Wispy to Voluminous: Prior-free Multi-view Capture of Strand-level Facial Hair

Researchers have introduced a pipeline for reconstructing strand-level facial hair from multi-view images, targeting beard, mustache, eyebrow, and eyelash capture. The key shift is that it converts an unstructured 3D Gaussian representation into explicit curves, which is much more useful for downstream production than hair that is fused into a volumetric face mesh.

The paper describes a four-stage process: first, it optimizes 3D Gaussians constrained by tracked head geometry to reduce sub-surface noise and force early ray termination; then it traces continuous strands through crossings and sharp curvature; next it grounds strands to the skin and resolves root/tip ambiguity with a physically motivated prior; finally, it refines density using opacity-driven control under photometric optimization.

For game teams, the practical value is editability. The authors say the recovered strands preserve the sparse orientation patterns typical of facial hair and are suitable for facial animation, physical simulation, grooming transfer, appearance editing, and physics-based rendering. That makes this more relevant to character pipelines than a pure research demo, especially for realistic avatars where facial hair is a major fidelity bottleneck.

This is an arXiv cs.GR paper submitted on 6 Jun 2026 by Jaeseong Lee, Giljoo Nam, Adrian Jarabo, and Carlos Aliaga. It’s also notable because the authors claim it is the first method to reconstruct high-fidelity facial hair strands from a 3D Gaussian representation, which suggests a broader trend: using Gaussian splats as an intermediate...