Pansharpening aims to fuse high-resolution panchromatic and low-resolution multispectral images to obtain high-resolution multispectral images. In this paper, an efficient general training framework based on zero-shot learning is proposed for this task. Early deep learning methods use a resolution downgrading process to generate suitable training samples, but the performance of models trained at a reduced scale on the target resolution is still an open issue. To address the inherent scale-variation problem, the proposed framework employs joint training, enabling single-stage cross-scale training while ensuring high-quality fusion performance at both full and reduced resolutions. The framework is compatible with various backbones and consistently enhances their performance. Furthermore, to support the cross-scale framework, a novel and simple reference-free quality loss function is proposed. Finally, extensive experiments are conducted on the QuickBird, GaoFen-2, and WorldView-2 datasets based on three types of classical deep learning pansharpening architectures. The experimental results demonstrate that the proposed method consistently outperforms other state-of-the-art methods at both reduced and full resolutions. The code is released in https://github.com/pengwei1996/Pansharpening .