Compositional Generation Failures in Diffusion Models Traced to Samplers

AI researcher Mark Neumann argues that diffusion models' compositional generation failures stem from the sampler rather than the model, and that MCMC-style samplers can fix it. He also shared references on conservative diffusion models and raised the question of scoring generated samples with one's own model in scientific research.

2026-09-17 ~ 2026-09-17 · 2 related posts

• Why compositional diffusion fails: sampler, not model — MCMC samplers fix it — MarkNeumannnn · 2026-09-17
• Scientific generative models: scoring samples under your own model matters most — MarkNeumannnn · 2026-09-17