KAO: Kernel-Adaptive Optimization in Diffusion for Satellite Image

TL;DR

KAO introduces Kernel-Adaptive Optimization within diffusion models to perform high-fidelity satellite image inpainting on very high-resolution data. By combining latent-space conditioning with Explicit Propagation and integrating with a Token Pyramid Transformer, KAO adaptively modulates diffusion kernels and propagates information across scales to prioritize structurally important regions. Empirical results on Massachusetts Roads and DeepGlobe show state-of-the-art performance in FID, precision, and recall, while reducing computational cost compared to baselines. The method demonstrates robust restoration under cloud/mist occlusions and offers practical strategies for efficient deployment in remote sensing workflows. This work advances satellite image restoration by uniting kernel-adaptive denoising with hierarchical latent representations for scalable, high-quality inpainting.

Abstract

Satellite image inpainting is a crucial task in remote sensing, where accurately restoring missing or occluded regions is essential for robust image analysis. In this paper, we propose KAO, a novel framework that utilizes Kernel-Adaptive Optimization within diffusion models for satellite image inpainting. KAO is specifically designed to address the challenges posed by very high-resolution (VHR) satellite datasets, such as DeepGlobe and the Massachusetts Roads Dataset. Unlike existing methods that rely on preconditioned models requiring extensive retraining or postconditioned models with significant computational overhead, KAO introduces a Latent Space Conditioning approach, optimizing a compact latent space to achieve efficient and accurate inpainting. Furthermore, we incorporate Explicit Propagation into the diffusion process, facilitating forward-backward fusion, which improves the stability and precision of the method. Experimental results demonstrate that KAO sets a new benchmark for VHR satellite image restoration, providing a scalable, high-performance solution that balances the efficiency of preconditioned models with the flexibility of postconditioned models.

Figures (8)

• Figure 1: Qualitative comparison of satellite image inpainting across seven models. Each column represents a different method: the occluded input, ground truth target, Stable Diffusion rombach2022high, RePaint lugmayr2022repaint, SatDiff panboonyuen2025satdiff, DPS chung2022diffusion, PSLD rout2023solving, and KAO (Ours). Each row shows a different satellite image sample with varying occlusion patterns and object structures. KAO consistently produces more accurate and visually realistic reconstructions, closely aligning with the ground truth. This demonstrates a deeper understanding of spatial context and structure, enabling faithful restoration of occluded regions in real-world satellite imagery.
• Figure 2: Qualitative comparison of satellite image inpainting across six diverse scenes (columns), each representing varying geographic contexts such as urban and agricultural regions. The first row shows input satellite images with occluded areas (masks), followed by rows 2–7 depicting outputs from state-of-the-art methods: Stable Diffusion rombach2022high, RePaint lugmayr2022repaint, SatDiff panboonyuen2025satdiff, DPS chung2022diffusion, PSLD rout2023solving, and our proposed method KAO. KAO consistently delivers reconstructions that are sharper, structurally coherent, and contextually aligned with real-world features—effectively recovering linear structures like roads in urban scenes and preserving texture continuity in agricultural areas. This highlights KAO’s enhanced spatial understanding and its advantage in producing high-fidelity inpaintings under complex terrain and occlusion patterns.
• Figure 3: Qualitative results on three agricultural satellite image samples, comparing inpainting performance across seven models. Each column corresponds to one model: occluded input, ground truth, Stable Diffusion rombach2022high, RePaint lugmayr2022repaint, SatDiff panboonyuen2025satdiff, DPS chung2022diffusion, PSLD rout2023solving, and our method KAO. Each row depicts a distinct occlusion scenario within farmland or cultivated landscapes. Competing methods often fail to restore fine-grained plot patterns or introduce unrealistic textures in structured agricultural layouts. In contrast, KAO consistently reconstructs terrain boundaries, vegetation rows, and irrigation lines with high fidelity—producing outputs that align visually and semantically with the real-world cultivated geometry. These examples validate KAO’s ability to reason about spatial context and agricultural structure, making it highly effective for earth observation tasks such as crop monitoring and rural land-use mapping.
• Figure 4: KAO (our proposed method) as an easy add-on to a diffusion model. With this addition, a pretrained unconditional diffusion model is conditioned for inpainting. KAO can be seamlessly integrated into any Token Pyramid Transformer (TPT) diffusion model to perform high-quality inpainting.
• Figure 5: Illustration of how KAO is applied twice within the existing structure of a diffusion TPT model. The method is introduced between the Input, Middle, and Output blocks, enhancing the model’s inpainting capabilities. The proposed approach is detailed in Algorithm \ref{['alg:kaoalgo']}.