KoalaGS: Complexity-Aware Continuous Level of Detail for Gaussian Splatting

Radu Beche; Raoul de Charette

cs.CVOct 20, 2026

KoalaGS: Complexity-Aware Continuous Level of Detail for Gaussian Splatting

Authors: Radu Beche, Raoul de Charette

Organizations: Work done while visiting the Astra-Vision group at Inria, Paris. · TUCN · Inria

Abstract

Deploying a trained 3D Gaussian Splatting scene under shrinking compute budgets requires one primitive set to remain effective when rendered at smaller prefixes. In this continuous level of detail (LoD) setting, the challenge is to decide which Gaussians should retain capacity as the budget decreases. A pilot study on 3DGS-MCMC reconstructions shows that image complexity, computed from the input views before optimization, predicts where reconstruction error concentrates in image and 3D space. It also reveals that simple regions contain too many Gaussians and degrade as much as complex ones at lower budgets, despite being inherently cheaper to represent. Subsequently, in KoalaGS, during training we use this complexity prior to bias relocation toward poorly reconstructed areas while ensuring smooth degradation across LoDs. A primitive importance score then enables inference at low-capacity, while preserving the scene structure. Tested across three datasets, KoalaGS maintains competitive full-capacity quality while significantly improving quality across the budget spectrum, yielding smoother transitions across LoDs, and enabling faster progressive streaming.

Figures & tables

Panel of Figure 1 — Figure 1: Analysis methodology. ( 1(a) ) We measure the spatial correlation between the complexity map of the input image (left), and the quality map (center), obtained after 3DGS-MCMC optimization. We then project and aggregate these 2D fields onto the Gaussian particles to obtain complexity/quality fields in 3D (right). ( 1(b) ) Complexity and quality inversely correlate both in the 2D image plane and in the 3D object space. Note that for easier visualization, we inverted colors of all complexity maps .

Domain	Complexity	PSNR	SSIM	PSNR	SSIM	Avg. Rank
		Pearson $\rho$		Spearman $r_{s}$
2D	Shannon	$-0.605$	$-0.259$	$-0.590$	$-0.346$	$3$
	Laplacian	$-0.546$	$-0.207$	$\mathbf{-0.649}$	$-0.365$	$2$
	Delentropy	$\mathbf{-0.647}$	$\mathbf{-0.379}$	$-0.638$	$\mathbf{-0.498}$	$\mathbf{1}$
	Average	$-0.599$	$-0.282$	$-0.626$	$-0.403$	–
3D	Shannon	$-0.588$	$-0.557$	$-0.170$	$-0.165$	$3$

Table 1: Correlation between local complexity and reconstruction quality. On DL3DV, across 140 scenes, complexity is negatively correlated with PSNR and SSIM in both the image plane and object space, indicating that more complex regions tend to reconstruct worse. Delentropy gives the strongest average rank. Best correlations are shown in bold.

Panel of Figure 2 — Figure 2: Relation between complexity and Gaussian density. ( 2(a) ) Each bin counts the Gaussians inside a voxel of that complexity aggregated across DL3DV scenes. On the y-axis, the straight rise of the percentiles indicates an exponential growth (note y-axis is log-scaled). ( 2(b) ) Considering an indoor scene reconstructed with 3DGS-MCMC, we lower capacity by dropping primitives in order of opacity, plotting the average PSNR reached at each complexity for every budget. We observe that the simplest regions lose almost as much quality as the most complex ones, so even flat surfaces are broken into multiple low opacity Gaussians.

			100% Capacity			10% Capacity			1% Capacity			0.1% Capacity
	Method	#GS [k]	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
DB	OctreeGS [ 27 ]	1,457	30.23	0.908	0.254	-	-	-	-	-	-	-	-	-
	FlexGS [ 5 ]	2,547	29.29	0.898	0.320	29.27	0.897	0.349	24.68	0.824	0.472	-	-	-
	3DGS MCMC [ 8 ]	2,950	29.70	0.907	0.236	24.36	0.792	0.362	16.81	0.656	0.520	12.67	0.601	0.599
	LightGaussian [ 15 ]	2,950	29.73	0.907	0.240	26.21	0.849	0.333	19.59	0.742	0.482	15.26	0.675	0.546
	MGS [ 2 ]	2,950	29.29	0.905	0.247	28.61	0.888	0.282	23.28	0.800	0.416	16.58	0.692	0.531

Table 2: Performance at log-spaced LoD targets. Metrics at 100%, 10%, 1%, and 0.1% budgets across three datasets. Among full-LoD methods, ours performs on par at full capacity and generally better at lower budgets. For partial-LoD methods, dashes (-) denote unreachable operating points.

Figure 3: Qualitative comparison across the LoD spectrum. Methods perform similarly at full capacity but diverge sharply at lower budgets. Ours better preserves scene coverage at low budgets and remains competitive with FlexGS while enabling prefix rendering.

Figure 4: Quality metrics across the LoD spectrum. Our method matches full-capacity baselines and degrades more gracefully as budgets shrink. Below 10%, 3DGS MCMC collapses and LightGaussian improves little. OctreeGS has limited discrete LoDs, FlexGS fluctuates between trained budgets, and MGS follows ours but with lower quality throughout. Results are on MipNerf.

		Total [0.1%, 100%]			[0.1%, 1%] ( $\times 100$ )			[1%, 10%] ( $\times 10$ )			[10%, 100%]
	Method	AUC PSNR $\uparrow$	AUC SSIM $\uparrow$	AUC LPIPS $\downarrow$	AUC PSNR $\uparrow$	AUC SSIM $\uparrow$	AUC LPIPS $\downarrow$	AUC PSNR $\uparrow$	AUC SSIM $\uparrow$	AUC LPIPS $\downarrow$	AUC PSNR $\uparrow$	AUC SSIM $\uparrow$	AUC LPIPS $\downarrow$
DB	FlexGS [ 5 ]	27.47	0.867	0.363	-	-	-	25.28	0.788	0.350	24.94	0.788	0.328
	3DGS MCMC [ 8 ]	27.96	0.870	0.274	13.69	0.573	0.497	19.49	0.664	0.378	25.87	0.798	0.231
	LightGaussian [ 15 ]	28.61	0.888	0.268	16.29	0.650	0.457	21.44	0.730	0.347	26.31	0.809	0.229
	MGS [ 2 ]	28.95	0.897	0.259	18.81	0.685	0.414	24.62	0.778	0.289	26.30	0.813	0.226
	Ours	29.51	0.902	0.254	19.19	0.690	0.413	24.86	0.778	0.288	26.83	0.817	0.221

Table 3: Area Under the Curve (AUC) across Capacity Intervals. For all intervals, our method outperforms all other prefix-renderable continuous LoD methods, exhibiting large gaps in smaller intervals

[0.1\%,1\%]

. Intervals are scaled for direct comparison.

Panel of Figure 5 — Figure 5: Variance analysis across LoD. For both MGS and ours, we report the inter and intra LoD variances on MipNerf( lower is better ). Our method consistently outperforms MGS.

	Components			AUC [0.1%, 100%]			100% Capacity
	Dens.	Score	Var.	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
None	✗	✗	✗	28.74	0.855	0.196	29.55	0.863	0.164
Single	✓	✗	✗	28.81	0.859	0.190	29.79	0.889	0.151
	✗	✓	✗	28.94	0.861	0.188	29.62	0.882	0.159
	✗	✗	✓	28.79	0.855	0.190	29.59	0.870	0.160
Pair	✓	✓	✗	28.99	0.862	0.187	29.75	0.885	0.154

Table 4: Full ablation. All combinations of complexity-aware densification (Dens.), importance scoring (Score), and the spatial variance loss (Var.) on MipNerf.

Figure 6: Time to quality in streaming. Rendering quality of Bonsai [ 34 ] , over a

50

Mbps link. Prefix methods send each primitive once, so every arrival refines the model, and ours reaches every quality level first. FlexGS resends the full set for each budget, so time is delayed and quality oscillates.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Value
Spatial-variance loss weight $\lambda_{\mathrm{var}}$	$2.0$
Importance-score weight $\lambda$	$2.0$
Complexity EMA rate $\eta$	$0.8$
Training LoD sampling distribution	$\tau\sim\mathcal{U}(0,100\%)$
Dropout rate on normalized complexity $\tilde{C}$	$0.2$
Training steps	30000

Appendix

Table 5: Hyperparameters used in our method.

Figure 7: Qualitative comparison on Garden. Renders of the outdoor Garden scene from MipNerf360 [ 34 ] at four budgets.

Figure 8: Qualitative comparison on Room. Renders of the indoor Room scene from MipNerf360 [ 34 ] at four budgets.

		100% Capacity			AUC [0.1%, 100%]
Dataset	$\delta$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
TandT [ 35 ]	0	24.52	0.856	0.153	23.91	0.829	0.193
	0.2	24.65	0.868	0.150	24.02	0.839	0.189
	0.4	24.46	0.852	0.152	23.86	0.828	0.192
	0.6	24.44	0.847	0.163	23.85	0.818	0.195
	0.8	24.39	0.841	0.171	23.80	0.811	0.199

Appendix

Table 6: Complexity dropout rate. Sweep of the dropout rate

\delta

on TandT and DB.

\delta=0

always applies the complexity term and

\delta=1

reduces to opacity sampling. Our default is shaded.

	Indoor				Outdoor
Method	$100\%$	$10\%$	$1\%$	$0.1\%$	$100\%$	$10\%$	$1\%$	$0.1\%$
FlexGS ∗ [ 5 ]	339	578	833	–	351	609	1021	–
LightGaussian [ 15 ]	300	379	697	1101	189	435	855	1364
3DGS MCMC [ 8 ]	300	568	688	732	189	805	1093	1177
MGS [ 2 ]	302	595	731	829	193	769	1050	1151
Ours	302	450	475	463	198	649	833	874

Appendix

Table 7: Rendering speed. Average frame rate (FPS

\uparrow

) on MipNerf360 at four levels of detail. Indoor scenes are rendered at half resolution and outdoor scenes at quarter resolution. FlexGS is excluded from the ranking.

Method	Statistic	$[0,1)$	$[1,2)$	$[2,3)$	$[3,4)$	$[4,5]$
MGS [ 2 ]	Count	22.9k	110.0k	240.1k	466.5k	460.5k
MGS [ 2 ]	Share	1.8%	8.5%	18.5%	35.8%	35.4%
Ours	Count	21.1k	100.7k	233.8k	471.5k	472.9k
Ours	Share	1.6%	7.7%	18.0%	36.3%	36.4%

Appendix

Table 8: Full capacity allocation. Number and share of Gaussians per complexity bin at a budget of 1300k. Our densification shifts Gaussians towards the most complex bins.

	Components			AUC [0.1%, 100%]			100% Capacity			10% Capacity			1% Capacity			0.1% Capacity
	Dens.	Score	Var.	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
None	✗	✗	✗	28.74	0.855	0.196	29.55	0.863	0.164	27.02	0.794	0.275	20.12	0.543	0.540	14.59	0.393	0.623
Single	✓	✗	✗	28.81	0.859	0.190	29.79	0.889	0.151	27.53	0.805	0.263	20.76	0.560	0.511	15.68	0.424	0.616
	✗	✓	✗	28.94	0.861	0.188	29.62	0.882	0.159	27.69	0.818	0.259	20.95	0.564	0.511	15.94	0.457	0.615
	✗	✗	✓	28.79	0.855	0.190	29.59	0.870	0.160	27.44	0.804	0.268	20.39	0.553	0.520	15.10	0.412	0.621
Pair	✓	✓	✗	28.99	0.862	0.187	29.75	0.885	0.154	27.75	0.819	0.259	20.99	0.565	0.510	16.08	0.457	0.615

Appendix

Table 9: Full ablation. All combinations of complexity aware densification (Dens.), importance scoring (Score), and the spatial variance loss (Var.) on MipNeRF360.

Table 10: Coverage across LoDs. Number of "hole" pixels, i.e. , pixels whose accumulated opacity is below

0.05

, on the test set of the indoor Bonsai and outdoor Bicycle scenes [ 34 ] . Lower is better.

3D Gaussian Splatting

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

See all 1,019 papers

Matryoshka Gaussian Splatting
Mar 19, 2026Zhilin Guo, Boqiao Zhang, Hakan Aktas +113D Gaussian Splatting 3DGS Training
University of Cambridge · Chalmers University of Technology · Google
RLSplat: Resource-Linear Budget Control for Feed-forward 3D Gaussian Splatting
Oct 20, 2026Injae Kim, Minseok Joo, Jihwan Park +33D Gaussian Splatting Feed-Forward 3D Reconstruction
KAIST · Korea University
ParetoGS: High-Quality Continuous-Budget 3D Gaussian Splatting with Sparse Adapted Anchors
Oct 25, 2026Xiaobin Deng, Changyu Diao, Min Li +13D Gaussian Splatting 3DGS Compression
Zhejiang University

Mar 19, 2026J/K move · Enter open · S save

Matryoshka Gaussian Splatting

Zhilin Guo, Boqiao Zhang, Hakan Aktas, Kyle Fogarty+10

University of Cambridge · Chalmers University of Technology · Google

3D Gaussian Splatting 3DGS Training 3DGS Acceleration Matryoshka Representation Learning

Open paper AskPDF arXiv

The ability to render scenes at adjustable fidelity from a single model, known as level of detail (LoD), is crucial for practical deployment of 3D Gaussian Splatting (3DGS). Existing discrete LoD methods expose only a limited set of operating points, while concurrent continuous LoD approaches enable smoother scaling but often suffer noticeable quality degradation at full capacity, making LoD a costly design decision. We introduce Matryoshka Gaussian Splatting (MGS), a training framework that enables continuous LoD for standard 3DGS pipelines without sacrificing full-capacity rendering quality. MGS learns a single ordered set of Gaussians such that rendering any prefix, the first k splats, produces a coherent reconstruction whose fidelity improves smoothly with increasing budget. Our key idea is stochastic budget training: each iteration samples a random splat budget and optimises both the corresponding prefix and the full set. This strategy requires only two forward passes and introduces no architectural modifications. Experiments across four benchmarks and six baselines show that MGS matches the full-capacity performance of its backbone while enabling a continuous speed-quality trade-off from a single model. Extensive ablations on ordering strategies, training objectives, and model capacity further validate the designs.

Abstract

3D Gaussian Splatting

Explore similar work

Matryoshka Gaussian Splatting

RLSplat: Resource-Linear Budget Control for Feed-forward 3D Gaussian Splatting

ParetoGS: High-Quality Continuous-Budget 3D Gaussian Splatting with Sparse Adapted Anchors