Why diffusion model samples drift toward the dataset center at high noise levels
YouJiacheng · x · 2026-09-09
- A neat property of diffusion models: when a sample is mostly noise (early in generation), the model's best guess is close to the conditional mean.
- As a result, early samples get pulled toward the center of the dataset—even where no data actually exists—before rebounding toward the data manifold.
- YouJiacheng notes the mechanism: under high noise, the model can only predict the (conditional) expectation, which is the optimal denoising target.
Related event: Diffusion Models Pull Early Noisy Samples Toward Dataset Center(2 posts)→
More from Research
- Boaz Barak revisits 'Math after AI': he predicted this moment, just not this fast — soumitrashukla9 · 2026-09-09
- Mathematicians crowdsourcing a Lean formalization of the Hodge conjecture, a missing Millennium Problem — AlexKontorovich · 2026-09-09
- AI-designed anti-aging drug Rentosertib shows human data with six aging clocks pointing younger — rand_longevity · 2026-09-09
- Distill's 2017 classic revisited: interfaces into model representations to augment human intelligence — enjalot · 2026-09-09
- Millennium Prize déjà vu: Perelman solved Poincaré amid credit disputes, like OpenAI's Navier–Stokes now — keenanisalive · 2026-09-09
- Terence Tao warns AI is 'non-renewably' mining open math problems, threatening open science — Simon Willison · 2026-09-09