HairLRM: Strand-based Hair Modeling via Large Reconstruction Models

HairLRM tackles a familiar pain point in hair reconstruction: strand-based methods often produce plausible-looking but wrong results when the input image doesn’t fully constrain the 3D shape. The paper argues that the core issue is ill-posed reconstruction, not just data scarcity, and that this shows up especially in global occlusion like ponytails and local directionality like curls.

The main change is to bring in a Large Reconstruction Model (LRM) mesh as a geometric scaffold before generating strands. On top of that, the authors use a Dual Orientation AutoEncoder to lift the coarse shape into high-fidelity hair strands, with latent-space optimization and surface-guided refinement to handle vector-field singularities. In plain terms: they’re trying to make the system respect the underlying head/hair structure instead of hallucinating strand flow from scratch.

For game developers, this matters most for character art pipelines, scan cleanup, and any workflow that needs believable hair from limited reference. If the method holds up outside the paper, it could reduce the amount of manual grooming needed after reconstruction and improve consistency on hard cases that usually break automated tools. That’s especially relevant for realistic characters, cinematics, and digital human pipelines.

The paper is submitted to arXiv on 13 Jun 2026 under cs.GR, with cross-listing in cs.CV. It claims a new benchmark for robustness and accuracy in hair reconstruction, but as with most reconstruction papers, the practical question is whether the method is fast, stable, and easy to...