Why compositional diffusion fails: sampler, not model — MCMC samplers fix it

MarkNeumannnn · x · 2026-09-17

Mark Neumann shared two references on "conservative diffusion": the 2023 arXiv paper Reduce, Reuse, Recycle (Yilun Du, Sander Dieleman, Jascha Sohl-Dickstein, Will Grathwohl et al.), which shows the sampler—not the model—is responsible for compositional generation failures and proposes MCMC-inspired, energy-based parameterizations with Metropolis-corrected samplers that markedly improve compositional generation on ImageNet-style tasks; plus an apparently unpublished discussion of why this isn't done in practice for images.

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

Original post →

More from Research

Research channel →