NeurIPS paper: diffusion models interpolate well but rarely extrapolate skills

liuzhuang1234 · x · 2026-11-19

A NeurIPS 2026 paper, "Do Diffusion Models Learn to Generalize Basic Visual Skills?", asks whether diffusion models generalize skills beyond training data or mostly interpolate. Since this can't be tested on today's text-to-image models (unknown training data), the authors trained small diffusion models from scratch on synthetic images with full control over every value seen. Finding: they interpolate well but rarely extrapolate.

Original post →

More from Research

Research channel →