With the wide application of digital scene modelling in the fields of urban planning, architectural visualization and virtual simulation, the existing modelling systems put forward higher requirements on realism, detail reproduction and interaction efficiency. In order to improve the modelling quality, GAF-Net algorithm is constructed, which integrates multi-view segmentation, edge perception and lighting optimization. The results show that GAF-Net outperforms traditional algorithms in terms of edge clarity (0.872), occlusion filling rate (93.6%), and texture transition entropy (1.62). Meanwhile, the average frame processing time is 112.4 ms, and the GPU utilisation rate is only 68.3%, which provides efficient computing capability. In the error evaluation, the GAF-Net boundary offset is 1.84 pixels, and the detail loss is controlled at 6.3%. After the system is applied, the average rating of practitioners on visual realism and ease of operation reaches 4.65 and 4.45, respectively. The strategy can achieve multi-dimensional optimisation in accuracy, efficiency and visualisation performance, and can provide stable and efficient technical support for digital modelling.