TriAnchor-ID: Multi-View Semantic Identity Anchoring for Pose-Aware Personalized Diffusion

This paper presents TriAnchor-ID, a prototype framework for identity-conditioned diffusion portrait generation. The work focuses on multi-view semantic identity anchoring using ArcFace/InsightFace embeddings, quality-weighted identity aggregation, landmark-derived geometric observations, and post-generation identity consistency evaluation.

The paper corrects an important distinction between semantic face identity embeddings and physical 3D morphable model parameters: the implemented 512-dimensional ArcFace vector is treated as a semantic identity anchor, not as a FLAME or 3DMM shape vector. The proposed Identity Capsule design separates implemented components from future extensions such as FLAME/DECA shape fitting, UV texture modeling, spatial ControlNet conditioning, and inference-time identity consistency guidance.

This version is released as a preprint/prototype research report. The qualitative examples are illustrative, and the paper outlines a controlled evaluation protocol using ArcFace similarity, identity drift score, landmark error, face detection failure rate, prompt alignment, repeated seeds, and confidence intervals.