Frontalization Helps Recognition: Diffusion-Conditioned Shared Latents for Pose-Robust Face Frontalization

Face frontalization, the synthesis of a frontal-view face from an arbitrary pose, is typically treated as an end in itself, evaluated on image quality rather than on whether it actually helps the recognition task it is meant to support. We revisit this assumption and ask whether frontalization can be made to directly improve pose-robust face recognition, rather than being trained and evaluated as a separate objective. We propose a diffusion-conditioned shared latent that unifies the two: a spatial representation built from a coarse 3D geometry and texture proxy and an identity embedding, occlusion-masked to handle hair, eyewear, and other self-occlusions that existing methods handle poorly. This latent feeds two heads trained end to end, a recognition head operating directly in feature space and a conditional diffusion decoder, distilled to a small number of sampling steps, that reconstructs the frontal image. Because both heads backpropagate through the same latent, the recognition objective is kept consistent with an explicit, reconstructable geometry, and the generative objective benefits from an identity-discriminative signal rather than pixel reconstruction alone. We evaluate on standard pose-robust recognition benchmarks and image-quality metrics against recent feature-space and image-level baselines, and test generalization across a broader pose distribution than commonly used small-scale benchmarks to probe an overfitting failure mode reported in prior work. We report inference-time comparisons against optimization-based frontalization methods, since removing per-image test-time optimization is a central goal of our design, and use ablations isolating the shared latent, occlusion masking, and diffusion decoder to attribute gains to specific architectural choices rather than to scale alone.