CVPR 2026

CGHair: Compact Gaussian Hair Reconstruction with Card Clustering

High-fidelity hair reconstruction with 200x less memory through hierarchical card clustering and shared Gaussian textures.

Haimin Luo1,2 Srinjay Sarkar1 Albert Mosella-Montoro1 Francisco Vicente Carrasco1 Fernando De la Torre1
1Carnegie Mellon University    2ShanghaiTech University
Paper
TL;DR — We present a compact pipeline for hair reconstruction from multi-view images. By clustering hair strands into representative cards and sharing Gaussian texture codebooks, we achieve comparable rendering quality with over 200x lower memory and 4x faster strand reconstruction than prior methods.
CGHair method overview showing hierarchical clustering from strands to hair cards to shared Gaussian textures
200x
Memory
Reduction
4x
Faster Strand
Reconstruction
Abstract

Overview

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.

Video Presentation

Key Contributions

What We Contribute

1. Generative Prior Accelerated Strand Reconstruction

An efficient method for reconstructing strand-level hair geometry from multi-view images, leveraging the PERM parametric model for a 4x speedup over prior approaches.

2. Hair Card-Guided Hierarchical Clustering

A compact hair modeling pipeline that clusters strands into representative hair cards, significantly reducing redundancy at the strand level.

3. Shared Gaussian Appearance Codebook

A shared Gaussian texture codebook for hair cards, enabling scalable and consistent appearance modeling across structurally similar hair regions with 200x memory reduction.

Method

Pipeline Overview

Given monocular video frames, we reconstruct hair strands with our efficient strand generator, group them by hair cards, and further cluster into card groups with shared Gaussian textures for compact appearance modeling.

Full pipeline overview
Stage 1

Preprocessing

Estimate camera poses from multi-view images, compute hair segmentation masks and orientation maps using Gabor filters for directional cues.

Stage 2

Gaussian Fitting & Strand Generation

Reconstruct head and hair Gaussians separately, fit a FLAME head model, then use PERM's generative prior to efficiently decode strand geometry from latent UV textures.

Stage 3

Hair Card Clustering

Cluster strands into groups, construct representative hair cards for each cluster, and extract per-card geometry textures for compact representation.

Stage 4

Compact Gaussian Appearance

Further cluster hair cards by appearance into groups with shared Gaussian texture codebooks, then optimize end-to-end with a tailored 3DGS scheme.

Results

Reconstruction Results

Novel-view rendering results of our compact CGHair representation.

Comparison

Qualitative comparison of our method against prior approaches.

CGHair Animation Results

Strand-level animations rendered with our compact CGHair representation across diverse hairstyles.

Citation

BibTeX

@inproceedings{cghair2026, title = {CGHair: Compact Gaussian Hair Reconstruction with Card Clustering}, author = {Luo, Haimin and Sarkar, Srinjay and Mosella-Montoro, Albert and Vicente Carrasco, Francisco and de la Torre, Fernando}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, year = {2026} }
Paper
Download Open in Tab