Carbon fiber reinforced polymer (CFRP) bolted joint structures, with their excellent detachability and lightweight properties, have exhibited great application potential in the aerospace field. However, the deformation analysis of such structures under the effect of multi-source assembly variables has been a technical bottleneck restricting its wide application. In view of this, this study proposes an analytical framework named “CFRP–bolted joint–generative adversarial network” (CFRP–BJ–GAN) for the deformation analysis of CFRP bolted structures. The framework first introduces a multi-scale geometric deviation modeling method based on statistical parameters of key features, which enables accurate capture of the structural deformation characteristics at different scales. Subsequently, by introducing an advanced ViT encoder architecture, it realizes the deep integration and efficient processing of multiple types of heterogeneous data, thus enhancing the accuracy and efficiency of deformation prediction. The experimental validation results show that the CFRP–BJ–GAN framework outperforms traditional numerical simulation methods in the calculation of all the proposed evaluation metrics, while a single prediction takes only 8 s, which significantly improves the analysis speed. Therefore, the CFRP–BJ–GAN framework proposed in this study provides an efficient, accurate and practical solution for the deformation analysis of CFRP bolted joint structures.