GeomHair: Reconstruction of Hair Strands from Colorless 3D Scans

TL;DR

This paper tackles reconstructing complete hair strand geometry from colorless 3D head scans by combining geometry-derived orientation cues with a text-conditioned diffusion prior. GeomHair fuses 3D crest-line orientations and 2D TEED-based orientations from scan renderings to guide a strand-based optimization that uses an unsigned distance function and a diffusion prior conditioned on a scan-specific description, yielding realistic hair for challenging hairstyles without surface color. The approach introduces a refined diffusion prior (HAAR) with a denoising schedule and text conditioning, and demonstrates scalability by producing Strands400, a large public dataset of real-world hair strands aligned to surface geometry. Empirically, GeomHair achieves competitive results with RGB-based methods on straight and curly hairstyles, while enabling dataset generation from mesh data and offering improved robustness to lighting and occlusions inherent to colorless scans.

Abstract

We propose a novel method that reconstructs hair strands directly from colorless 3D scans by leveraging multi-modal hair orientation extraction. Hair strand reconstruction is a fundamental problem in computer vision and graphics that can be used for high-fidelity digital avatar synthesis, animation, and AR/VR applications. However, accurately recovering hair strands from raw scan data remains challenging due to human hair's complex and fine-grained structure. Existing methods typically rely on RGB captures, which can be sensitive to the environment and can be a challenging domain for extracting the orientation of guiding strands, especially in the case of challenging hairstyles. To reconstruct the hair purely from the observed geometry, our method finds sharp surface features directly on the scan and estimates strand orientation through a neural 2D line detector applied to the renderings of scan shading. Additionally, we incorporate a diffusion prior trained on a diverse set of synthetic hair scans, refined with an improved noise schedule, and adapted to the reconstructed contents via a scan-specific text prompt. We demonstrate that this combination of supervision signals enables accurate reconstruction of both simple and intricate hairstyles without relying on color information. To facilitate further research, we introduce Strands400, the largest publicly available dataset of hair strands with detailed surface geometry extracted from real-world data, which contains reconstructed hair strands from the scans of 400 subjects.

Figures (10)

• Figure 1: Quantitative evaluation of different methods and ablations. For straight hair, our method outperforms both Neural Haircut neuralhaircut and Gaussian Haircut GH in both precision and F-score metrics. For curly hair, our method achieves superior recall and F-score compared to both approaches, demonstrating better overall hair strand recovery.
• Figure 2: Overview of our GeomHair framework. Our method consists of two main stages: orientations extraction (left) and strands reconstruction (right). In the orientation extraction stage, we extract complementary orientation signals by combining 3D orientations from crest lines with 2D orientations obtained from TEED features applied to rendered shading of the scans. During reconstruction, we optimize a geometry texture to generate hair strands while enforcing multiple constraints: the orientation loss ($\mathcal{L}_\text{orient}$) ensures strand growth aligns with our extracted 3D+2D orientation field, volume loss ($\mathcal{L}_\text{volume}$) keeps strands near the surface, and Chamfer distance ($\mathcal{L}_\text{chamfer}$) promotes uniform coverage of the hair volume. To enhance realism, we additionally incorporate a diffusion prior ($\mathcal{L}_\text{prior}$) conditioned on hairstyle descriptions generated by a VQA model analyzing the input scan.
• Figure 3: Comparison of our method with state-of-the-art hair reconstruction methods across five different scenes. For Neural Haircut neuralhaircut, Gaussian Haircut GH, and MonoHair MH, the input is a $360^\circ$ RGB video, and for our method, the input is a corresponding 3D scan without color of the same person, closely following NPHM dataset giebenhain2023learning setup. Faces blurred for anonymity purposes.
• Figure 4: Results on 3D-designed assets. Here, we demonstrate the results of fitting to a synthetic, hand-carved 3D mesh from a 3D stock (CGTrader, author: ZStuff). Our method is capable of reconstructing a plausible collection of strands, even when not so many guiding strands can be observed.
• Figure 5: Ablation over the various components of the GeomHair pipeline. Here, we demonstrate the most important components of the pipeline that provide the largest improvement. The comparison is done over the scans by https://www.cemyuksel.com/research/hairmodels/ made from hand-drawn hair strands. 3D Gabor corresponds to replacing Crest Lines algorithm with more simple Gabor filtering (corresponds to the w/ 3DO Gabor ablation in Table \ref{['tab:comparison_suppmat']}).