GoDe: Gaussians on Demand for Backbone-Agnostic Progressive Compression

Francesco Di Sario; Riccardo Renzulli; Marco Grangetto; Akihiro Sugimoto; Enzo Tartaglione

cs.CVJan 23, 2025

GoDe: Gaussians on Demand for Backbone-Agnostic Progressive Compression

Authors: Francesco Di Sario, Riccardo Renzulli, Marco Grangetto, Akihiro Sugimoto, Enzo Tartaglione

Organizations: Computer Science, University of Turin, Corso Svizzera, Turin, 10149, Italy. · National Institute of Informatics, 2 Chome-1-2 Hitotsubashi, Tokyo, 101-8430, Japan. · LTCI, T´el´ecom Paris, Institut Polytechnique de Paris, 19 Pl. Marguerite Perey, Paris, 91120, France.

Abstract

Splat-based radiance fields have rapidly diversified in primitive type, kernel support, and appearance parameterization. This diversity makes compression difficult: most codecs are single-rate, requiring a separately optimized model for each target bitrate, while progressive codecs remain tied to representation-specific assumptions that do not transfer across backbones. We introduce GoDe (Gaussians on Demand), a progressive compression framework whose central contribution is backbone agnosticism. It reorganizes any trained differentiable splat representation into a single scalable bitstream without modifying the backbone or re-deriving the compression structure for each new primitive type. GoDe builds a nested level-of-detail hierarchy via masked-gradient importance ranking, refines all levels with one quantization-aware fine-tuning stage, and stores the base layer and enhancement layers as progressive chunks compressed with shared entropy dictionaries. Any decoded prefix is directly renderable, so one training pipeline provides multiple rate-distortion operating points without per-rate retraining. Across five heterogeneous backbones and three standard benchmarks, GoDe achieves competitive rate-distortion performance while enabling faster rendering and substantially faster progressive decoding than existing methods. Project page: https://gaussians-on-demand.github.io/

Figures & tables

Figure 1: One training run, a full rate curve. GoDe constructs a progressive hierarchy from any trained splat representation; each decoded prefix yields a distinct quality–size–speed operating point with no per-rate retraining.

Method	Prog. render.	Pruning	Codec	Prog. comp.	Agnostic
Progressive rendering / LoD – no compressed bitstream
CLoD-GS	✓	✗	✗	✗	✗
PRoGS	✓	✗	✗	✗	✗
LapisGS	✓	✗	✗	✗	✗
Octree-GS	✓	✗	✗	✗	✗
Hierarchical-GS	✓	✗	✗	✗	✗

Table 1: Method positioning. Qualitative comparison of GoDe against existing approaches.

Figure 2: GoDe pipeline. From a trained representation: (1) primitives are organized into a nested LoD hierarchy via iterative gradient-informed ranking; (2) all LoDs are jointly fine-tuned under simulated quantization, sampling one LoD per iteration; (3) the base layer and each enhancement layer are entropy-coded as independent chunks.

Method	Rate	Mip-NeRF360					Tanks&Temples					DeepBlending
Method	Rate	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$	FPS $\uparrow$
Non-scalable – Neural
HEMGS † [ 24 ]	Low	27.75	0.806	0.248	12.50	–	24.42	0.848	0.192	6.00	–	30.24	0.908	0.266	3.00	–
HEMGS † [ 24 ]	High	27.93	0.813	0.230	21.00	–	24.58	0.856	0.176	10.10	–	30.37	0.911	0.253	6.70	–
HAC++ [ 8 ]	Low	27.60	0.803	0.253	8.34	70	24.22	0.849	0.190	5.18	79	30.16	0.907	0.266	2.91	152
HAC++ [ 8 ]	High	27.82	0.811	0.231	18.48	71	24.32	0.854	0.178	8.63	78	30.34	0.911	0.254	5.28	154

Table 2: Main quantitative comparison. Each row of a non-scalable method corresponds to an independently optimized operating point, whereas scalable methods obtain multiple rates from a single trained model. For scalable methods, we show three representative decoded prefixes spanning the range of competing operating points; rate labels denote the selected operating points, not LoD indices. GoDe uses

8

levels in the main experiments.

\dagger

denotes methods whose public implementations were unavailable at submission time.

Figure 3: Rate–distortion and speed–distortion curves. Each point is averaged over the dataset; solid lines denote scalable methods, dashed lines denote non-scalable ones.

Method	Scalable	Train $\downarrow$		Encoding $\downarrow$			Decoding $\downarrow$
		Per run	All	Low	Medium	High	Low	Medium	High
Context-GS	✗	1.3	4.0	52.7	87.8	97.2	53.1	82.5	88.1
HAC++	✗	1.3	3.9	8.9	10.1	16.4	12.3	13.4	19.0
RDO	✗	1.3	3.9	1.1	2.9	4.8	2.2	7.4	19.0
Comp-GS	✗	1.6	4.9	10.8	10.8	10.9	9.1	9.0	8.8
PCGS †	✓	1.83	1.83	74.49			42.3	46.3	52.1

Table 3: Training, encoding, and decoding time. Training is in hours; coding is in seconds. † PCGS uses one shared scalable encoding for all bitrate levels.

Figure 4: Cross-backbone transfer on Mip-NeRF360. GoDe and independently optimized same-backbone models matched by primitive count. Results are averaged over the nine scenes.

Figure 5: Level-of-detail progression. High-resolution crops show how coarse levels preserve global structure while enhancement layers progressively restore fine geometry and view-dependent appearance.

Figure 6: Smooth level transitions. Opacity interpolation progressively activates an enhancement layer, reducing visible popping between adjacent LoDs.

Figure 7: Importance-criterion ablation on Mip-NeRF360. Our criterion consistently preserves higher reconstruction quality at matched active primitive counts than established Gaussian-pruning criteria.

Figure 8: Cost of scalability. BD-PSNR on Mip-NeRF360 comparing scalable GoDe with a matched-budget non-scalable variant.

		Mask only		$+$ Fine-tuning		$+$ Compression (GoDe)
Rate	Retained	PSNR $\uparrow$	Size [MB] $\downarrow$	PSNR $\uparrow$	Size [MB] $\downarrow$	PSNR $\uparrow$	Size [MB] $\downarrow$
0	0.13	23.45	70.95	27.06	70.95	26.97	11.31
1	0.16	24.22	91.57	27.41	91.57	27.32	14.29
2	0.19	25.05	119.42	27.60	119.42	27.50	18.16
3	0.24	25.88	157.32	27.70	157.32	27.60	23.21
4	0.30	26.67	209.19	27.77	209.19	27.66	29.77

Table 4: Impact of multi-level fine-tuning and compression on Mip-NeRF360 with 3DGS. Top: masked LoDs before and after joint fine-tuning. Bottom: quantitative results and final compression. Retained indicates the fraction of Gaussians relative to the full model; GoDe denotes the final compressed representation.

Figure S1 : Score-normalization ablation on Mip-NeRF360. PSNR at matched active primitive counts for the proposed normalized multi-parameter score, its unnormalized counterpart, and single-attribute criteria.

Figure S2 : Number of levels. PSNR versus active primitive count on the Mip-NeRF360 bonsai scene for

L\in\{4,8,16,32\}

Figure S3 : Lower hierarchy bound. PSNR versus active primitive count on Mip-NeRF360 for different values of

n_{\min}

Retained	1.00	0.90	0.80	0.75	0.65	0.60	0.55	0.50	0.45	0.40	0.35	0.30	0.25	0.20	0.63 (Adaptive)
PSNR $\uparrow$	27.82	27.82	27.82	27.82	27.81	27.79	27.72	27.59	27.37	27.03	26.53	25.85	24.97	23.83	27.82
SSIM $\uparrow$	0.830	0.830	0.830	0.830	0.830	0.829	0.827	0.824	0.819	0.812	0.802	0.789	0.771	0.746	0.830
LPIPS $\downarrow$	0.239	0.239	0.239	0.240	0.240	0.242	0.244	0.248	0.254	0.261	0.272	0.285	0.304	0.329	0.239

Table S1 : Adaptive upper-bound analysis on Mip-NeRF360. Dataset-averaged reconstruction quality after retaining a fixed fraction of the pretrained Gaussians, compared with the proposed scene-adaptive criterion. The adaptive rule retains

62.864\%

of primitives on average while preserving full-model quality at the reported precision.

Dataset	Method	Wall-clock time (s)	Ranking time (s)	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Mip-NeRF360	PUP	361	16	26.55	0.784	0.271
	GoDe	315	9	26.88	0.795	0.263
Tanks&Temples	PUP	215	17	22.45	0.797	0.251
	GoDe	183	6	22.88	0.804	0.247
DeepBlending	PUP	313	15	29.57	0.904	0.260
	GoDe	279	9	29.58	0.902	0.262

Table S2 : Comparison under the PUP-3DGS protocol. Both methods use the same pretrained checkpoints,

90\%

total pruning, and a

10

k-iteration optimization budget.

Figure S4 : Multi-level convergence on Mip-NeRF360. PSNR across the eight decoded prefixes at different optimization iterations, averaged over the nine scenes.

Rate	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
0	$26.963\pm 0.015$	$0.7812\pm 0.0003$	$0.3219\pm 0.0002$
1	$27.329\pm 0.016$	$0.7969\pm 0.0002$	$0.3021\pm 0.0002$
2	$27.493\pm 0.014$	$0.8071\pm 0.0001$	$0.2858\pm 0.0003$
3	$27.606\pm 0.008$	$0.8129\pm 0.0001$	$0.2742\pm 0.0003$
4	$27.653\pm 0.011$	$0.8171\pm 0.0002$	$0.2650\pm 0.0001$
5	$27.697\pm 0.010$	$0.8188\pm 0.0003$	$0.2581\pm 0.0001$

Table S3 : Seed robustness on Mip-NeRF360. Mean and standard deviation over five independent runs of GoDe (3DGS) across all eight decoded levels.

Figure S5 : Cross-backbone transfer across datasets. Rate–distortion comparison of GoDe with independently optimized same-backbone models for BS+SV, Triangle Splatting, and Octree-GS on Mip-NeRF360, Tanks&Temples, and DeepBlending. Independent models are shown before and after the same quantization and compression pipeline used by GoDe; the original model is included with and without lossless Zstandard compression for reference.

Dataset	Rate	GoDe (BS+SV)				Independent training w/ compression				Independent training w/o compression
Dataset	Rate	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$
Mip-NeRF360	2	27.53	0.785	0.323	10.01	27.87	0.798	0.310	11.11	27.90	0.799	0.307	73.97
	4	28.23	0.818	0.269	23.54	28.35	0.821	0.259	24.26	28.38	0.822	0.257	183.58
	7	28.38	0.826	0.239	64.63	28.47	0.827	0.231	67.32	28.57	0.829	0.227	569.66
Tanks&Temples	2	24.37	0.841	0.240	8.00	24.53	0.847	0.230	9.42	24.60	0.849	0.228	64.03
	4	24.69	0.856	0.204	15.57	24.75	0.859	0.198	16.42	24.82	0.862	0.193	131.43

Table S4 : Cross-backbone transfer with BS+SV. Independent reference models are separately trained for

30

k iterations at the primitive counts corresponding to GoDe levels R2, R4, and R7, and are reported before and after applying the same quantization and compression pipeline used by GoDe.

Dataset	Rate	GoDe (Triangle-Splatting)				Independent training w/ compression				Independent training w/o compression
Dataset	Rate	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$
Mip-NeRF360	2	24.68	0.685	0.425	9.75	25.41	0.718	0.385	10.21	25.42	0.718	0.384	79.00
	4	25.91	0.743	0.348	24.83	26.15	0.756	0.328	23.15	26.17	0.756	0.327	186.79
	7	26.42	0.772	0.304	87.91	26.76	0.783	0.276	86.73	26.75	0.783	0.275	691.51
Tanks&Temples	2	21.97	0.786	0.301	9.44	22.03	0.793	0.303	8.04	22.04	0.793	0.301	68.44
	4	22.43	0.811	0.265	18.14	22.52	0.817	0.256	16.68	22.55	0.818	0.254	138.91

Table S5 : Cross-backbone transfer with Triangle Splatting. Independent reference models are separately trained for

30

k iterations at the primitive counts corresponding to GoDe levels R2, R4, and R7, and are reported before and after applying the same quantization and compression pipeline used by GoDe.

Dataset	Rate	GoDe (Octree-GS)				Independent training w/ compression				Independent training w/o compression
Dataset	Rate	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$
Mip-NeRF360	2	26.75	0.779	0.309	11.21	27.51	0.806	0.283	13.61	27.67	0.808	0.279	53.78
	4	27.77	0.815	0.265	21.72	27.82	0.817	0.262	21.90	27.91	0.819	0.259	87.60
	7	27.84	0.819	0.257	36.22	27.91	0.819	0.253	36.10	28.04	0.821	0.250	149.80
Tanks&Temples	2	23.45	0.830	0.288	8.16	23.36	0.828	0.284	6.93	23.35	0.829	0.280	27.18
	4	24.14	0.862	0.232	14.69	24.01	0.858	0.236	12.72	24.04	0.859	0.231	50.32

Table S6 : Cross-backbone transfer with Octree-GS. Independent reference models are separately trained for

30

k iterations at the primitive counts corresponding to GoDe levels R2, R4, and R7, and are reported before and after applying the same quantization and compression pipeline used by GoDe.

		Mip-NeRF360					Tanks&Temples					DeepBlending
Backbone	LoD	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size [MB] $\downarrow$	FPS $\uparrow$
Scaffold-GS	0	27.29	0.792	0.310	9.81	104	23.17	0.812	0.281	4.13	171	30.00	0.901	0.351	3.92	183
	1	27.62	0.802	0.297	11.93	96	23.68	0.829	0.259	5.10	160	30.27	0.905	0.343	4.63	178
	2	27.79	0.808	0.288	14.57	89	23.98	0.840	0.242	6.30	148	30.40	0.908	0.337	5.48	171
	3	27.87	0.811	0.282	17.85	83	24.16	0.847	0.229	7.80	137	30.49	0.909	0.332	6.49	166
	4	27.92	0.813	0.278	21.95	78	24.25	0.851	0.220	9.66	125	30.53	0.910	0.329	7.70	159

Table S7 : Full per-LoD results of GoDe. Dataset-averaged results for all eight decoded levels and all five backbones. Size is cumulative and includes all chunks required to decode the corresponding level.

Panel of Figure 25 — Figure S6 : Full LoD progression on DeepBlending. All eight decoded levels using the 3DGS-MCMC backbone.

3DGS Compression

3D Gaussian Splatting compression reduces storage and rendering costs while preserving the quality of reconstructed views and scenes.

See all 80 papers

ProGS: Towards Progressive Coding for 3D Gaussian Splatting
Mar 10, 2026Zhiye Tang, Lingzhuo Liu, Shengjie Jiao +43DGS Compression 3D Gaussian Splatting
College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, 518060, China · Department of Computer Science, City University of Hong Kong, Kowloon, Hong Kong SAR, China · School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan, 430074, China
CodecSplat: Ultra-Compact Latent Coding for Feed-Forward 3D Gaussian Splatting
May 25, 2026Pengpeng Yu, Runqing Jiang, Qi Zhang +33DGS Compression 3D Gaussian Splatting
1Sun Yat-sen University, China · 3Pengcheng Laboratory, China · 2Peking University, China
PLEDGE: Progressive Lossy Error-Bounded Compression of 3D Gaussian Splatting
Oct 13, 2026Mingze Xia, Yuxiao Li, Baixi Sun +6Quantization 3DGS Compression
Oregon State University · The Ohio State University · Argonne National Laboratory +1

Mar 10, 2026J/K move · Enter open · S save

ProGS: Towards Progressive Coding for 3D Gaussian Splatting

Zhiye Tang, Lingzhuo Liu, Shengjie Jiao, Qiudan Zhang+3

College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, 518060, China · Department of Computer Science, City University of Hong Kong, Kowloon, Hong Kong SAR, China · School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan, 430074, China

3DGS Compression 3D Gaussian Splatting Entropy Coding

Open paper AskPDF arXiv

Progressive transmission of 3D Gaussian Splatting (3DGS) requires each completed transmission stage to be decodable from received data and directly renderable. This work presents ProGS, a progressive codec that organizes anchor-based 3DGS as parent-closed octree prefixes. ProGS combines parent-causal entropy coding, level-balanced anchor growth, bounded multi-prefix training, and lightweight parent-anchor refinement to improve early-prefix quality without altering the complete-model rendering path. One fixed-

λ

training run yields five deployable rate--quality points from a single bitstream. Experiments on 17 scenes from three datasets evaluate rate--distortion performance, rendering speed, and transmission efficiency against progressive and single-rate baselines. On one representative scene per dataset, ProGS reaches a common quality target with 30.7

\sim

60.7% fewer bytes than HAC-Rand and 22.6

\sim

53.4% fewer bytes than HAC++-Rand. Across the three dataset averages, ProGS-LR uses 4.7

\sim

6.1% fewer bytes than HAC-high while improving SSIM by 0.002

\sim

0.041 and reducing LPIPS by 4.8

\sim

51.2%. The parent-closed syntax makes every prefix causally decodable and directly renderable without future topology. ProGS-HR also yields higher endpoint SSIM and lower LPIPS than PCGS across all three dataset averages, and all five prefixes render in real time. Code is available at https://github.com/ZhiyeTang/ProGS-Official

Abstract

3DGS Compression

Explore similar work

ProGS: Towards Progressive Coding for 3D Gaussian Splatting

CodecSplat: Ultra-Compact Latent Coding for Feed-Forward 3D Gaussian Splatting

PLEDGE: Progressive Lossy Error-Bounded Compression of 3D Gaussian Splatting