A multi-path collaboration-based and spatial-spectral prior-based fusion method was proposed for hyperspectral and multispectral images, to address the challenges of insufficient global modeling and local detail capture in hyperspectral-multispectral image fusion, as well as the difficulty in exploring correlations between adjacent spectral bands. Firstly, the backbone network integrated a Local Bottleneck Control Unit and a Transformer in a parallel architecture. The Local Bottleneck Control Unit learned local structures while suppressing redundant features, whereas the Transformer handled long-range dependencies. A bidirectional interactive fusion mechanism was adopted to enhance the comprehension of both local details and global contexts. Secondly, the spatial-spectral joint prior module employed a dual-path pooling strategy for spatial attention and introduced an intra-spectral grouped attention mechanism to quantify inter-band correlations. Finally, the multi-path aggregation network consolidated features through residual blocks and a progressive fusion strategy. Experimental results demonstrated that the proposed method achieved average improvements of 4.5% in PSNR and 0.7% in SSIM compared to eight other methods on the CAVE dataset, exhibiting superior performance in capturing local-global features and integrating spatial-spectral prior information.