Rethinking Pan-sharpening: Principled Design, Unified Training, and a Universal Loss Surpass Brute-Force Scaling

Abstract: The field of pan-sharpening has recently seen a trend towards increasingly large and complex models, often trained on single, specific satellite datasets. This approach, however, leads to high computational overhead and poor generalization on full resolution data, a paradigm we challenge in this paper. In response to this issue, we propose PanTiny, a lightweight, single-step pan-sharpening framework designed for both efficiency and robust performance. More critically, we introduce multiple-in-one training paradigm, where a single, compact model is trained simultaneously on three distinct satellite datasets (WV2, WV3, and GF2) with different resolution and spectral information. Our experiments show that this unified training strategy not only simplifies deployment but also significantly boosts generalization on full-resolution data. Further, we introduce a universally powerful composite loss function that elevates the performance of almost all of models for pan-sharpening, pushing state-of-the-art metrics into a new era. Our PanTiny model, benefiting from these innovations, achieves a superior performance-to-efficiency balance, outperforming most larger, specialized models. Through extensive ablation studies, we validate that principled engineering in model design, training paradigms, and loss functions can surpass brute-force scaling. Our work advocates for a community-wide shift towards creating efficient, generalizable, and data-conscious models for pan-sharpening. The code is available at https://github.com/Zirconium233/PanTiny .