Diffusion models have demonstrated great potential in generating realistic image details. However, existing diffusion models are primarily trained on natural images, making their application to remote sensing image super-resolution highly challenging. Moreover, these models typically require dozens or even hundreds of iterative sampling steps during inference, resulting in high computational costs and limited practicality. To address these issues, this paper proposes a degradation-aware adaptive estimation-based single-step remote sensing image super-resolution diffusion model (RS-AdaDiff), which balances reconstruction performance and inference efficiency. Specifically, we propose a degradation-aware timestep estimation module that adaptively estimates the diffusion timestep for the diffusion model by assessing the degradation level of the input image. This approach reconstructs the iterative denoising process into a single-step reconstruction from low-resolution to high-resolution images, thereby significantly accelerating inference. Meanwhile, we integrate trainable lightweight LoRA layers into a pre-trained diffusion model and fine-tune it on a remote sensing image dataset to mitigate the domain gap caused by data distribution differences. Additionally, to fully leverage the image priors of the pre-trained model, we introduce distribution contrastive matching distillation. By regularizing the KL divergence, the reconstructed super-resolved images are brought closer to high-resolution images and farther from low-resolution images in the feature space, thereby improving generation quality. Finally, we propose a feature-edge joint perceptual similarity loss to enhance the perception of structural information and mitigate issues such as edge blur and texture distortion. Extensive experimental results demonstrate that the proposed RS-AdaDiff outperforms existing state-of-the-art methods on multiple public remote sensing datasets, achieving significant improvements in both quantitative metrics and visual quality, and producing super-resolved remote sensing images with clearer structures and richer details.

Existing dehazing methods face challenges in generalization due to the lack of paired real-world training data and tailored models. Recently, some semi-supervised/unsupervised schemes have been explored, achieving impressive performance. However, their performance still depends heavily on synthetic training data and the introduced prior-based strong constraints do not always hold. In this paper, we first introduce RealHQ-HAZE, a new dataset with 200 collected real-world hazy images, 200 corresponding carefully rendered haze-free images, and an additional 1000 varicolored hazy images transferred from the collected images. We also propose a prior-compensated multi-stage dehazing network, PMDN, which can learn different levels of real-world haze distribution through multi-stage progressive learning. To utilize prior knowledge effectively, we introduce a prior-based feature compensation module, guiding intermediate results with an adaptive weight. Additionally, we propose a MixCut consistent dehazing strategy to mix paired and derived images using a cross-cutting scheme, reinforcing dehazing through consistency principles. Extensive experiments demonstrate the effectiveness of our dataset and the superiority of PMDN compared to existing state-of-the-art dehazing methods.