A visual fidelity enhancement model supported by differentiable layout optimization is proposed to solve the problems of element occlusion, structural deviation, semantic break and visual fidelity degradation in multi-dimensional graphic element rearrangement. In the model, elements such as text, images, ICONS, buttons and charts are transformed into a unified parametric representation, and the spatial adjacency, alignment, hierarchy and semantic proximity relationships are depicted by relational graphs. On this basis, a joint loss function of structure preservation, occlusion penalty, boundary constraint, visual balance and semantic consistency is constructed to realize the continuous derivable update of layout parameters. Experimental results show that the structural similarity of the proposed model reaches 0.914, the occlusion rate is reduced to 2.3%, the semantic preservation rate is 94.2%, and the average optimization time is 0.42 s, which is better than the regular layout, force-directed layout, genetic algorithm and deep generative layout models. The results show that the model can improve the structural stability, visual clarity and semantic continuity of complex graphics after reordering, and provide technical support for graphics editing, interface reconstruction and automatic typesetting.
