Scene graph generation (SGG) constructs structured semantic representations by detecting entities and predicting their pairwise relations. However, prevailing datasets exhibit severely long-tailed relation distributions, biasing models toward head-class relations while neglecting informative tail-class relations. We propose a post-processing residual correction framework that adjusts pretrained relation logits without modifying the backbone. The framework combines nonlinear semantic-geometric interactions, three residual components with class-adaptive scaling, and a hybrid loss for long-tail optimization. Experiments on VG150 and GQA200 show consistent improvements in mR@K and F@K across the evaluated backbones, with corresponding reductions in R@K.
