Approximate Maximum a Posteriori Training of Diffusion Models for Image Restoration
2026 (English)In: IEEE Access, E-ISSN 2169-3536, Vol. 14, p. 102614-102634
Article in journal (Refereed) Published
Abstract [en]
Inverse problems such as super-resolution, deblurring, and inpainting are commonly solved using iterative Maximum a Posteriori (MAP) solvers with pretrained priors. Recently, Diffusion Models (DMs) have achieved state-of-the-art restoration quality through a generative sampling process conditioned on the degraded observation, leveraging pretrained priors similar to those used in MAP estimation. Motivated by this connection, previous work has trained DMs as deep regularizers in an iterative MAP solver, optimizing them on the mean squared error (MSE) between the estimate and the ground truth. However, MSE trains an estimator to seek the posterior mean rather than the mode, introducing a mismatch with MAP estimation that can result in blurry images. In this work, we present an extended investigation of such MAP training for DMs and introduce a loss designed to address the mismatch, providing a theoretical justification for its use in MAP training. Our experiments show that MAP training with MSE or the proposed loss improves robustness to hyperparameter selection and helps reduce restoration failures in certain cases. Furthermore, we observe that the proposed loss reduces training time and can produce sharper restorations with fewer artifacts when compared to MAP training using MSE.
Place, publisher, year, edition, pages
Institute of Electrical and Electronics Engineers (IEEE), 2026. Vol. 14, p. 102614-102634
Keywords [en]
Training, Modeling, Inverse problems, Lamps, Noise reduction, Superresolution, Deblurring, PSNR, Image restoration, Weighted sum model, Deep equilibrium models, diffusion models, image restoration, inverse problems, maximum a posteriori estimation
National Category
Computer Vision and Learning Systems
Identifiers
URN: urn:nbn:se:miun:diva-58204DOI: 10.1109/access.2026.3708219ISI: 001821036400044Scopus ID: 2-s2.0-105043732011OAI: oai:DiVA.org:miun-58204DiVA, id: diva2:2089674
Funder
Interreg Aurora, 20366448Mid Sweden University, 2020/14132026-08-042026-08-042026-09-01Bibliographically approved