BoundedGS: Memory-Efficient High-Resolution Training for 3D Gaussian Splatting

Xiaobin Deng; Changyu Diao; Min Li; Duanqing Xu

cs.CVOct 25, 2026

BoundedGS: Memory-Efficient High-Resolution Training for 3D Gaussian Splatting

Authors: Xiaobin Deng, Changyu Diao, Min Li, Duanqing Xu

Organizations: Zhejiang University

Abstract

High-resolution 3D Gaussian Splatting training increases pixel and Gaussian-tile intersection counts, expanding rasterization caches that coexist with projection and gradient state at peak memory. A complete-view update aggregates gradients across the view; each region's forward and backward caches can be released after its contribution is collected. BoundedGS separates rasterization and projection backward passes and bounds the active local rasterization working set according to its measured load. Candidate-indexed projection and compact gradient storage address the view-level peaks exposed by this change, while CPU-assisted structure rebuilding and bounded image supply manage resources outside rendering. In a resolution sweep at a fixed count of about 6M Gaussians, the measured rendering working set stays within 1.073-1.077 GiB from 1K to 16K; at 11K, a warm pass takes 19.79 seconds versus 21.76 for full-frame execution. Across five native 5K outdoor scenes, full training reduces the mean rendering working set from 9.97 to 1.76 GiB (82.3%), retains similar reconstruction quality, and shortens training time from 36.7 to 33.9 minutes. Project page: https://xiaobin2001.github.io/BoundedGS-Web/

Figures & tables

Figure 1: Original working-set trace for one view of Bicycle’s 15K SH3 model. The event-order axis and stacked bands show co-resident projection, gradient, and rasterization/supervision/loss state; the red marker gives the 9.55 GiB peak.

Dataset	Method	PSNR $\!\uparrow$	SSIM $\!\uparrow$	LPIPS $\!\downarrow$	Time $\!\downarrow$	Mem $\!\downarrow$
Mip-NeRF 360 Outdoors (5K)	Original	24.44	0.739	0.314	36.7	9.97
	BoundedGS	24.42	0.739	0.314	33.9	1.76
Mip-NeRF 360 Indoors (3K)	Original	31.52	0.932	0.239	16.5	4.43
	BoundedGS	31.64	0.932	0.239	15.9	0.87
Tanks and Temples (4K)	Original	22.92	0.764	0.294	23.2	7.65
	BoundedGS	22.91	0.764	0.294	23.0	1.79

Table 1: Full-training quality, time, and rendering working set. Groups equally average five Mip-NeRF 360 outdoor, four indoor, and four Tanks and Temples scenes. Time is in minutes; Mem is the scene mean of temporary working sets at recorded rendering-stage peaks, in GiB (Appendix A.2 ). Bold marks lower Mem.

Figure 2: High-resolution reconstruction. Columns show ground truth, Original, and BoundedGS; rows show Bicycle view 4 and Train view 7. Insets enlarge marked regions. BoundedGS retains similar local detail.

Configuration	$\mathrm{SSIM}_{\mathrm{Orig}}\uparrow$	Working set (GiB) $\downarrow$	Time (s) $\downarrow$
Original	1.000000	9.98	12.19
Stage separation and early release	0.999991	7.64	11.35
+ Auto Leaf (384 MiB)	0.999991	5.05	12.04
+ Candidate-indexed projection	0.999990	3.27	11.23
+ Compact gradients (BoundedGS384)	0.999991	1.66	10.46
BoundedGS ( $3\times 3$ groups, 128 MiB)	0.999991	1.31	12.69

Table 2: Component ablation on five 5K outdoor scenes. Working set and time are second-pass scene means;

\mathrm{SSIM}_{\mathrm{Orig}}

compares final renderings with Original, averaged over views then scenes. Rows add components sequentially; the last changes both Group layout and Leaf budget.

Figure 3: Component working-set traces for one view of Bicycle’s 15K model. A is Original; B separates backward stages; C–E add Auto Leaf, candidate projection, and compact gradients; F changes Group and Leaf budgets. All variants retain 3DGS’s direct gradient accumulation. The event-order axis and stacked bands show concurrent state; red markers give peaks. A–C and D–F use different vertical scales.

Figure 4: Resolution scaling on Bicycle with the Gaussian count fixed at about 6M (1K = 1024 pixels wide). Left and center show one view’s peak cache composition; right shows 169-view warm time and relative

L_{1}

difference. Original is measured through 11K and reaches OOM at 12K; its higher-resolution dashed curve is extrapolated (Appendix A ). BoundedGS is measured through 16K. Memory excludes persistent state.

Configuration	Render	Rebuild	Host RSS	Time
BoundedGS (Full)	1.76	8.68	13.37	33.95
Repeated GPU rebuild	1.76	12.83	6.30	30.39
Predecoded host images	1.76	8.68	17.60	32.19

Table 3: Full-training resource ablation on five 5K outdoor scenes. Render subtracts persistent state; Rebuild is the total allocated structure-update peak; Host is peak process RSS. All runs use the same rendering path. Memory is in GiB and time in minutes.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Fixed-pixel blocks and Auto Leaf across resolution. All three paths use compact projection and gradients: A updates after each block, B accumulates gradients across fixed blocks for one update per view, and C uses a 384 MiB Auto Leaf with one update per view. The first three panels show peak-time temporary rendering state for the same representative Bicycle view; D shows synchronized second-pass time over 169 training views. Memory for A/B and C was recorded on separate RTX 4090 cards of the same model; all three times were measured on one card.

Figure 6: Independently updated fixed grids on Bicycle’s 11.98M-Gaussian model. The second pass spans 169 views with one Adam update per block. The upper row repeats full projection; the lower row uses stage separation, candidate projection, and compact gradients. Warm gives the all-view peak and time; Trace peak marks one view. Colors show concurrent projection, gradient, and rasterization/loss states.

Scene	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time $\downarrow$	Mem $\downarrow$
Mip-NeRF 360 Outdoors (5K)
bicycle	Original	24.89	0.7583	0.2962	35.1	9.56
	BoundedGS	24.94	0.7585	0.2962	32.0	1.67
flowers	Original	21.34	0.6205	0.3636	37.8	10.38
	BoundedGS	21.33	0.6205	0.3635	35.6	1.97
garden	Original	26.87	0.8201	0.2119	40.4	10.20

Appendix

Table 4: Per-scene results for Original and BoundedGS across 13 scenes. Time is pure training time in minutes and Mem is the temporary rendering working set in GiB; source values are rounded for display.

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

EffGS: Efficient and High-Fidelity Gaussian Splatting
Sep 30, 2026Changbai Li, Shuo Yang, Yichen Yang +23D Gaussian Splatting 3D Gaussian Pruning
Beihang University · Nanyang Technological University
TideGS: Scalable Training of Over One Billion 3D Gaussian Splatting Primitives via Out-of-Core Optimization
May 19, 2026Chonghao Zhong, Linfeng Shi, Hua Chen +43D Gaussian Splatting 3D Scene Representation
Hong Kong University of Science and Technology · Great Wall Motor · Tsinghua University +1
Matryoshka Gaussian Splatting
Mar 19, 2026Zhilin Guo, Boqiao Zhang, Hakan Aktas +113D Gaussian Splatting 3DGS Training
University of Cambridge · Chalmers University of Technology · Google

Sep 30, 2026J/K move · Enter open · S save

EffGS: Efficient and High-Fidelity Gaussian Splatting

Changbai Li, Shuo Yang, Yichen Yang, Shuwei Shao+1

Beihang University · Nanyang Technological University

3D Gaussian Splatting 3D Gaussian Pruning 3DGS Acceleration Adaptive Gaussian Densification Novel View Synthesis 3DGS Training

Open paper AskPDF arXiv

3D Gaussian Splatting (3DGS) enables real-time novel view synthesis, but existing general-purpose acceleration methods suffer severe rendering quality degradation when extended to more complex, large-scale scenes. To address this issue, we propose EffGS, a more general acceleration framework that improves training and rendering efficiency while maintaining reconstruction quality comparable to or better than vanilla 3DGS across bounded and large-scale scenes. EffGS combines frequency-aware guidance, localized density control, and adaptive primitive scale modulation. First, an importance scoring mechanism combines pixel-wise reconstruction errors with a difference-of-Gaussians mask scheduled over training to provide stage-dependent spatial guidance. Second, localized densification and pruning restricts density modifications to Gaussians with valid projected footprints in the sampled views. Third, learnable per-Gaussian scale modulation adjusts effective primitive extent during optimization while retaining the Compact Box rasterization rule. Extensive experiments on bounded and large-scale scene datasets demonstrate a favorable balance between reconstruction quality, training time, and primitive count. Component ablations and matched-primitive-budget comparisons further support the effectiveness of the framework.

Abstract

3D Gaussian Splatting

Explore similar work

EffGS: Efficient and High-Fidelity Gaussian Splatting

TideGS: Scalable Training of Over One Billion 3D Gaussian Splatting Primitives via Out-of-Core Optimization

Matryoshka Gaussian Splatting