Diffusion model collapse isn't all bad: new sampler extracts population atlases without retraining
kwangmoo_yi · x · 2026-09-29
A new arXiv paper (Shi, Femiani, Wonka), "Atlases Are Already Inside," introduces an inference-time sampler that gives pretrained diffusion models a capability they were never trained for: constructing the atlas of the population they synthesize.
- No retraining: any diffusion model that has learned a coherent population (including released ones) yields its intrinsic atlas in a single inference pass, without deformable registration
- Multi-domain: works on brain MRI, chest X-ray, faces, and 3D shapes
- Subpopulations: an age-conditioned model yields atlases at any age in its training range, reproducing CSF expansion of healthy aging
Evaluated as a registration target, the intrinsic atlas is best or second-best on every dataset against classical and learned templates, and the most central template on held-out brain MRI cohorts. The paper reframes atlas construction as a byproduct of generative modeling.
Related event: Deliberately Collapsing Diffusion Models Recovers Population Atlases(2 posts)→
More from Multimodal
- NUS rethinks DiT residual connectivity: 1.73x fewer training iterations, 1.39 FID — NationalUniversityofSingapore · 2026-09-29
- Seedance 2.5 makes fake street-footage viral videos: fat cat on scooter prompt shared — LudovicCreator · 2026-09-29
- Character Swap LoRA fixes MiniMax H3 video consistency, tests show 88% denoise sweet spot — not_food · 2026-09-29
- Open-source ComfyUI enhancer node visibly improves Krea2 turbo 8-step image quality — No-Bad-4838 · 2026-09-29
- ControlNet Depth with end_percent below 0.5 produces surprisingly good results — MaleMaldives · 2026-09-29
- AI-Generated 'Harry Potter and the Order of the GYM' Goes Viral on Reddit — iquizuanswer · 2026-09-29