Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CGHair: Compact Gaussian Hair Reconstruction with Card Clustering

Haimin Luo, Srinjay Sarkar, Albert Mosella-Montoro, Francisco Vicente Carrasco, Fernando De la Torre

Abstract

We present a compact pipeline for high-fidelity hair reconstruction from multi-view images. While recent 3D Gaussian Splatting (3DGS) methods achieve realistic results, they often require millions of primitives, leading to high storage and rendering costs. Observing that hair exhibits structural and visual similarities across a hairstyle, we cluster strands into representative hair cards and group these into shared texture codebooks. Our approach integrates this structure with 3DGS rendering, significantly reducing reconstruction time and storage while maintaining comparable visual quality. In addition, we propose a generative prior accelerated method to reconstruct the initial strand geometry from a set of images. Our experiments demonstrate a 4-fold reduction in strand reconstruction time and achieve comparable rendering performance with over 200x lower memory footprint.

CGHair: Compact Gaussian Hair Reconstruction with Card Clustering

Abstract

We present a compact pipeline for high-fidelity hair reconstruction from multi-view images. While recent 3D Gaussian Splatting (3DGS) methods achieve realistic results, they often require millions of primitives, leading to high storage and rendering costs. Observing that hair exhibits structural and visual similarities across a hairstyle, we cluster strands into representative hair cards and group these into shared texture codebooks. Our approach integrates this structure with 3DGS rendering, significantly reducing reconstruction time and storage while maintaining comparable visual quality. In addition, we propose a generative prior accelerated method to reconstruct the initial strand geometry from a set of images. Our experiments demonstrate a 4-fold reduction in strand reconstruction time and achieve comparable rendering performance with over 200x lower memory footprint.

Paper Structure

This paper contains 28 sections, 7 equations, 17 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Hair Strand Generation
  5. Hair Card Generation.
  6. Compact Gaussian Hair Representation Based on Card Clustering.
  7. Hair Appearance Modeling from Images.
  8. Experiments
  9. Implementation Details.
  10. Comparisons with State-of-the-art.
  11. Ablation Studies.
  12. Conclusion
  13. Additional Results
  14. Computational time analysis
  15. Strands Reconstruction
  16. ...and 13 more sections

Figures (17)

  • Figure 1: Full pipeline. Given monocular video frames, we first reconstruct hair strands with our efficient strand generator. The strands are grouped by hair cards, which are further clustered into card groups. Each card group is assigned to a shared Gaussian Texture for compact appearance modeling, a specialized global MLP is used to compute Spherical Harmonics (SH) from the Gaussian Textures.
  • Figure 2: Strand Generation Pipeline. We use the parametric hair geometric model (PERM) to synthesize a hair texture and decode it into strands. We then attach cylindrical Gaussian primitives to the strands and optimize them from images using the 3DGS.
  • Figure 3: (a) Hair card geometry construction; (b) Strand texture generation.
  • Figure 4: We model each hair card cluster with a shared Gaussian appearance texture and a lightweight globally-shared decoder, achieving over 270-fold compression.
  • Figure 5: Qualitative comparison of reconstructed hair strands with GaussianHaircut and GaussianHair. Our method achieves higher geometric fidelity, over 4× speedup (vs. GaussianHaircut zakharov2024human), and a fully automatic pipeline (unlike GaussianHair luo2024gaussianhair).
  • ...and 12 more figures