Towards Unified 3D Hair Reconstruction from Single-View Portraits

TL;DR

This work tackles single-view 3D hair reconstruction across braided and un-braided hairstyles by proposing a unified pipeline based on a flexible 3D Gaussian hair representation. The authors learn two hair-specific diffusion priors, HairSynthesizer and HairEnhancer, from a new large-scale synthetic dataset SynMvHair to perform coarse-to-fine optimization that yields a consistent, texture-rich 3D hair model. A two-stage refinement—view-wise to improve multi-view texture consistency and pixel-wise to sharpen strand-level details—produces Theta^2, a Gaussian-based hair representation suitable for fast multi-view rendering. Across synthetic and real portraits, the approach achieves state-of-the-art results and demonstrates strong generalization, enabling new capabilities such as multi-view hair rendering from a single image and hairstyle customization for 3D avatars. The work significantly advances unified, texture-aware 3D hair reconstruction and provides a scalable dataset and priors to support future research in hair modeling and digital humans.

Abstract

Single-view 3D hair reconstruction is challenging, due to the wide range of shape variations among diverse hairstyles. Current state-of-the-art methods are specialized in recovering un-braided 3D hairs and often take braided styles as their failure cases, because of the inherent difficulty to define priors for complex hairstyles, whether rule-based or data-based. We propose a novel strategy to enable single-view 3D reconstruction for a variety of hair types via a unified pipeline. To achieve this, we first collect a large-scale synthetic multi-view hair dataset SynMvHair with diverse 3D hair in both braided and un-braided styles, and learn two diffusion priors specialized on hair. Then we optimize 3D Gaussian-based hair from the priors with two specially designed modules, i.e. view-wise and pixel-wise Gaussian refinement. Our experiments demonstrate that reconstructing braided and un-braided 3D hair from single-view images via a unified approach is possible and our method achieves the state-of-the-art performance in recovering complex hairstyles. It is worth to mention that our method shows good generalization ability to real images, although it learns hair priors from synthetic data.

Figures (20)

• Figure 1: Overview of our method. Given only a single-view portrait, our method can reconstruct its corresponding 3D hair in the representation of 3D Gaussian and render high-quality images from arbitrary views.
• Figure 2: Illustration of a sample in our SynMvHair dataset.
• Figure 3: With the aid of HairSynthesizer, we perform image-wise Gaussian refinement on optimized coarse 3D Gaussian. The refined 3D Gaussian will be obtained with better texture.
• Figure 4: With the aid of HairEnhancer, we perform pixel-wise Gaussian refinement on optimized coarse 3D Gaussian. The enhanced 3D Gaussian will be obtained with better texture.
• Figure 5: Comparison with (a) zheng2023hairstep, (b) hu2017avatar, and (c) sun2021single on braided hairstyles. From left to right: input images, results of previous methods and ours.