Coloring the Noise: Adversarial Sobolev Alignment for Faithful Image Super Resolution
The paper presents a new framework called ASASR for image super-resolution that addresses the limitations of existing generative methods. It focuses on aligning the generative flow with the natural image manifold by using a Sobolev-induced Riemannian geometry. The results indicate that ASASR significantly improves spectral consistency and structural fidelity compared to leading generative baselines.
- ▪ASASR aims to overcome the spectral misalignment in image super-resolution.
- ▪The framework utilizes a parametric adversary based on the Riesz Representation Theorem.
- ▪Extensive evaluations show that ASASR outperforms existing generative methods in preserving image quality.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computer Vision and Pattern Recognition arXiv:2605.23264 (cs) [Submitted on 22 May 2026] Title:Coloring the Noise: Adversarial Sobolev Alignment for Faithful Image Super Resolution Authors:Hongbo Wang, Huaibo Huang, Pin Wang, Jinhua Hao, Chao Zhou, Ran He View a PDF of the paper titled Coloring the Noise: Adversarial Sobolev Alignment for Faithful Image Super Resolution, by Hongbo Wang and 5 other authors View PDF HTML (experimental) Abstract:Generative priors in Image Super-Resolution (SR) often compromise faithful restoration, we attribute this limitation to a fundamental spectral misalignment between isotropic objectives and the intrinsic natural image manifold.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
