Joint Super-Resolution and Spectral Reconstruction Via Linear Initialization of a Pretrained RGB Model

Hyperspectral reconstruction from low-resolution RGB images is a challenging task that requires joint spatial and spectral recovery. This paper presents an efficient method for the 2026 ICASSP Hyper-Object Challenge (Track 2), involving 2× spatial super-resolution and 61-band reconstruction. Our approach adapts the HAT backbone, pretrained for RGB super-resolution. We modify the output layer to produce 61 bands, initializing the new weights via a novel strategy based on linear RGB-to-band priors. This yields a zero-shot model with high spatial fidelity that accelerates fine-tuning, converging in roughly 55 minutes. By incorporating an MST++ module to marginally refine band accuracy, we provided a final boost that secured one of the winning scores.