| Title |
Trust-region Diffusion Models with Entropic Constraints for Zero-shot Inverse Problems |
| Authors |
이환(Hwan Lee) ; 임홍기(Hong-ki Lim) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.1.29 |
| Keywords |
Inverse problem; Diffusion model; KL-divergence |
| Abstract |
Recently, research has been conducted on inverse problem reconstruction using posterior sampling, which combines the likelihood term and prior distribution derived from Bayes' theorem. However, this approach faces challenges due to the step size and the accuracy of the degradation operator. This paper proposes A Trust-RegIon Diffusion Sampler with an ENTropic Boundary (TRIDENT) to overcome the limitations of existing diffusion-based techniques for inverse problem reconstruction. TRIDENT introduces a trust region based on Kullback-Liebler (KL) divergence, constraining each likelihood update to remain within the dense region of the prior distribution. This ensures zero-shot operator invariance and stable convergence. Various experiments demonstrate improved reconstruction performance and perceptual quality compared to existing methods. |