FieldMorph3D for Training-Free Structure-Coherent 3D Morphing via Attention-Corrected Progress Fusion

3D morphing generates continuous 3D transformations from a source object to a target object. However, large category or structural differences often produce duplicated parts, orientation inconsistency, and unstable local structures in intermediate states. Existing methods mainly focus on the interpolation or fusion of source and target features. They pay limited attention to structural evidence reliability and structure evolution consistency in the Sparse Structure (SS) stage. The SS stage determines spatial occupancy and coarse geometry. Errors at this stage can directly propagate to subsequent 3D decoding. Based on this observation, we propose FieldMorph3D, a training-free framework for 3D morphing. The framework first improves the reliability of structural evidence before feature fusion. It then enforces global, local, and temporal consistency during sparse structure evolution. We introduce Cross-Attention Feature Correction Gating (CFCG) to suppress overloaded query-key responses and unreliable attention outputs. This operation reduces duplicated and low-confidence structural evidence before fusion. We further propose Latent Correction Fusion (LCF) to stabilize intermediate sparse structures. LCF first resolves orientation ambiguity and establishes a consistent spatial reference. It then uses neighborhood information to correct abnormal local deviations while preserving the principal source-to-target morphing direction. Finally, it adjusts token-wise morphing progress according to endpoint occupancy and historical structure states. This regulation improves the birth, death, and continuous evolution of local structures. FieldMorph3D requires neither additional training nor explicit topology correspondence. It produces more stable, coherent, and structurally plausible 3D morphing sequences.