Constrained Diffusion Keeps Sampling Feasible

nandofioretto · x · 2026-07-10

The post points out that diffusion model sampling rarely lands naturally on hard constraint sets, as the score function itself has no incentive to respect boundaries.

To address this, the author proposes projecting the trajectory onto a feasible manifold during the reverse diffusion step, keeping the entire sampling process feasible. A related course on constrained diffusion is also mentioned.

Original post →

More from Research

Research channel →