2602.19736

Total: 1

#1 InfScene-SR: Spatially Continuous Inference for Arbitrary-Size Image Super-Resolution [PDF] [Copy] [Kimi] [REL]

Authors: Shoukun Sun, Zhe Wang, Xiang Que, Jiyin Zhang, Xiaogang Ma

Image Super-Resolution (SR) aims to recover high-resolution (HR) details from low-resolution (LR) inputs, a task where Denoising Diffusion Probabilistic Models (DDPMs) have recently shown superior performance compared to Generative Adversarial Networks (GANs) based approaches. However, standard diffusion-based SR models, such as SR3, are typically trained on fixed-size patches and struggle to scale to arbitrary-sized images due to memory constraints. Applying these models via independent patch processing leads to visible seams and inconsistent textures across boundaries. In this paper, we propose InfScene-SR, a framework enabling spatially continuous super-resolution for large, arbitrary scenes. We adapt the iterative refinement process of diffusion models with a novel guided and variance-corrected fusion mechanism, allowing for the seamless generation of large-scale high-resolution imagery without retraining. We validate our approach on remote sensing datasets, demonstrating that InfScene-SR not only reconstructs fine details with high perceptual quality but also eliminates boundary artifacts, benefiting downstream tasks such as semantic segmentation.

Subject: Computer Vision and Pattern Recognition

Publish: 2026-02-23 11:34:59 UTC