| Title |
Stein’s Unbiased Risk Estimation based Posterior Sampling: Trajectory Correction for Diffusion based Inverse Problems |
| Authors |
김민우(Minwoo Kim) ; 임홍기(Hongki Lim) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.3.119 |
| Keywords |
Diffusion model; Stein’s Unbiased risk estimate; Inverse problem |
| Abstract |
Diffusion models have emerged as powerful learned priors for solving inverse problems. However, current iterative approaches that alternate between diffusion sampling and data-consistency steps often require hundreds to thousands of iterations to achieve high-quality reconstructions due to accumulated errors. To address this issue, we propose a method that corrects deviations of the sampling trajectory using gradient updates derived from Stein’s Unbiased Risk Estimate (SURE) and PCA-based noise estimation. By mitigating no ise-induced errors during the critical early and middle stages of sampling, the method enables more accurate posterior sampling and reduces error accumulation. Consequently, it maintains high reconstruction quality with fewer than 100 Neural Function Evaluations (NFEs). Extensive evaluations across diverse inverse problems show that the method consistently outperforms existing approaches in the low-NFE regime. |