Multimodal large language models meet self-supervised diffusion for real-world aerial image super-resolution

Abstract

Real-world aerial image super-resolution (SR) remains particularly challenging because degradations in remote-sensing imagery involve random combinations of anisotropic blur, signal-dependent noise, and unknown downsampling kernels. Most existing SR methods either rely on simplified degradation assumptions or lack semantic perception of degradation, resulting in limited generalization to real-world conditions. To address these gaps, we propose a novel diffusion-based SR framework that integrates Multi-modal Large Language Models (MLLMs) and self-supervised contrastive learning for extracting degradation-insensitive representation. Specifically, we introduce a contrastive learning strategy into a ControlNet module, where the HR and LR counterparts of the same image are regarded as positive pairs, while representations from different images serve as negative pairs, enabling the network to learn degradation-insensitive structural features. To further enhance semantic awareness of degradation, an MLLM-generated change caption is incorporated into the diffusion process as textual guidance, allowing the model to explicitly perceive and reconstruct different degradation types. Moreover, a classifier-free guidance (CFG) distillation strategy compresses the original dual-branch diffusion model into a single lightweight network, substantially improving inference efficiency while maintaining high reconstruction fidelity. Extensive experiments conducted on various datasets have showcased the superior performance of our proposed model compared to existing state-of-the-art methods. Furthermore, our distillation algorithm achieves a twofold reduction in inference time compared to its non-distilled counterpart, making it more feasible for real-time and resource-limited applications.