CGHair: Compact Gaussian Hair Reconstruction with Card Clustering

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

Figure 2: 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 3: 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 4: (a) Hair card geometry construction; (b) Strand texture generation.

Figure 5: We model each hair card cluster with a shared Gaussian appearance texture and a lightweight globally-shared decoder, achieving over 270-fold compression.

Figure 6: Qualitative comparison of reconstructed hair strands with GaussianHaircut and GaussianHair. Our method achieves higher geometric fidelity, over 4× speedup (vs. GaussianHaircut [ 78 ] ), and a fully automatic pipeline (unlike GaussianHair [ 36 ] ).

Figure 7: Novel-view rendering comparisons on three challenging real-world hairstyles against GaussianHair [ 36 ] and GaussianHaircut [ 78 ] . CGHair achieves comparable quality with 200x fewer parameters. Refer to the supplementary material for quantitative results.

Figure 8: We integrate our CGHair module with GaussianHair [ 36 ] (Ours-GH), and achieve superior hair details.

Figure 9: Strand generator ablations, showing reconstructed hair geometries. Refer to the supplementary material for more results.

Metric	Ours	Only $\mathcal{D}_{PCA}$	Freeze $\alpha,\beta$	Only $\mathcal{D}_{style}$	Only $\mathcal{D}_{guide}$
Pos. Error $\downarrow$	0.148	0.162	0.159	0.182	0.170
Cur. Error $\downarrow$	6.73	9.12	6.98	7.48	7.19

Table 1: Quantitative ablation studies on strand generation.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size (MB) $\downarrow$	Ratio $\uparrow$
Unique	33.99	0.982	0.019	163.7	1.0
W/o card clustering	28.00	0.938	0.042	22.42	7.30
Single-SH	30.48	0.958	0.037	0.46	355.9
Multi-SH	33.50	0.978	0.021	11.95	13.70
Latents(full)	32.40	0.970	0.026	0.71	230.6

Table 2: Quantitative ablation on key modules of CGHair.

$K$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size (MB) $\downarrow$	Ratio $\uparrow$
10	32.40	0.970	0.026	0.71	230.6
30	32.40	0.970	0.026	1.72	95.17
60	32.36	0.969	0.026	3.22	50.84
90	32.10	0.968	0.026	4.72	34.68

Table 3: Codebook size

K

ablations.

$D$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size (MB) $\downarrow$	Ratio $\uparrow$
32	32.39	0.969	0.026	0.60	272.8
64	32.40	0.970	0.026	0.71	230.6
128	32.18	0.969	0.026	0.93	176.0
256	31.94	0.967	0.027	1.37	119.5

Table 4: Latent strand feature dimensionality

D

ablations.

$N_{T}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size (MB) $\downarrow$	Ratio $\uparrow$
16	31.49	0.964	0.030	0.60	272.8
32	31.84	0.967	0.027	0.64	255.8
64	32.40	0.970	0.026	0.71	230.6
96	32.57	0.971	0.025	0.79	207.2
400	33.04	0.975	0.022	1.53	108.0

Table 5: Quantitative ablations on texture clusters number

N_{T}

.

Figure 10: We show the ablations of our strand generator on several real captured hair styles.

Model	Recon. Time	Ratio
GaussianHairCut	11h	4.14
GaussianHair	8h	3.01
Ours	2h40m	1.00

Table 6: Comparison of the strand reconstruction time .

Model	Recon. Time	Ratio
GaussianHairCut	15h	3.00
GaussianHair	10.5h	2.10
Ours	5h	1.00

Table 7: Comparison on Time for End-to-End Reconstruction .

	Appearance Features		File size
Model	Size (Mb)	Ratio	Size (Mb)	Ratio
GaussianHairCut	544	464	93	77.51
GaussianHair	505	431	90	75
Ours	1.17	1.00	1.20	1.00

Table 8: Comparison of storage cost .

	Curly			Short			Long
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
GaussianHairCut	30.07	0.9225	0.0565	31.03	0.9041	0.0774	27.78	0.8719	0.1010
GaussianHair	31.56	0.9290	0.0466	32.07	0.9020	0.0597	29.12	0.8791	0.0749
Ours	30.34	0.9153	0.0528	31.15	0.8928	0.0696	28.60	0.8696	0.0815

Table 9: Quantitative comparison on appearance . Across three hair types (Curly, Short, and Long) from GaussianHair dataset.

Figure 11: Qualitative ablation results using a carefully designed synthetic hairstyle, illustrating the algorithm’s robustness in complex scenarios where hair exhibits multiple color tones.

Figure 12: More hairstyles from the MonoHair dataset. We compare our reconstructed strands against MonoHair and GaussianHaircut.

Figure 13: More Hairstyles appearance rendering comparisons. We compare our appearance rendering results against GaussianHaircut on 3 hairstyles from the MonoHair dataset and one from the H3DS dataset.

Figure 14: We also compare our reconstructed strands with GaussianHaircut on the H3DS hairstyle; note that GaussianHaircut failed.

Figure 15: Animated hair: On the left, original reconstruction, on the right, random timestamp frame from a strand-based animation.

Figure 16: Our explicit strand representation naturally enables strand-based dynamic simulation for a synthetic hairstyle (top), a real captured hairstyle from GaussianHair dataset (middle), and the other one from MonoHair Dataset (bottom). The simulated Gaussian strands are rendered using the standard Gaussian rendering pipeline. Please refer to the project webpage for a better visualization of the dynamic motions.

Figure 17: We show several example representative strand textures for grouped strands, which are represented by a hair card. The results are reconstructed from the hairstyle big wavy from the Monohair dataset.

Figure 18: Visualization of our additional free-view renderings of real-captured hairstyles.

CGHair: Compact Gaussian Hair Reconstruction with Card Clustering

Abstract

Figures & tables

3D Reconstruction

HairCS: Reconstructing Strand-Based Hair from Hair Cards

Wispy to Voluminous: Prior-free Multi-view Capture of Strand-level Facial Hair

Real-Time Neural Hair G-Buffer Anti-Aliasing

Abstract

3D Reconstruction

Explore similar work

HairCS: Reconstructing Strand-Based Hair from Hair Cards

Wispy to Voluminous: Prior-free Multi-view Capture of Strand-level Facial Hair

Real-Time Neural Hair G-Buffer Anti-Aliasing