Complexity-guided Regularization for Generalizable Human Gaussian Splatting

Gun Ryu; Gi-Mun Um; Won-Sik Cheong; Wonjun Kim

cs.CVOct 26, 2026

Complexity-guided Regularization for Generalizable Human Gaussian Splatting

Authors: Gun Ryu, Gi-Mun Um, Won-Sik Cheong, Wonjun Kim

Organizations: Konkuk University · Electronics and Telecommunications Research Institute

Abstract

Recent studies on generalizable human Gaussian splatting have achieved the notable progress in synthesizing novel views of unseen human subjects from multi-view images. Despite this progress, existing methods still struggle to represent fine-grained details of human subjects, which inherently involve high geometric and textural complexity. This is because previous methods do not impose a constraint on how densely Gaussians are placed, which often leads to blurry rendering results in complex regions. To address this problem, we propose a complexity-guided regularization scheme. The key idea is to leverage a complexity prior computed via the local entropy in geometric and textural characteristics of a clothed human. This effectively prevents Gaussians from being coarsely placed in complex regions, which facilitates the faithful reconstruction of intricate structures. Furthermore, we sample local patches predominantly from complex regions according to this complexity prior, which allows the model to be sufficiently supervised by local details. Experimental results on benchmark datasets demonstrate that our method delivers state-of-the-art performance while maintaining the real-time inference speed in generalizable human Gaussian splatting. The code and model are publicly available at https://github.com/DCVL-3D/CGR-GS_release.

Figures & tables

Figure 1: Distributions of Gaussian centers and the corresponding rendering results. (a) Novel view image. (b) GPS-Gaussian [ 10 ] . (c) RoGSplat [ 15 ] . (d) Proposed method.

Figure 2: Overall architecture of the proposed method. The complexity prior of the clothed human guides the model to allocate Gaussians densely in complex regions. Note that Gaussians are randomly supervised by either the entire image or complexity-aware local patches with the equal probability at each iteration.

Figure 3: Complexity prior and complexity-guided regularization scheme. (a) RGB image. (b) Per-vertex complexity prior $c$ . (c) Geometric complexity $\bar{H}^{\text{geo}}$ . (d) Textural complexity $\bar{H}^{\text{tex}}$ . (e) Overview of the complexity-guided regularization process, where the local Gaussian density is computed from the average distance of $k$ -nearest neighbor Gaussians and regularized to be positively correlated with the complexity prior via a pairwise ranking regularization scheme.

Figure 4: Qualitative comparison of in-domain novel view synthesis on THuman2.0 [ 24 ] , 2K2K [ 25 ] , and X-Humans [ 26 ] datasets (from top to bottom). (a) Input images. (b) Results by GPS-Gaussian [ 10 ] . (c) Results by RoGSplat [ 15 ] . (d) Results by GHGS-MVSC [ 12 ] . (e) Results by the proposed method. (f) Ground truth.

Methods	THuman2.0			2K2K			X-Humans			FPS $\uparrow$
Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FPS $\uparrow$
GPS-Gaussian [ 10 ]	26.54	0.9539	0.0610	26.46	0.9483	0.0611	26.29	0.9458	0.0687	56.31
RoGSplat [ 15 ]	28.83	0.9615	0.0429	30.67	0.9697	0.0310	30.31	0.9656	0.0371	5.42
GHGS-MVSC [ 12 ]	30.93	0.9710	0.0334	30.45	0.9678	0.0326	30.60	0.9668	0.0391	2.79
Ours	32.42	0.9766	0.0201	31.68	0.9717	0.0229	31.96	0.9703	0.0230	33.96

Table 1: Quantitative comparison of in-domain novel view synthesis on THuman2.0 [ 24 ] , 2K2K [ 25 ] , and X-Humans [ 26 ] datasets. The best result is shown in bold.

Figure 5: Qualitative comparison of cross-domain novel view synthesis on 2K2K [ 25 ] and X-Humans [ 26 ] datasets (from top to bottom). (a) Input images. (b) Results by GPS-Gaussian [ 10 ] . (c) Results by GHG [ 13 ] . (d) Results by RoGSplat [ 15 ] . (e) Results by the proposed method. (f) Ground truth.

Methods	THuman2.0 $\to$ 2K2K			THuman2.0 $\to$ X-Humans
Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
GPS-Gaussian [ 10 ]	24.47	0.9361	0.0721	24.34	0.9330	0.0795
GHG [ 13 ]	26.74	0.9548	0.0518	26.06	0.9494	0.0580
RoGSplat [ 15 ]	26.96	0.9490	0.0489	26.57	0.9496	0.0507
GHGS-MVSC [ 12 ]	28.59	0.9527	0.0480	25.17	0.9375	0.0674
Ours	29.52	0.9580	0.0412	27.53	0.9499	0.0474

Table 2: Quantitative comparison of cross-domain novel view synthesis on 2K2K [ 25 ] and X-Humans [ 26 ] datasets. The best result is shown in bold.

Figure 6: Ablation studies on the THuman2.0 dataset. We visualize the rendering result and the local Gaussian density. (a) Results by the baseline. (b) Results without patch supervision. (c) Results without $\mathcal{L}_{\text{comp}}$ . (d) Results by the proposed method. (e) Ground truth.

$\mathcal{L}_{\text{comp}}$	Patch sup.	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
✗	✗	30.32	0.9664	0.0316
✓	✗	30.79	0.9695	0.0269
✗	✓	32.06	0.9752	0.0209
✓	✓	32.42	0.9766	0.0201

Table 3: Ablation studies on the THuman2.0 dataset.

# Neighbors ( $k$ )	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
$k=3$	30.55	0.9684	0.0299
$k=5$	30.68	0.9690	0.0275
$k=7$	30.79	0.9695	0.0269
$k=9$	30.75	0.9696	0.0271

Table 4: Performance analysis according to different

k

values on the THuman2.0 dataset.

Strategy	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Random sampling	31.63	0.9731	0.0227
Gradient-based sampling	31.81	0.9744	0.0216
Complexity-aware sampling	32.06	0.9752	0.0209

Table 5: Comparison of patch sampling strategies on the THuman2.0 dataset.

Complexity component	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Geometric complexity ( $\bar{H}^{\text{geo}}$ )	31.65	0.9727	0.0232
Textural complexity ( $\bar{H}^{\text{tex}}$ )	32.24	0.9752	0.0214
Average operation ( $\mathrm{mean}(\bar{H}^{\text{geo}},\bar{H}^{\text{tex}})$ )	32.31	0.9760	0.0205
Maximum operation ( $\max(\bar{H}^{\text{geo}},\bar{H}^{\text{tex}})$ )	32.42	0.9766	0.0201

Table 6: Ablation study of complexity components on the THuman2.0 dataset.

Complexity-guided reg.	Correlation coefficient $\uparrow$
✗	0.1086
✓	0.4262

Table 7: Spearman rank correlation between local complexity and density of Gaussians on the THuman2.0 dataset.

Methods	Depth reg.	Textured mesh	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
GPS-Gaussian [ 10 ]	✓	✗	26.54	0.9539	0.0610
RoGSplat [ 15 ]	✓	✗	28.83	0.9615	0.0429
GHGS-MVSC [ 12 ]	✗	✗	30.93	0.9710	0.0334
Ours	✗	✗	31.83	0.9737	0.0231
	✓	✗	32.04	0.9744	0.0224
	✓	✓	32.42	0.9766	0.0201

Table 8: Quantitative comparison under different regularization settings on the THuman2.0 dataset.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Attribute	Per-token Dim.	Per-Gaussian Dim.	Activation
Depth	64	1	Sigmoid
3D spatial offset	192	3	-
Scale	192	3	Exp
Rotation	256	4	Norm.
Opacity	64	1	Sigmoid
Color	192	3	-

Appendix

Table 9: The output dimension and the activation function of each Gaussian attribute.

$\mathcal{L}_{\text{comp}}$	Patch sup.	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
✗	✗	29.80	0.9632	0.0357
✓	✗	30.41	0.9665	0.0327
✗	✓	30.96	0.9702	0.0279
✓	✓	32.01	0.9750	0.0212

Appendix

Table 10: Ablation studies on the THuman2.0 dataset without the depth regularization scheme

\mathcal{L}_{\text{depth}}

Loss formulation	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Baseline (w/o complexity-guided regularization scheme)	30.32	0.9664	0.0316
Pearson correlation formulation $\mathcal{L}_{\text{pearson}}$	30.31	0.9649	0.0325
Pairwise ranking formulation Eq. ( 7 )	30.79	0.9695	0.0269

Appendix

Table 11: Comparison of loss formulations for the complexity-guided regularization scheme on the THuman2.0 dataset.

Perturbed regularization	Input masks	PSNR $\uparrow$	$\Delta$ PSNR
✗	Clean	31.68	–
✗	Jitter	31.41	$-0.27$
✓	Clean	31.46	$-0.22$
✓	Jitter	31.28	$-0.40$

Appendix

Table 12: Robustness to degraded training-time regularization and inference-time mask perturbation on the 2K2K dataset.

Table 19

Figure 7: Qualitative comparison of in-domain novel view synthesis on THuman2.0 [ 24 ] with different numbers of input views (3-view, 4-view, and 5-view settings from top to bottom). (a, g) Input images. (b, h) Results by GPS-Gaussian [ 10 ] . (c, i) Results by RoGSplat [ 15 ] . (d, j) Results by GHGS-MVSC [ 12 ] . (e, k) Results by the proposed method. (f, l) Ground truth.

Methods	3-view			4-view			5-view
Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
GPS-Gaussian [ 10 ]	-	-	-	-	-	-	26.54	0.9539	0.0610
EVA-Gaussian [ 11 ]	-	-	-	26.31	0.9555	0.0391	27.54	0.9614	0.0297
GHG [ 13 ]	24.84	0.9232	0.0719	25.84	0.9462	0.0633	26.53	0.9571	0.0398
RoGSplat [ 15 ]	26.32	0.9478	0.0530	28.83	0.9615	0.0429	30.98	0.9711	0.0341
GHGS-MVSC [ 12 ]	27.81	0.9521	0.0517	30.93	0.9710	0.0334	31.54	0.9729	0.0269

Appendix

Table 15: Quantitative comparison of in-domain novel view synthesis on THuman2.0 [ 24 ] with different numbers of input views. The best result is shown in bold.

Methods	In-domain			THuman2.0 $\to$ THuman2.1
Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
GPS-Gaussian [ 10 ]	25.42	0.9430	0.0705	23.44	0.9315	0.0793
GHG [ 13 ]	27.83	0.9628	0.0498	25.07	0.9418	0.0651
RoGSplat [ 15 ]	30.57	0.9717	0.0300	26.64	0.9524	0.0523
GHGS-MVSC [ 12 ]	31.11	0.9735	0.0297	28.39	0.9621	0.0384
Ours	35.49	0.9851	0.0132	29.65	0.9623	0.0267

Appendix

Table 16: Quantitative comparison of novel view synthesis on the THuman2.1 [ 24 ] dataset under in-domain and cross-domain settings. The best result is shown in bold.

Figure 8: Qualitative comparison of in-domain novel view synthesis on the THuman2.1 [ 24 ] dataset. (a) Input images. (b) Results by GPS-Gaussian [ 10 ] . (c) Results by RoGSplat [ 15 ] . (d) Results by GHGS-MVSC [ 12 ] . (e) Results by the proposed method. (f) Ground truth.

Figure 9: Qualitative comparison of cross-domain novel view synthesis on the THuman2.1 [ 24 ] dataset. (a) Input images. (b) Results by GPS-Gaussian [ 10 ] . (c) Results by GHG [ 13 ] . (d) Results by RoGSplat [ 15 ] . (e) Results by the proposed method. (f) Ground truth.

Figure 10: Qualitative comparison of cross-domain novel view synthesis on DNA-Rendering [ 29 ] and ZJU-MoCap [ 30 ] datasets (from top to bottom). (a) Input images. (b) Results by GPS-Gaussian [ 10 ] . (c) Results by GHG [ 13 ] . (d) Results by RoGSplat [ 15 ] . (e) Results by the proposed method. (f) Ground truth.

Methods	THuman2.0 $\to$ DNA-Rendering			THuman2.0 $\to$ ZJU-MoCap
Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
GPS-Gaussian [ 10 ]	23.22	0.8821	0.1296	26.26	0.9296	0.0829
GHG [ 13 ]	23.31	0.8852	0.0937	26.72	0.9357	0.0712
RoGSplat [ 15 ]	25.68	0.9180	0.0803	27.66	0.9508	0.0547
GHGS-MVSC [ 12 ]	17.76	0.8274	0.1397	23.10	0.8924	0.1056
Ours	26.47	0.9186	0.0788	28.99	0.9525	0.0536

Appendix

Table 17: Quantitative comparison of cross-domain novel view synthesis on real-world datasets, i.e., DNA-Rendering [ 29 ] and ZJU-MoCap [ 30 ] . The best result is shown in bold.

Methods	THuman2.0					X-Humans
Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CSIM $\uparrow$	AKD $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CSIM $\uparrow$	AKD $\downarrow$
GPS-Gaussian [ 10 ]	20.20	0.798	0.240	0.426	3.19	19.52	0.712	0.327	0.388	3.69
RoGSplat [ 15 ]	18.18	0.748	0.259	0.473	2.76	19.84	0.703	0.228	0.484	2.99
GHGS-MVSC [ 12 ]	23.23	0.856	0.184	0.549	2.25	23.18	0.785	0.219	0.513	2.72
Ours	24.73	0.888	0.111	0.699	1.26	25.48	0.831	0.137	0.620	2.04

Appendix

Table 18: Quantitative comparison on the face region of THuman2.0 [ 24 ] and X-Humans [ 26 ] datasets under in-domain settings. The best result is shown in bold.

Figure 11: Qualitative comparison on the face region of THuman2.0 [ 24 ] and X-Humans [ 26 ] datasets (from top to bottom). (a) Results by GPS-Gaussian [ 10 ] . (b) Results by RoGSplat [ 15 ] . (c) Results by GHGS-MVSC [ 12 ] . (d) Results by the proposed method. (e) Ground truth.

Methods	PSNR $\uparrow$	LPIPS ${}^{*}\downarrow$	FID $\downarrow$	FPS $\uparrow$
LIFe-GoM [ 14 ]	25.10	108.75	45.17	1.0
Ours	29.22	49.44	7.91	7.7

Appendix

Table 19: Quantitative comparison with LIFe-GoM [ 14 ] on the THuman2.0 dataset. The best result is shown in bold.

Datasets	Methods	# of input views	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
THuman2.1	IDOL [ 40 ]	1	22.12	0.9257	0.112
THuman2.1	Ours	4	33.51	0.9802	0.017
2K2K	LHM [ 41 ]	1	20.73	0.9084	0.122
2K2K	Ours	4	31.72	0.9722	0.022

Appendix

Table 20: Quantitative comparison with single-image human reconstruction methods. The best result is shown in bold.

3D Gaussian Splatting

3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.

See all 1,013 papers

Generalizable Human Gaussian Splatting via Multi-view Semantic Consistency
Apr 28, 2026Jingi Kim, Wonjun KimMulti-View Consistency 3D Gaussian Splatting
Konkuk University
CLOTH-HUGS: Cloth Aware Human Gaussian Splatting
Apr 17, 2026Sadia Mubashshira, Nazanin Amini, Kevin Desai3D Gaussian Splatting 3D Reconstruction
HeroGS: Hierarchical Guidance for Robust 3D Gaussian Splatting under Sparse Views
Mar 1, 2026Jiashu Li, Xumeng Han, Zhaoyang Wei +5View-Dependent 3DGS 3D Gaussian Splatting
University of the Chinese Academy of Sciences

Apr 28, 2026J/K move · Enter open · S save

Generalizable Human Gaussian Splatting via Multi-view Semantic Consistency

Jingi Kim, Wonjun Kim

Konkuk University

Multi-View Consistency 3D Gaussian Splatting Cross-View Attention Sparse-View 3D Reconstruction Gaussian Splatting

Open paper AskPDF arXiv

Recently, generalizable human Gaussian splatting from sparse-view inputs has been actively studied for the photorealistic human rendering. Most existing methods rely on explicit geometric constraints or predefined structural representations to accurately position 3D Gaussians. Although these approaches have shown the remarkable progress in this field, they still suffer from inconsistent feature representations across multi-view inputs due to complex articulations of the human body and limited overlaps between different views. To address this problem, we propose a novel method to accurately localize 3D Gaussians and ultimately improve the quality of human rendering. The key idea is to unproject latent embeddings encoded from each viewpoint into a shared 3D space through predicted depth maps and recalibrate them belonging to the same body part based on cross-view attention. This helps the model resolve the spatial ambiguity occurring in highly textured regions as well as occluded body parts, thus leading to the accurate localization of 3D Gaussians. Experimental results on benchmark datasets show that the proposed method efficiently improves the performance of generalizable human Gaussian splatting from sparse-view inputs.

Abstract

3D Gaussian Splatting

Explore similar work

Generalizable Human Gaussian Splatting via Multi-view Semantic Consistency

CLOTH-HUGS: Cloth Aware Human Gaussian Splatting

HeroGS: Hierarchical Guidance for Robust 3D Gaussian Splatting under Sparse Views