Diffusion Models for Hyperspectral Image Superresolution: From noise prediction to direct image prediction

Diffusion models have recently attracted increasing attention in hyperspectral image super-resolution owing to their ability to model complex data distributions. However, most existing diffusion-based methods formulate hyperspectral image super-resolution as a noise prediction problem. In these methods, the high-resolution image is recovered through an indirect denoising target. This formulation deviates from the true goal of reconstructing a high-resolution hyperspectral image. This issue is particularly critical for hyperspectral images, which are characterized by high spectral dimensionality and strong inter-band correlations. Under this condition, using noise as the direct network prediction target makes the reconstruction objective indirect with respect to the target high-resolution hyperspectral image and may weaken the explicit preservation of spatial–spectral structures during reconstruction. To address this issue, we develop a direct image prediction paradigm for hyperspectral image super-resolution. This paradigm operates directly on the hyperspectral image distribution. It not only breaks the long-standing reliance of diffusion-based hyperspectral reconstruction on noise prediction, but also allows diffusion modeling to better match the physical characteristics and spatial–spectral structures of hyperspectral data. Based on this paradigm, we develop an implementation framework composed of the spectral-preserving embedding module, spatial–spectral transformer encoding module, and explicit spectral reconstruction module. We evaluate the framework on both remote sensing imagery and laboratory scenes. Experimental results show that the proposed framework outperforms state-of-the-art methods in both quantitative metrics and visual quality. This work provides a physically meaningful diffusion modeling perspective for hyperspectral image super-resolution. It also establishes a novel connection between generative diffusion modeling and hyperspectral data characteristics. This paradigm offers a potential direction for high-fidelity hyperspectral image reconstruction. The source code has been made publicly accessible at https://gitee.com/YinghaoXU/DPCHSISR