High-Fidelity Face Normal Estimation with Flow Matching
Abstract
High-fidelity face normal estimation is a fundamental task for photorealistic 3D reconstruction, relighting, and avatar creation. Previous methods formulate this problem as a direct pixel-wise regression, which often leads to over-smoothed results, fails to capture fine-scale geometric details, and exhibits sensitivity to challenging in-the-wild lighting conditions. In this paper, we challenge this paradigm by reframing face normal estimation as a conditional generative modeling problem. We introduce FNFlow, a novel framework that generates precise normal maps by leveraging the power and efficiency of latent flow matching. This formulation departs from conventional regression paradigms and enables more expressive modeling of fine-scale geometric details. By learning the entire conditional distribution instead of a single point estimate, FNFlow implicitly learns a strong geometric prior of human faces, enabling it to robustly generate sharp, consistent, and detailed normal maps even from ambiguous inputs. Extensive quantitative and qualitative evaluations on benchmark datasets demonstrate that FNFlow achieves state-of-the-art performance. Our code is available at https://github.com/zhangquanbo/FNFlow.