Deep-learning-assisted adaptive optimization of spatial art layout parameters

Spatial art layout in two-dimensional visual design requires coordinated control of information communication, visual balance, element alignment, and geometric feasibility. Existing methods mainly generate compositions or assign global scores, leaving element-level repair underexplored. This study formulates spatial art layout as typed elements on a normalized canvas and proposes a deep-learning-assisted adaptive bounding-box optimization framework. A graph attention network (GAT) learns relation-aware representations, while a pooled random forest (RF) models global statistics; their equal-weight ensemble supplies the quality term for constrained local search. Rubric-assisted labels are constructed for 500 layouts derived from CGL-Dataset v2, and translation, scaling, and alignment perturbations define a reproducible recovery task. Across three random seeds, RF–GAT achieves an overall-score root mean square error of 0.4018 ± 0.0112 and a Spearman correlation of 0.7864 ± 0.0386, improving the RF values of 0.4271 ± 0.0026 and 0.7204 ± 0.0042. On 30 perturbed layouts, ensemble-guided optimization reaches a bounding-box recovery of 0.5841 ± 0.0368, exceeding RF-guided search by 0.0641 (paired Wilcoxon p = 0.0197). In a 75-layout RF-controlled ablation, adaptive and fixed weighting achieve comparable recovery ( p = 0.3157), while adaptive weighting redistributes repair pressure across geometric objectives. Budget, optimizer, label-source, and failure-mode analyses identify candidate allocation, constraint scheduling, and rubric construction as determinants of performance. The learned score represents composite-rubric conformity, with independent human-preference validation remaining outside the evaluation. Within this reference-anchored setting, the framework provides an interpretable means of coordinating relation-aware quality estimation with explicit geometric constraints for local design refinement.