related papers are less relevant there. He also shares two references: a conservative diffusion model for images, and an (apparently unpublished) discussion of why this approach isn't used in practice, at least for images.

Related event: Compositional Generation Failures in Diffusion Models Traced to Samplers(2 posts)→

Original post →

More from Research

Research channel →