PaperlayerView exact revision on arXiv

Source-linked AI summary

3DGS.zip: A survey on 3D Gaussian Splatting Compression Methods

Milena T. Bagdasarian, Paul Knoll, Yi-Hsin Li, Florian Barthel, Anna Hilsmann, Peter Eisert, Wieland Morgenstern

arXiv:2407.09510v5cs.CV

TL;DR

3DGS's storage and memory demands constrain its practicality, motivating compression and compaction methods. This survey reviews and compares those approaches; it reports that leading methods reduce file size by more than 40-fold while enhancing visual quality, and notes that standardized benchmarks remain needed.

  • Problem

    3DGS's storage and computational costs constrain its practicality, while varied datasets, metrics, and experimental conditions make compression methods difficult to compare directly.

    Source: §7 · p. 18, §6 · p. 18

  • Method

    The survey reviews 3DGS compression and compaction methods, compares their approaches and results, and proposes basic comparison guidelines.

    Source: §1 · p. 2, §6 · p. 18

  • Results

    More than 40-fold file-size reduction with enhanced visual quality is reported for the most effective methods on Tanks and Temples scenes.

    Source: §6 · p. 18

  • Takeaways & Limitations

    A well-implemented adaptive density-control strategy can improve most methods, while the best compression choice varies across datasets and quality metrics.

    Source: §4.1.3 · p. 12, §6 · p. 18

    Different datasets, metrics, and experimental conditions impede direct comparison, and standardized datasets and evaluation protocols remain needed.

    Source: §6 · p. 18

Abstract

from arXiv · show

3D Gaussian Splatting (3DGS) has emerged as a cutting-edge technique for real-time radiance field rendering, offering state-of-the-art performance in terms of both quality and speed. 3DGS models a scene as a collection of three-dimensional Gaussians, with additional attributes optimized to conform to the scene's geometric and visual properties. Despite its advantages in rendering speed and image fidelity, 3DGS is limited by its significant storage and memory demands. These high demands make 3DGS impractical for mobile devices or headsets, reducing its applicability in important areas of computer graphics. To address these challenges and advance the practicality of 3DGS, this survey provides a comprehensive and detailed examination of compression and compaction techniques developed to make 3DGS more efficient. We classify existing methods into two categories: compression, which focuses on reducing file size, and compaction, which aims to minimize the number of Gaussians. Both methods aim to maintain or improve quality, each by minimizing its respective attribute: file size for compression and Gaussian count for compaction. We introduce the basic mathematical concepts underlying the analyzed methods, as well as key implementation details and design choices. Our report thoroughly discusses similarities and differences among the methods, as well as their respective advantages and disadvantages. We establish a consistent framework for comparing the surveyed methods based on key performance metrics and datasets. Specifically, since these methods have been developed in parallel and over a short period of time, currently, no comprehensive comparison exists. This survey, for the first time, presents a unified framework to evaluate 3DGS compression techniques. We maintain a website that will be regularly updated with emerging methods: https://w-m.github.io/3dgs-compression-survey/ .

1 Introduction

3DGS offers an explicit, editable alternative to NeRF, but its large files and memory demands motivate methods that reduce representation size while preserving quality. This survey reviews compression and compaction approaches and compares them within a defined scope.

1 Introduction

  • NeRF's implicit representation requires billions of network evaluations per image and complicates direct scene editing, despite speed advances such as Instant-NGP.3DGS is introduced as an explicit alternative with quality close to Zip-NeRF.Source: §1 · p. 2
  • 3DGS represents scenes with editable Gaussian ellipsoids and supports real-time rendering, but full-precision attributes can produce files of several gigabytes.Compression can reduce transmission and memory demands and support rendering on lower-end devices.Source: §1 · p. 2
  • The survey reviews and evaluates techniques for compacting 3DGS, including classical methods adapted to Gaussian splatting and newly developed approaches.It aims to help researchers identify effective strategies and areas for further exploration.Source: §1 · p. 2
  • The survey identifies a lack of comprehensive comparison among parallel methods and argues that a consistent standard could improve compatibility and usability.It notes that the community is closing the compression-efficiency gap between 3DGS and NeRFs.Source: §1 · p. 2

Scope of this Survey

  • The report focuses on optimizing 3DGS size or memory footprint, compares common datasets and pipelines, and includes conference and journal papers as well as arXiv preprints.It does not cover all 3DGS methods and applications; NeRF-specific compression is referred to [28].Source: Scope of this Survey · pp. 2–3

2 Fundamentals of 3D Gaussian Splatting (3DGS)

3DGS models a scene with Gaussian primitives whose attributes are optimized to match scene geometry and appearance. Rendering projects them into image-plane splats and blends their contributions by depth; attribute distributions can also suggest compression or pruning opportunities.

Fundamentals of 3D Gaussian Splatting (3DGS)

  • Each 3D Gaussian has 59 attributes: position, covariance, opacity, and spherical-harmonic color coefficients.Position uses three values; covariance uses scale and rotation, opacity one value, and color 48 coefficients.Source: §2 · p. 3
  • Rendering projects each 3D Gaussian onto the image plane, then alpha-blends overlapping splats in depth order to determine each pixel's color.A splat's contribution is the product of its opacity and Gaussian falloff.Source: §2 · p. 3
  • The bicycle-scene histograms show many very-low-opacity Gaussians, potential pruning candidates, and few large Gaussians after adaptive density control.The figure plots opacity and the first scaling attribute for the scene's Gaussians.Source: §2 · p. 3, Figure 1 · p. 3

3 Fundamentals of Compression and Compaction for 3DGS

This section surveys coding and spatial-organization concepts for reducing 3DGS storage while retaining useful scene information. It describes how methods exploit redundancy and spatial structure, then weighs their implementation and memory trade-offs.

3 Fundamentals of Compression and Compaction for 3DGS; 3.1 Vector Quantization

  • Lossless coding reduces redundancy, while lossy compression removes imperceptible information; vector quantization replaces attribute vectors with indices to shared codebook representatives.Quantization can operate on full multidimensional vectors or separate dimensions, allowing different attributes to receive different precision.Source: §3 · p. 4, §3.1 · p. 4

3.2 Structuring and Dimensionality Reduction; 3.2.1 Octrees

  • Octrees recursively partition space and allocate memory to occupied regions, while structural methods exploit spatial relationships rather than encoding attributes independently.This organization can skip empty space and use coordinates relative to sub-cubes, reducing average bit-length.Source: §3.2.1 · p. 5, §3.2 · p. 5

3.2.2 Anchor-based Representations; 3.2.3 Multi-Resolution Hash Grids

  • Anchor-based representations derive Gaussian attributes from representative points, and multiresolution hash grids decode stored feature vectors into individual primitive attributes.Hash-grid table size, feature-vector size, and level count trade off memory, performance, and quality; an MLP implicitly handles collisions.Source: §3.2.2 · p. 5, §3.2.3 · p. 6

3.2.4 Z-order Curves

  • Z-order curves interleave coordinate bits to order splats by position, enabling spatially coherent run-length or predictive encoding without allocating storage to empty regions.Their linear mapping does not always keep high-dimensional neighbors close; Hilbert curves can sometimes provide better locality.Source: §3.2.4 · p. 6

3.2.5 Tri-planes and K-planes

  • Tri-plane factorization projects 3D positions onto three 2D planes, storing features that are combined and decoded by an MLP into Gaussian attributes.The planes provide explicit storage while the decoder is an implicit component; K-Planes extends the factorization with extra planes for additional dimensions [15].Source: §3.2.5 · pp. 6–7

3.2.6 Self-Organizing Gaussians

  • Self-Organizing Gaussians learn a scene-specific 2D layout that preserves local relationships and supports standard image compression techniques.Unlike fixed Z-order curves and tri-planes, the learned organization can model neighbors across dimensions.Source: §3.2.6 · p. 7, Figure 8 · p. 7

3.2.7 Region Adaptive Hierarchical Transform (RAHT)

  • RAHT hierarchically transforms color attributes along an octree, propagating low-frequency coefficients and encoding high-frequency coefficients at each level.Weight-adapted pairwise operations preserve small-scale detail with lower complexity than graph-based transforms; the coefficients are then quantized and entropy coded.Source: §3.2.7 · p. 7

3.2.8 Discussion

  • Representation choices trade off spatial coherence, storage, and computation: feature-based methods require MLP decoding, while explicit methods expose Gaussian attributes directly.Explicit-method VRAM tracks primitive count, whereas feature planes, hash tables, and codebooks have memory use that does not further decrease when primitive count is reduced.Source: §3.2.8 · p. 8

3.3 Attribute Pruning

Attribute pruning reduces per-Gaussian storage by tailoring or removing spherical-harmonic coefficients, while compaction dynamically adjusts Gaussian density to balance scene detail and model size.

3.3 Attribute Pruning

  • SH coefficients can account for 48 of 59 Gaussian attributes, so reducing them targets a major storage and processing cost.Degree-3 SH for RGB uses 48 coefficients, or 81% of the 59 attributes.Source: §3.3 · p. 8
  • Per-view transmittance guides SH reduction or removal, while NeRF-like MLPs or classical shading may replace SH with custom, potentially slower rendering.Removing SH bands also removes higher-frequency color components, sacrificing fine view-dependent lighting and shading detail.Source: §3.3 · p. 8

3.4 Compaction

  • Adaptive Density Control clones, splits, or prunes Gaussians using criteria such as gradients, pixel coverage, and saliency.Splitting replaces a Gaussian with a default of two children within the parent’s bounds; the children inherit its attributes and reduced scaling.Source: §3.4 · pp. 8–9, Figure 9 · p. 9
  • Opacity pruning shows minimal visual loss from 5.6 to 4.1 million Gaussians, but 3.0 million reveals spoke changes and 1.0 million damages bicycle and grass reconstruction.ADC operates during training and rendering to refine Gaussian allocation as the scene evolves.Source: §3.4 · p. 9

4 Efficient Strategies for Compression and Compaction in 3D Gaussian Splatting

The section distinguishes compression, which reduces storage, from compaction, which reduces the number and distribution of Gaussians. Both strategies aim to retain essential scene information while improving 3DGS efficiency.

4 Efficient Strategies for Compression and Compaction in 3D Gaussian Splatting

  • Compression reduces storage through methods such as vector quantization and structured representations, while compaction optimizes Gaussian number and distribution.Structured forms include grids or anchor points.Source: §4 · p. 9

4.1 Compression

Compression methods trade file size against visual quality, with different approaches favored by different priorities. The section surveys attribute and structural representations before comparing methods across datasets and metrics.

4.1.1 Efficient Representation of Gaussian Attributes

  • Vector quantization groups similar Gaussian attributes into codebooks, while adaptive SH reduction or removal saves storage at the cost of fine view-dependent color detail.LightGaussian uses significance to avoid quantizing more significant SH; removing higher SH bands also sacrifices subtle lighting and shading variation.Source: §4.1.1 · pp. 9–10

4.1.2 Structured Representation of Gaussians

  • Several compression methods structure unordered Gaussians to exploit correlations between neighboring Gaussians, using anchor-based representations, hash grids, structured 2D grids, feature planes, or sorting by quantized indices.Anchor-based methods predict Gaussian attributes from anchor features and viewing conditions; 2D grids allow standard image compression, while CodecGS uses standard video coding for feature-plane compression.Source: §4.1.2 · p. 11

4.1.3 Quantitative Comparison of Compression Methods

  • ContextGS-lowrate ranks first overall, but dataset and metric variation means no compression strategy wins universally.CodecGS reports 7.8 MB on Tanks and Temples with PSNR and SSIM near the top tier; HAC-lowrate, SOG without SH, and ContextGS-lowrate are under 10 MB there with acceptable visual quality.Source: §4.1.3 · p. 12
  • For LPIPS, gsplat, ContextGS-highrate, and HAC-highrate are strong choices, while original 3DGS-30K has the best values on two datasets at up to 70 times the size.The two datasets are Mip-NeRF 360 and Deep Blending.Source: §4.1.3 · p. 12
  • ContextGS, HAC, CodecGS, and gsplat balance PSNR, SSIM, LPIPS, and size, with quality or storage priorities shifting the preferred method.Context modeling and hierarchical approaches tend to lead on image quality, while CodecGS and low-rate variants are robust for compression.Source: §4.1.3 · p. 12

4.2 Compaction

Compaction changes Gaussian distribution through densification and pruning, often reducing memory needs when similar quality can be achieved with fewer primitives. The comparison shows that the most compact choice depends on scene characteristics and quality priorities.

4.2 Compaction

  • Compaction divides into densification, which adds Gaussians for fidelity, and pruning, which removes low-contribution primitives to avoid over-reconstruction and inefficiency.Some methods improve quality without adding primitives, implicitly reducing memory needs at similar quality; incomplete Gaussian-count evaluations exclude some methods from the comparison.Source: §4.2 · p. 12

4.2.1 Densification

  • Densification methods use cues including color, frequency, pixel-aware gradients, depth, normals, and global scores to choose where to add or split Gaussians.MCMC instead samples a fixed number from a learned distribution, adds positional noise, and respawns low-opacity Gaussians near high-opacity ones.Source: §4.2.1 · pp. 12, 14, 15

4.2.2 Pruning

  • Pruning methods remove Gaussians judged redundant by masks, importance scores, overlap, or spatial graphs; SUNDAE uses a CNN to recover fine detail after graph-based down-sampling.Source: §4.2.2 · p. 15

4.2.3 Quantitative Comparison of Compaction Methods

  • Compaction rankings average PSNR, SSIM, LPIPS, and Gaussian-count ranks across available datasets, giving each quality metric one-sixth and count one-half of the total.The comparison covers Tanks and Temples, Mip-NeRF 360, and Deep Blending.Source: §4.2.3 · p. 15
  • Octree-GS and Mini-Splatting use the fewest Gaussians while retaining competitive PSNR; Taming3DGS (Big) and GaussianPro favor higher quality with more Gaussians.Octree-GS does not lead on every dataset: it excels on sparse texture-rich regions, whereas Taming3DGS (Big) performs better in texture-dense scenes.Source: §4.2.3 · p. 15

5 Datasets and Comparison Metrics

The survey evaluates compression and compaction methods across real-world and synthetic datasets, using image-quality and model-size metrics. It also standardizes testing protocols and shows that training and evaluation resolution affect PSNR.

5.1 Datasets

  • The evaluation uses Tanks and Temples, Mip-NeRF 360, and Deep Blending real-world scenes plus Synthetic NeRF scenes, aligned with the original 3DGS benchmarks.The selected scenes include outdoor and indoor environments, alongside detailed synthetic scenes.Source: §5.1 · pp. 15–16

5.2 Comparison Metrics

  • The survey assesses rendered quality with PSNR, SSIM, and LPIPS, and representation efficiency with model file size or Gaussian count.Higher PSNR and SSIM indicate better similarity; lower LPIPS and smaller model size indicate better perceptual quality and efficiency, respectively.Source: §5.2 · p. 16

5.3 Testing Protocols

  • To improve comparability, the survey solicited standardized results from method authors and recommended established 3DGS scene, image-sizing, and test-split conventions.The protocol specifies the nine Mip-NeRF 360 scenes, two Tanks and Temples scenes, a 1600-pixel maximum side, and dataset-specific test splits.Source: §5.3 · p. 16
  • Matching evaluation resolution to training resolution yields the best PSNR; higher training and evaluation resolutions also improve PSNR, and 1 million Gaussians outperform 360k in overall quality.Table 3 compares training and evaluation sizes for Mip-NeRF 360 scenes.Source: §5.3 · pp. 16, 18, Table 3 · p. 18

6 Discussion

The discussion reports substantial compression gains and argues that compaction can complement many methods. It highlights unresolved trade-offs in quality, efficiency, and evaluation.

6 Discussion

  • The strongest compression methods reduce the original 3DGS model size by more than 40-fold while improving visual quality.The original implementation on Tanks and Temples scenes exceeded 400 MB.Source: §6 · p. 18
  • A well-implemented compaction strategy can improve most methods, while combining vector quantization with structured representations remains an open research direction.The discussion names octrees and hash grids as spatial structures that could pair with vector quantization.Source: §6 · p. 18
  • Compression choices trade file size against fidelity and compute: aggressive pruning or quantization can create artifacts, while neural feature encoding reduces storage but adds rendering computation.Explicit representations use more storage but allow real-time rendering with minimal overhead.Source: §6 · p. 18
  • Benchmarking remains difficult because studies use different datasets, metrics, and experimental conditions; the survey proposes comparison guidelines but calls for standardized protocols.It also recommends evaluating real-time performance and energy efficiency, especially for mobile and embedded deployment.Source: §6 · p. 18
  • HAC-highrate favors quality with greater computational overhead, whereas SOG w/o SH reduces memory substantially but may degrade quality in fine-detail scenes.The discussion frames flexible, application-tuned solutions as a need rather than a settled approach.Source: §6 · p. 18

7 Conclusion and Future Directions

The conclusion presents compression and compaction as ways to address 3DGS storage and computational costs, while identifying adaptability and standardization as continuing needs. It points to dynamic compression and multiresolution rendering as future directions.

7 Conclusion and Future Directions

  • 3DGS offers explicit scene control and high rendering fidelity, but storage and computational costs constrain its practicality and platform adaptability.The conclusion calls for continued refinement and standardization of compression and compaction techniques.Source: §7 · p. 18
  • Most current 3DGS compression techniques target static scenes, while real-world applications increasingly require dynamic and interactive capabilities.The conclusion identifies pretrained-model dependence as a constraint on adaptability.Source: §7 · p. 18
  • Future work could explore multiresolution models with level-of-detail scaling, though seamless transitions without artifacts remain a challenge.Octrees or feature grids are identified as scalable options; quantization-aware training and shared codebooks may further improve compression efficiency.Source: §7 · pp. 18–19

8 Short Summaries of Included Compression Approaches

The survey catalogs methods that compress or compact 3DGS through context modeling, quantization, pruning, structured representations, and learned features. The approaches differ in how they balance storage, rendering, and visual quality.

8.1 ContextGS; 8.2 HAC

  • ContextGS predicts uncoded anchors from coarser coded levels and adds quantized hyperprior features; HAC predicts anchor distributions from hash-grid features and adaptively quantizes and masks anchors and offsets.ContextGS applies to Scaffold-GS and vanilla 3DGS; HAC is built on Scaffold-GS.Source: §8.1 · p. 19, §8.2 · p. 19

8.3 CodecGS; 8.4 gsplat

  • CodecGS trains progressive tri-planes after standard 3DGS training and encodes transformed feature planes with a standard video codec; gsplat combines 3DGS-MCMC with grid-organized parameters and clustered spherical harmonics.CodecGS uses two-phase training to address sparse-signal instability; gsplat stores clustered spherical-harmonic values as FP16.Source: §8.3 · p. 19, §8.4 · p. 19

8.5 Compact3D; 8.6 CompGS

  • Compact3D uses K-means vector quantization, compact codebooks, and run-length-style index coding, while CompGS predicts coupled-primitive attributes from anchors and jointly optimizes distortion and bit rate.Compact3D also regularizes opacity before pruning low-opacity Gaussians.Source: §8.5 · p. 19, §8.6 · p. 19

8.7 RDO-Gaussian

  • RDO-Gaussian jointly optimizes rate and distortion using dynamic Gaussian pruning, adaptive spherical-harmonic pruning, and entropy-constrained vector quantization.It discretizes covariance and color attributes with vector quantization.Source: §8.7 · pp. 19–20

8.8 Reduced3DGS

  • Reduced3DGS combines scale- and resolution-aware pruning, adaptive spherical-harmonic bands, K-means codebooks, and half-float quantization to reduce storage.Its SH-band adjustment uses cross-view color consistency to assess view-dependent effects.Source: §8.8 · p. 20

8.9 SOG; 8.10 MesonGS

  • SOG sorts Gaussians into a locally smooth 2D grid and trains with a smoothness loss; MesonGS combines pruning, attribute transforms, quantization, octree geometry coding, LZ77 packing, and quality-restoring fine-tuning.SOG's training-time smoothness is intended to improve compressibility.Source: §8.9 · p. 20, §8.10 · p. 20

8.11 Compressed3D; 8.12 Compact3DGS

  • Compressed3D combines sensitivity-aware codebooks, quantization-aware fine-tuning, and entropy and run-length coding along a space-filling curve; Compact3DGS uses volume masks, grid-based color fields, geometry codebooks, and post-processing quantization and entropy coding.Compressed3D also proposes GPU-based sorting and rasterization for real-time novel-view synthesis on low-end devices.Source: §8.11 · p. 20, §8.12 · p. 20

8.13 EAGLES; 8.14 Scaffold-GS; 8.15 LightGaussian

  • EAGLES quantizes color and rotation attributes and prunes low-influence Gaussians; Scaffold-GS predicts view-adaptive Gaussian attributes from structured anchors, while LightGaussian prunes and recovers Gaussians, distills spherical harmonics, and quantizes redundant attributes.EAGLES identifies color and rotation as over 80% of memory use; Scaffold-GS initializes anchors from voxelized SfM points.Source: §8.13 · p. 20, §8.14 · p. 20, §8.15 · pp. 20–21

9 Short Summaries of Included Compaction Approaches

The surveyed compaction methods refine where Gaussians are added or removed to better capture scene details while reducing redundant representation. They use different strategies, including depth guidance, level of detail, scoring, geometry-aware optimization, and color cues.

9.1 Mini-Splatting: Representing Scenes with a Constrained Number of Gaussians (Mini-Splatting)

  • Mini-Splatting [13] refines blurred regions and repositions Gaussians using depth points, then simplifies the scene while preserving visual impact, geometry, and rendering quality.Its simplification uses Intersection Preserving and Sampling.Source: §9.1 · p. 21

9.2 Octree-GS: Towards Consistent Real-time Rendering with LOD-Structured 3D Gaussians (Octree-GS)

  • Octree-GS [45] organizes anchor Gaussians by level of detail and selects levels according to the observation view for rendering.The model uses standard image reconstruction and volume regularization losses.Source: §9.2 · p. 21

9.3 Taming 3DGS: High-Quality Radiance Fields with Limited Resources (Taming3DGS)

  • Taming 3DGS scores Gaussians using gradient, pixel coverage, per-view saliency, and core attributes, then splits or clones the top B to densify the scene.The composite score targets areas of structural complexity and visual importance.Source: §9.3 · p. 21

9.4 AtomGS: Atomizing Gaussian Splatting for High-Fidelity Radiance Field (AtomGS)

  • AtomGS [31] uses uniformly sized isotropic Atom Gaussians for fine detail, merges large Gaussians over smooth surfaces, and applies geometry-guided optimization.Its Edge-Aware Normal Loss combines a normal map from pre-optimized 3DGS with an edge map from ground-truth RGB gradients; it also uses multi-scale SSIM.Source: §9.4 · p. 21

9.5 Color-cued Efficient Densification Method for 3D Gaussian Splatting (Color-cued GS)

  • Color-cued GS uses view-independent 0th spherical-harmonics coefficient gradients for color cues and coarser 2D position gradients to refine structures SfM misses.Source: §9.5 · p. 21

9.6 GaussianPro: 3D Gaussian Splatting with Progressive Propagation (GaussianPro)

  • GaussianPro [10] propagates depth and normal information from neighboring pixels to initialize Gaussians where geometric filtering identifies a need for additions.A planar loss and L1 and angular losses encourage surface alignment and consistency between rendered and propagated normals.Source: §9.6 · p. 21

B Additional figures for compression and compaction.

The additional figures show attribute distributions and correlations for the bicycle scene, alongside quality-versus-size or quality-versus-Gaussian-count plots for compression methods across four datasets.

B Additional figures for compression and compaction.

  • Figure 14 plots stored, unactivated splat attributes for the bicycle scene, showing only the first 12 of 48 spherical-harmonics attributes.Source: Figure 14 · p. 24
  • Figure 15 shows correlations among bicycle-scene attributes, including cross-block spherical-harmonics correlations and the space occupied by third-degree SH attributes.The caption notes that SH correlations diminish at higher degrees.Source: Figure 15 · p. 25
  • Figures 16–20 compare method quality against model size or Gaussian count across Tanks and Temples, Mip-NeRF 360, Deep Blending, and Synthetic NeRF.Figures 16–17 plot SSIM or LPIPS against model size; Figures 18–20 plot PSNR, SSIM, or LPIPS against Gaussian count.Source: Figure 16 · p. 26, Figure 17 · p. 26, Figure 18 · p. 27

Related papers

Loading 2407.09510v5…