InfScene-SR: Arbitrary-Size Image Super-Resolution via Iterative Joint-Denoising

Abstract

While diffusion models have achieved state-of-the-art performance in Image Super-Resolution (SR), their prohibitive computational and memory demands restrict their training and inference to fixed-size inputs. The standard workaround to super-resolve larger images relies on partitioning the image, super-resolving patches independently, and stitching them together -- a process that inevitably introduces severe boundary artifacts and spatial inconsistencies in large-scale scenes. To achieve spatially continuous, arbitrary-size image super-resolution, we propose InfScene-SR, a diffusion-based SR approach. Building upon SR3, our approach leverages Variance-Corrected Fusion (VCF) to perform joint-denoising across overlapping patches. VCF guarantees continuous transitions while preserving the stochastic variance crucial for high-fidelity texture reconstruction. To overcome the prohibitive synchronization overhead of scaling joint-denoising to gigapixel imagery, we introduce Spatially-Decoupled Variance Correction (SDVC). SDVC reformulates the global fusion process into independent, atomic patch operations, drastically reducing memory complexity to O(1) and naturally enabling fully distributed, parallelized inference. Extensive experiments on large-scale remote sensing datasets demonstrate that InfScene-SR strictly eliminates boundary seams, achieves superior perceptual quality, and significantly boosts performance in downstream semantic segmentation task.

Reproductions