NeurIPS 2026 paper shows noise-conditioning-free diffusion models implicitly implement Riemannian gradient flow

docmilanfar · x · 2026-09-25

The author's sole submission to NeurIPS 2026 has been accepted. The paper tackles a counterintuitive finding: diffusion models typically require exact noise level schedules, yet some recent models work without explicit noise conditioning. The authors show that these time-invariant fields implicitly implement a Riemannian gradient flow on an energy landscape, explaining how denoising succeeds without explicit time/noise conditioning.

Original post →

More from Research

Research channel →