Atlases Are Already Inside: Recovering Population Templates from Pretrained Diffusion Models

The paper introduces an inference-time sampler for pretrained diffusion models that extracts an intrinsic atlas of the population the model synthesizes, converging from any random seed to the population’s central anatomy. This method requires no retraining, works across multiple domains such as brain MRI, chest X-ray, faces, and 3D shapes, and can generate atlases for subpopulations like different ages. Evaluations show the intrinsic atlas performs as well or better than classical and learned templates for registration tasks and is the most central template for held‑out brain MRI cohorts.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Computer Vision.

• diffusion