SGPnP and SNaP: making classical image reconstruction methods generative with diffusion priors

prof_kamilov · x · 2026-11-04

Ulugbek Kamilov's group presents two works that turn classical image reconstruction generative. SGPnP embeds pretrained score-based diffusion models directly as priors inside Plug-and-Play algorithms without reverse-diffusion sampling: injecting noise before denoising matches the score model's training regime, smooths the objective to avoid strict saddle points, and with annealing converges to stationary points. It works with ADMM, HQS, and PGM-style updates, with decoupled noise levels, outperforming standard PnP baselines on large-mask inpainting and multi-coil MRI while staying competitive with diffusion solvers. SNaP upgrades feedforward U-Net reconstruction to enable one-step posterior sampling.

Original post →

More from Multimodal

Multimodal channel →