Denoising as Projection: Constrained Optimization with Gradient-Guided Diffusion
The paper treats denoising as an approximate projection step for keeping guided diffusion on a learned feasible geometry.
Zhang and coauthors propose putting the external objective gradient inside the denoising update, using only a pretrained denoiser plus gradient evaluations at inference time. They analyze the method as an inexact projected-gradient scheme for constrained optimization. The guarantees cover linear manifolds, compact convex sets, and compact Riemannian submanifolds, with descent and finite-time convergence results. Experiments are described as supporting the view that the update trades off objective improvement against staying on the learned geometry. ArXiv · AI/CL/LG's note
Zhang and coauthors propose putting the external objective gradient inside the denoising update, using only a pretrained denoiser plus gradient evaluations at inference time. They analyze the method as an inexact projected-gradient scheme for constrained optimization. The guarantees cover linear manifolds, compact convex sets, and compact Riemannian submanifolds, with descent and finite-time convergence results. Experiments are described as supporting the view that the update trades off objective improvement against staying on the learned geometry. ArXiv · AI/CL/LG's note
score 4