TopoRefine: Topology-Aware Correspondence and Residual Refinement for Training-Free Subject-Consistent Generation

Subject-consistent generation (SCG) aims to preserve subject identity across diverse text-driven contexts. Recent training-free methods move beyond global sharing via correspondence-aware identity transfer. However, semantic-driven correspondence may transfer identity features to appearance-similar but structurally incompatible regions. Meanwhile, extended attention, commonly used in existing methods, enhances identity consistency but may perturb the emerging target structure, from global pose and layout to local structural details. Together, these issues can cause reference-pose over-copying, local structural artifacts, and degraded structural fidelity. We view training-free SCG as controlled identity injection, where the pretrained model determines the target structure, while the reference image provides subject identity. We propose TopoRefine, a training-free framework with Topology-aware Subject Correspondence (TSC) and Residual Identity Refinement (RIR). TSC regularizes semantic matching with subject-internal geodesic relations for structurally compatible identity transfer, while RIR injects reference identity as a foreground-gated, magnitude-aligned residual to preserve the emerging target structure. Both qualitative and quantitative results show that TopoRefine improves subject consistency while better preserving target-driven structure.