New Theory for Diffusion Training: Optimal Noise Sampling at Discrete Levels

LucaAmb · x · 2026-08-11

The author shares a theoretical study on optimal training for diffusion models: From Atoms to Entropy: Optimal Noise Allocation for Diffusion Training in the Convex Regime.

The theory expresses the optimization problem through its spatio-temporal Fisher geometry in the convex regime. It reaches a striking conclusion: under optimal training strategies, sample noise should be allocated at discrete noise levels, and the optimal schedule possesses a finite set of atoms.

Related event: New Theory Proposes Optimal Noise Allocation for Diffusion Model Training(2 posts)→

Original post →

More from Research

Research channel →