ProGS: Towards Progressive Coding for 3D Gaussian Splatting

Zhiye Tang; Lingzhuo Liu; Shengjie Jiao; Qiudan Zhang; Junhui Hou; You Yang; Xu Wang

cs.CVMar 10, 2026

ProGS: Towards Progressive Coding for 3D Gaussian Splatting

Authors: Zhiye Tang, Lingzhuo Liu, Shengjie Jiao, Qiudan Zhang, Junhui Hou, You Yang, Xu Wang

Organizations: 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

Abstract

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

Figures & tables

Fig. 1 : ProGS structures a 3D scene as a forest of layered octrees in a shared coordinate system. Each level-of-detail (LoD) $l$ contains the cumulative octree layers through $l$ . With context-based progressive coding, every completed level chunk yields a decodable and renderable model at a larger received payload.

Fig. 2 : Overview of ProGS. Stage 1 constructs parent-closed octree prefixes via level-balanced growth and multi-LoD supervision, then entropy-codes anchor attributes with priors and quantization steps predicted from the decoded parent prefix and current-level structure (AE/AD: arithmetic encoding/decoding). Stage 2 freezes the base codec and predicts bounded attribute residuals for selected low-LoD anchors from uniformly quantized per-anchor tokens, leaving octree positions and topology unchanged; refinement is bypassed at the complete LoD. A 2D quadtree is shown for clarity; the method uses octrees in 3D.

Fig. 3 : Geometric operations in level-balanced strict-octree growth (2D quadtree planes). (i) Same-level candidates are generated from Gaussian offsets. (ii) Finer-level candidates are quantized to child cells; a retained child without a parent queues the missing parent before materializing. (iii) Low-opacity anchors are pruned only as leaves, preserving parent closure. Prefix-aware scoring, closure penalties, and the top- $B_{l}$ budget are specified in Algo. 1 .

Fig. 4 : Training schedule of ProGS. Stage 1 optimizes the main codec for 40,000 iterations: the parent-causal entropy context is enabled at iteration 100, and level-balanced growth with multi-LoD supervision runs from iteration 1,500 to 25,000, with growth alone paused from 13,000 to 14,000 per the HAC-style densification schedule. From iteration 25,000 the octree is frozen and only the full-LoD path is optimized. Stage 2 then freezes the base codec and refines low LoDs for 5,000 iterations.

Method / category	Rate organization	Cumulative chunks	Coded-prefix renderable	Parent-closed octree	Decoding dependency
Reference representations	Single endpoint	No	No	No	–
General single-rate compression	Independent models	No	No	N/R	Method-specific
Entropy-coded single-rate compression	Independent models	No	No	No	Non-progressive
Octree-GS (official) [ 13 ]	Adaptive LoDs	No	N/A	No	View-dependent LoD
Octree-GS-EC	Cumulative LoD groups	Yes	Yes	No	Decoded previous LoD
PCGS [ 14 ]	Cumulative levels	Yes	Yes	No	Previous layers

TABLE I: Rate organization of the compared methods. Independent rate points require separate optimization. Octree-GS-EC is the entropy-coded adaptation of Octree-GS used in this comparison. Here, a coded prefix denotes a serialized cumulative bitstream prefix; N/A for official Octree-GS indicates that it reports renderable LoDs rather than coded prefixes. N/R: property not uniformly reported within a group.

	Mip-NeRF360 [ 41 ]				BungeeNeRF [ 42 ]				Tanks and Temples [ 43 ]
Method / point	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$
Reference representations
3DGS [ 1 ]	27.46	.812	.222	750.90	24.87	.841	.205	1616.00	23.69	.844	.178	431.00
Scaffold-GS [ 11 ]	27.50	.806	.252	253.90	26.62	.865	.241	183.00	23.96	.853	.177	86.50
General single-rate compression
SOG [ 19 ]	26.56	.791	.241	16.70	22.43	.708	.339	48.25	23.15	.828	.198	9.30

TABLE II : Main quantitative comparison. Top three entries per column are bold , underlined , and italic ; ties share a rank. Octree-GS-EC denotes the entropy-coded adaptation of Octree-GS used in this comparison. For Octree-GS-EC and ProGS, LR/HR use

\lambda=0.004/0.0005

; full denotes the highest available LoD. Official Octree-GS and the reference representations report checkpoint storage; PCGS, Octree-GS-EC, and ProGS report cumulative coded payloads; the remaining compression methods report complete-model storage.

Fig. 5 : Dataset-average SSIM–rate (top) and PSNR–rate (bottom). The two blue ProGS curves are the lv0–lv4 prefixes of fixed-

\lambda

LR and HR models. Solid lines connect only successive points of one cumulative progressive stream (ProGS, PCGS, and Octree-GS-EC); isolated markers are separately optimized single-rate models. Scenes with fewer occupied Octree-GS-EC levels retain their final LoD. Piecewise-linear Y-axes compress the sparse low-quality range; a kink on each left spine marks the scale transition, and ticks remain in the original units.

Fig. 6 : Progressive PSNR versus cumulative received size on matched scenes. ProGS and PCGS use native prefixes; Octree-GS-EC-LR uses decoded cumulative LoD groups. HAC/HAC++ random packetizations report ten-seed mean and standard deviation. The common target

Q_{s}^{*}

is defined over these five methods in Eq. ( 33 ). Piecewise-linear Y-axes compress the sparse low-quality range; a kink on each left spine marks the scale transition, and ticks remain in dB.

Panel of Figure 9 — Fig. 7 : Progressive decoding on representative test views; each row shows the scene context, ground-truth detail, and the crop from lv0 to lv4. Red boxes mark the detail region; gray tags report full-view PSNR (dB) and cumulative size (MB).

Panel of Figure 10 — Fig. 8 : Early-prefix comparison on representative test views. Each row shows the scene context and a shared detail crop. Red boxes mark the detail region; gray tags report full-view PSNR (dB) and cumulative size (MB). The selected PCGS prefixes/decoded Octree-GS-EC-LR LoDs are lv1/LoD2, lv2/LoD0, and lv2/LoD0, respectively.

Setting	Closure	Par. feat. context	Entropy context	Level balance	Prefix superv.	Refine.
Loose hierarchy	Loose	Off	Off	Off	Bounded	Off
Strict hierarchy	Strict	Off	Off	Off	Bounded	Off
w/o entropy context	Strict	On	Off	On	Bounded	On
w/o level balance	Strict	On	On	Off	Bounded	On
w/o refinement	Strict	On	On	On	Bounded	Off
w/o prefix supervision	Strict	On	On	On	None	On

TABLE III : Configurations of the ablation variants. Bounded prefix supervision runs from iteration 1,500 to 25,000; None uses only the full-model loss.

Setting	Base full size (MB)	$\Delta$ full size (MB)	lv0 prefix size (MB)	$\Delta$ lv0 size (MB)	Full PSNR (dB)	$\Delta$ full (dB)	lv0 PSNR (dB)	$\Delta$ lv0 (dB)
Loose hierarchy	16.718	$-5.047$	4.199	$-6.636$	26.696	$-0.322$	21.877	$-2.277$
Strict hierarchy	34.724	$+12.958$	19.482	$+8.647$	27.300	$+0.282$	22.696	$-1.458$
Full ProGS	21.766	–	10.747	–	27.018	–	24.154	–
w/o entropy context	22.412	$+0.647$	10.778	$+0.032$	27.054	$+0.036$	24.135	$-0.019$
w/o level balance	24.376	$+2.610$	13.470	$+2.723$	27.156	$+0.138$	24.321	$+0.167$
w/o refinement	21.766	$0.000$	10.365	$-0.382$	27.018	$0.000$	22.602	$-1.552$

TABLE IV : Absolute ablation results. Prefix size is cumulative through lv0 and includes the refinement side stream when enabled; base full size excludes it to isolate the Stage 1 codec. Each delta is the variant value minus Full ProGS.

Fig. 9 : Average parameter sensitivity. From left to right and top to bottom, the panels scan the growth threshold, level exponent, pruning threshold, balance ratio, minimum budget, multi-LoD schedule, token dimension, and active ratio; the two-setting quantization comparison is reported in Tab. V . The gray band marks the default setting in each panel; full denotes the base complete path and refined lv0–lv2 the frozen-base refined prefixes.

Fig. 10 : Average bitstream allocation. Left: shares by parameter type. Right: shares by progressive chunk. The refinement side stream follows lv0, is shown separately, and remains in subsequent cumulative payloads.

Metric	lv0	lv1	lv2	lv3	lv4
Size (MB)	10.75	15.42	18.23	20.62	22.15
Cum. Dec. (s)	12.84	19.62	23.72	26.68	28.49
FPS	120.2	114.1	105.0	143.6	149.8
$\mathrm{TTFF}_{50}$ (s)	14.57	22.10	26.65	29.99	32.04
Train: 3.35 h			Ref.: 0.382 MB; anchors: 555k

TABLE VI : Mean ProGS prefix efficiency on one RTX 4090 (17 scenes;

\lambda_{\mathrm{e}}=0.004

). lv0–lv4 are the first through fifth renderable levels.

\mathrm{TTFF}_{50}

represents the time-to-first-frame from the start of transmission over a 50-Mbps link.

3DGS Compression

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

See all 80 papers

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
MesonGS++: Post-training Compression of 3D Gaussian Splatting with Hyperparameter Searching
Apr 29, 2026Shuzhao Xie, Junchen Ge, Weixiang Zhang +103DGS Compression 3D Gaussian Splatting
Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China · Harbin Institute of Technology, Shenzhen, China +6
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

May 25, 2026J/K move · Enter open · S save

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

Pengpeng Yu, Runqing Jiang, Qi Zhang, Dingquan Li+2

1Sun Yat-sen University, China · 3Pengcheng Laboratory, China · 2Peking University, China

3DGS Compression 3D Gaussian Splatting Representation Learning Feed-Forward 3D Reconstruction Gaussian Splatting Compression

Open paper AskPDF arXiv

While feed-forward 3D Gaussian splatting reconstructs renderable Gaussian primitives from sparse context views without per-scene optimization, existing pipelines do not provide a compact scene representation for storage or transmission. A natural solution is to apply existing 3DGS compression methods to the generated Gaussian primitives. However, this approach operates on the final irregular 3D representation and is decoupled from the internal feature-to-Gaussian generation process, which limits compression efficiency. To address this, we introduce CodecSplat, an ultra-compact latent coding framework for feed-forward 3D Gaussian splatting. CodecSplat first encodes an intermediate 2D Gaussian-generation feature into an entropy-coded scene bitstream. At the decoder, the latent feature is reconstructed and used to predict depth and Gaussian parameters, which are then mapped to 3D Gaussian primitives. By integrating compression into the feed-forward Gaussian generation pipeline, CodecSplat avoids inefficient compression over irregular 3D Gaussian primitives and allows the codec to exploit the structured intermediate feature representation. We instantiate CodecSplat on a feed-forward Gaussian splatting backbone with depth-guided multi-view feature refinement and a hierarchical learned feature codec. On DL3DV and RealEstate10K datasets, CodecSplat achieves 23.56-26.36 dB and 24.76-27.05 dB PSNR with only 20.00-107.77 KiB and 3.37-12.51 KiB per scene, respectively. This is roughly one order of magnitude smaller than compressing feed-forward generated Gaussian primitives, while preserving controllable rate-distortion behavior. Code is available at https://github.com/pengpeng-yu/CodecSplat.

Abstract

3DGS Compression

Explore similar work

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

MesonGS++: Post-training Compression of 3D Gaussian Splatting with Hyperparameter Searching

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