HairLRM: Strand-based Hair Modeling via Large Reconstruction Models

How to read this

Category

research paper, strand level hair reconstruction from images using large reconstruction model priors

Contributions

• Argues that prior strand based hair reconstruction failures come not just from limited training data but from the fundamentally ill posed nature of regressing 3D strand orientation fields directly from 2D images without structural constraints
• Anchors the strand generation pipeline on an intermediate 3D surface mesh produced by a pretrained Large Reconstruction Model, using that mesh as a geometric scaffold rather than regressing strands from images alone
• Introduces a Dual Orientation AutoEncoder, conditioned on both a uniform point cloud and a salient point cloud sampling curls and flow boundaries, to resolve directional ambiguity in complex hair flows
• Adds a closed loop iterative refinement strategy that feeds extracted strands back into the encoder to progressively correct pathologies such as mesh interpenetration
• Supports single view and multi view input without pipeline changes, and reports outperforming prior single view methods on global silhouette accuracy and difficult topologies such as ponytails

Context

This paper bridges two lines cited directly in its related work, strand from image regression methods such as Im2Haircut and Gaussian Haircut, which the authors argue over smooth curls and lose global structure in occluded regions like ponytails, and large reconstruction model work such as Dora and FaceLift, which produces high fidelity 3D surface meshes from sparse images but is not directly compatible with strand level hair geometry. No builds_on ancestors are listed in the archive dataset. The core move is using the LRM mesh as a scaffold, then a dedicated dual orientation network to lift it into strands.

Correctness

The claims rest on comparison against prior single view and multi view strand reconstruction baselines across varied domains, cartoons and realistic portraits, and difficult styles, curly hair and ponytails, which the paper describes as extensive validation, though this guide reviewed only the introduction and method description rather than the full results tables. As with most strand reconstruction work, ground truth strand geometry for real photographs is inherently unverifiable, so results are judged largely by silhouette and perceptual plausibility rather than exact strand level accuracy.

Clarity

A clearly written full SIGGRAPH conference paper, accessible to anyone with background in hair capture literature and diffusion or VAE style latent architectures, though the Dual Orientation AutoEncoder formulation in Section 3.1 assumes familiarity with cross attention point cloud encoders such as 3DShape2VecSet.

How to read it

Five minute pass: read the abstract and Figure 1 to see the range of inputs, cartoon to realistic portraits, and outputs, ponytails and curls, the method handles. Second pass: read the introduction and Section 3's overview of the DOAE pipeline, sampling, feature encoding, geometry decoding, to see how the LRM mesh scaffold and the orientation field combine. A third pass is worthwhile for anyone building a hair capture or grooming pipeline, since the related work section is a genuinely useful map of the current strand reconstruction and LRM landscape, and the closed loop refinement strategy is likely reusable elsewhere.