PaperlayerView exact revision on arXiv

Source-linked AI summary

KISS-GS: 3D Gaussian Splatting Compression Kept Simple

Wieland Morgenstern, Friedrich Elias Branschke, Florian Fleischmann, Adrian Szatmari, Paul Schlack, Florian Barthel, Peter Eisert, Anna Hilsmann

arXiv:2608.26948v1cs.CV

TL;DR

Uncompressed 3DGS scenes are difficult to store and deploy, and combined compression systems obscure the contribution of individual stages. KISS-GS decouples compaction, image-based encoding, and optional adaptation into modular stages with simple web-compatible decoding. Across benchmarks, it reports substantially higher reduction factors than HAC++ on Tanks and Temples and Mip-NeRF 360 at the INRIA quality reference.

  • Problem

    Large raw 3DGS files hinder storage, streaming, and deployment, while combining compression strategies obscures gain attribution and limits component reuse across training pipelines.

  • Method

    KISS-GS separates compaction, encoding, and adaptation into independently replaceable stages operating on standard 3DGS files, using SOG-XT for browser-compatible image-based encoding.

  • Results

    KISS-GS reaches the INRIA quality reference at substantially higher reduction factors than HAC++ on Tanks and Temples and Mip-NeRF 360.

  • Takeaways & Limitations

    The modular pipeline makes each stage's rate-distortion contribution measurable and permits reconstruction improvements to be adopted without changing compression.

  • Takeaways & Limitations

    The authors acknowledge that KISS-GS falls short of simplicity because its encoder is a sophisticated modular pipeline containing POPSpa and PRAS.

Abstract

from arXiv · show

Scene reconstruction with 3D Gaussian Splatting (3DGS) has become common, however deployment remains painful as the uncompressed file sizes can be massive. Current 3DGS compression systems combine multiple strategies for file size reduction, which can obscure where gains come from and limit component reuse across training pipelines. To make the gains more transparent, we propose KISS-GS, a modular compression pipeline named after the principle of keeping things simple, designed to decouple compression entirely from training. Given a 3DGS scene reconstructed with vanilla 3DGS, we are able to reduce it through compaction by 15.7x using a combination of state-of-the-art pruning schemes. Then we encode it into an image-based format designed for simple, ubiquitous decoding. With the SOG-XT format, we propose a novel extension to Self-Organizing Gaussians with two main contributions: (i) Self-organizing 2D Codebooks and (ii) Parallel Representative Assignment Smoothing (PRAS), which leverages the symmetry of quaternion and scale parameterizations to produce 2D attribute grids more amenable to encoding. This encoding reduces scene size by 6.6x. We show that optional encoding-aware fine-tuning yields a further 2.2x. Across standard 3DGS benchmarks, our simple and modular approach thus achieves a total of 85x to 319x reductions in the size of the scene over uncompressed vanilla 3DGS, setting new benchmarks for real-world scenes and surpassing tightly integrated methods in rate-distortion. Decoding relies solely on web-native image formats, and the modular design makes each stage easy to combine with future advances in reconstruction and compaction. Code and project page: https://fraunhoferhhi.github.io/KISS-GS/

1 Introduction

KISS-GS addresses large 3DGS files and unclear attribution by separating compaction, encoding, and adaptation into modular, independently analyzable stages. It targets simple, browser-compatible decoding while evaluating whether this modular approach can match integrated compression methods.

  • 3DGS .ply files are large and inefficient for storage, streaming, and deployment, creating obstacles for web, mobile, and AR/VR applications.
  • Combining compaction, encoding, and adaptation can obscure where rate-distortion gains originate and hinder fair comparison and component reuse.
  • KISS-GS restricts decoding to a deliberately simple pipeline backed by widely supported image codecs.
  • The paper asks whether modularization closes the attribution gap, simple image-based encoding matches integrated methods, and codec-aware fine-tuning remains beneficial.
  • KISS-GS separates compaction, encoding, and adaptation into distinct stages that can be replaced, ablated, and analyzed independently.
  • SOG-XT extends Self-Organizing Gaussians with Self-organizing 2D codebooks and PRAS while keeping the decoder simple.

2 Related work

Prior 3DGS storage-reduction work is organized around compaction, encoding, and adaptation, alongside methods that integrate several strategies. Image-based layouts offer simple decoding through standard codecs, while integrated approaches combine more specialized representations and losses.

  • Compaction: Compaction reduces the number of active Gaussians, directly affecting storage footprint and rendering cost while aiming to preserve reconstruction quality.
  • Compaction: Compaction methods use score-based sampling, reconstruction error, rendering contribution, sampling strategies, or sparsity-constrained optimization to select primitives.
  • Encoding: Encoding reduces bits per primitive through quantization and entropy coding, including vector quantization for view-dependent spherical harmonics.
  • Encoding: Structured encodings replace explicit attributes with anchors, codebooks, MLPs, octrees, triplanes, or hash grids to exploit spatial correlations.
  • Image-based layouts: Morton curves and Self-Organizing Gaussians arrange primitives into locally smooth 2D attribute layouts that standard image codecs can decode with hardware acceleration.
  • Adaptation: Adaptation steers scene parameters toward encoding-friendly configurations using quantization, rate, entropy, masking, hash-grid, or rendering-fidelity losses.

3 Modules of the KISS-GS pipeline

KISS-GS is a modular post-training pipeline that accepts standard 3DGS scenes, compacts primitives with POPSpa, encodes them using SOG-XT, and can optionally fine-tune with the codec in the loop. SOG-XT improves grid compressibility through structured layouts, codebooks, and covariance-parameter symmetry resolution.

  • 3.1 3D Reconstruction: KISS-GS accepts any standard 3DGS .ply file and preserves the common means, quaternion-scale, spherical-harmonic, and opacity representation.
  • 3.1 3D Reconstruction: Decoupling compression from reconstruction avoids training-format coupling and allows future reconstruction improvements without changing the compression pipeline.
  • 3.2 Compaction with POPSpa: POPSpa combines score-based GaussianPOP pruning, GaussianSpa optimize-sparsify refinement, and effective-rank regularization as a replaceable compaction module.
  • 3.2 Compaction with POPSpa: Effective-rank regularization discourages needle-shaped Gaussians and favors disk-shaped Gaussians better suited for surface reconstruction.
  • 3.2 Compaction with POPSpa: POPSpa estimates per-primitive rendering error, prunes low-scoring primitives, refines the survivors, then performs a second pruning pass and fine-tuning.
  • 3.3 Encoding with SOG-XT: SOG-XT applies PLAS-based 2D layouts to attributes and replaces raw spherical-harmonic slices with a k-means codebook and per-primitive cluster indices.
  • 3.4 Covariance Symmetry and PRAS: PRAS searches valid quaternion-scale representatives across 48 equivalent parameter combinations to smooth grids while preserving covariances and grid coordinates.

4 Evaluation

The evaluation measures KISS-GS’s modular compaction, encoding, and optional adaptation under standard 3DGS benchmarks and compares quality–size trade-offs with established methods. Results show strong compaction, single-digit-megabyte encoding, improved rate-distortion with fine-tuning, and leading reductions relative to INRIA 3DGS.

  • Compaction: POPSpa consistently outperforms competing post-training compaction methods across all four benchmarks.The compaction comparison also includes compaction-aware training methods under extended iteration budgets.
  • Compaction: 1.7 average rank is achieved by POPSpa while using significantly fewer Gaussians than competing methods across most datasets.In Deep Blending, competitive quality is reached with only 56% of the primitives of the best-performing method.
  • Encoding: SOG-XT compresses test scenes to single-digit Megabyte sizes across a broad range of primitive counts and also works on the INRIA 40k baseline.This supports the modularity of applying the same encoding method to different compacted or baseline scenes.
  • Encoding: 2D codebooks and UV label packing provide size reductions without quality loss, while PRAS trades a small size increase for large quality gains.Removing SH AC codebooks causes the largest quality loss, whereas omitting them would nearly triple file size.
  • Adaptation: 2.2× further reduction at INRIA quality is obtained on Mip-NeRF 360 through encoding-aware fine-tuning over post-training compression.Fine-tuning is optional and is most useful when a GPU and training views are available during encoding.
  • Comparison with state of the art: KISS-GS achieves the highest file-size reductions on Tanks and Temples and Mip-NeRF 360, plus two of three Synthetic NeRF metrics, at INRIA-Q.Table 3 reports interpolated reduction factors at the quality of the INRIA 3DGS 40k reference.

5 Have we kept it simple?

KISS-GS keeps deployment simple by separating compression stages, but its encoder is deliberately sophisticated to make each gain measurable. It reaches substantial compression while preserving a simple decoding path.

  • Simplicity and modularity: KISS-GS separates reconstruction, compaction, encoding, and encoding-aware fine-tuning so each stage can be measured independently.The authors acknowledge that this modularity requires a sophisticated encoder, including POPSpa and PRAS.
  • Compression payoff: 319× size reduction on Tanks and Temples and 228× on Mip-NeRF 360 establish record-setting compression on two real-world datasets.The reported reductions reach single-digit megabytes at vanilla 3DGS reconstruction quality.
  • Deployment simplicity: SOG-XT uses standard image decoding with deterministic rescaling and reshaping operations to reconstruct standard 3DGS attributes.This keeps decoders easy to port across languages and platforms.
  • Scope boundary: Deep Blending remains the main boundary because anchor-based learned representations better absorb capture inconsistencies than standard-splat reconstructions.KISS-GS takes standard-splat reconstructions as input, limiting its scope on this dataset.

6 Appendix to KISS-GS: 3D Gaussian Splatting Compression Kept Simple

The appendix examines runtime, decoder behavior, fine-tuning sensitivity, and compaction ablations. It finds that SOG-XT encoding is inexpensive relative to earlier stages, while larger models shift cost toward SH codebook clustering.

  • Runtime Performance: Encoding with SOG-XT is inexpensive compared with reconstruction and compaction, whereas codec-in-the-loop fine-tuning is more costly.On Bicycle at 512k Gaussians, reconstruction takes 1215 s, POPSpa 277 s, SOG-XT encoding 34 s, and fine-tuning 441 s.
  • Runtime Performance: Post-hoc and fine-tuned encoding decode identically; fine-tuning trades longer encoding time for better compression without increasing deployment complexity.The codec is evaluated in the training loop only for the adaptation stage.
  • Runtime Performance: SH codebook clustering dominates encoding cost for larger models, making it the clearest target for future encoder optimization.Components have similar cost for small models.
  • Encoding-aware Fine-tuning Sensitivity: The fine-tuning sweep shows a stable rate-distortion region rather than a fragile optimum on Bicycle.Points are means over seven runs with one-standard-deviation error bars.
  • Encoding-aware Fine-tuning Sensitivity: Small PSNR gains from disabling TV regularization or increasing quaternion precision come with larger files, while stronger TV and extra steps do not improve the trade-off.The SH codebook size is the clearest optional rate-control knob, but its best setting is scene dependent; K = N/8 is used by default.
  • Compaction Ablation: POPSpa ablations generally improve results on two of three real-world datasets, while synthetic data stagnates.Adding Effective Rank Regularization prepares compacted data for compression without quality loss and reduces Gaussian spikiness.

6.5 Covariance Symmetries

The appendix explains covariance parameterization symmetries and evaluates protocol and compute effects in HAC++ comparisons. It also specifies the held-out, multi-dataset evaluation used to assess compressed scene quality.

  • Covariance Symmetries: 48 equivalent parameterizations represent the same Gaussian covariance: six scale permutations combined with four valid eigenvector orientations.These symmetries arise from the decomposition Σ = R S S^T R^T with diagonal scales and rotation matrices.
  • HAC++ Results: HAC++ comparisons report original 30k results, recomputed matched-resolution runs, and full-protocol runs using 1600 px test resolution.The 44k comparison matches KISS-GS compute by adding 10k compaction and 4k fine-tuning steps.
  • HAC++ Results: Scene size decreases slightly at 40k iterations and then stabilizes, while additional steps do not improve rendering quality.The full evaluation protocol lowers quality metrics slightly while leaving file size nearly unchanged.
  • HAC++ Results: At high λ, extended training forces further compression without recovering quality; at low λ, quality remains stable across iteration counts.Here λ controls the rate-distortion trade-off in the HAC++ loss.
  • Evaluation Protocol: The evaluation covers 21 scenes across four datasets and reports PSNR, SSIM, and LPIPS on held-out validation or designated test splits.Training and evaluation resolutions, backgrounds, and LPIPS preprocessing vary according to dataset and protocol.
  • Evaluation Protocol: Held-out-image evaluation measures novel-view generalization, testing whether bitrate reduction preserves scene structure for unseen viewpoints.Mip-NeRF 360 is especially informative because it combines diverse real-world scenes with differing training and evaluation resolutions.

6.8 Compression Ablation

SOG-XT combines quantization, PRAS, grid handling, and image-based decoding to reduce storage while preserving reconstruction quality. Its reference decoder reverses these transforms and writes a standard .ply scene.

  • Quantization and PRAS: Uniform min–max quantization stores attributes at 100, 256, or 65,536 discrete levels while retaining extrema for dequantization.The corresponding table labels are q100_u8, u8, and u16.
  • Quantization and PRAS: 100-level quaternion quantization is enabled by PRAS; without PRAS, recovering quality requires 256-level u8 quantization.The ablation compares q100_u8 with the “w/o PRAS, quaternion full u8 quant” setting.
  • Grid construction: SOG-XT pads square grids with dummy primitives and applies an active mask at decoding instead of pruning valuable post-POPSpa Gaussians.Opacity-based pruning gives only a tiny size benefit but causes a clear quality loss.
  • Decoding pipeline: The decoder extracts quantization ranges from metadata, dequantizes stored attributes, and untile-reconstructs the 45-channel spherical-harmonics centroid table.The helper functions support metadata parsing, inverse quantization, and 3×5 feature-grid unpacking.
  • Decoding pipeline: The decoder reads image attributes, applies inverse transforms, reconstructs per-primitive values, and writes an INRIA-style .ply file.The reference implementation uses metadata-defined quantization ranges and image files for decoding.
  • Decoded attributes: Decoded attributes include opacity, scales, means, quaternions, direct spherical-harmonic coefficients, and 45 remaining coefficients before active-mask filtering.Inverse transforms include sigmoid, exponential, signed-log inversion, and centroid lookup.

6.10 Comparison with State of the Art

On Mip-NeRF 360, KISS-GS remains competitive with recent 3DGS compression methods across the rate-distortion range, with particularly strong LPIPS and file-size results.

  • Mip-NeRF 360: KISS-GS remains competitive across the rate-distortion range on Mip-NeRF 360.The benchmark includes real-world scenes, held-out views, and multiple training and evaluation resolutions.
  • Mip-NeRF 360: KISS-GS shows particularly strong results in LPIPS and file size on this broader comparison.

6.11 Rate-Distortion Curves

Rate-distortion curves reveal dataset-dependent behavior: KISS-GS has a quality margin on Tanks and Temples and Mip-NeRF 360, but not Deep Blending, where the limitation originates mainly in the input reconstruction.

  • Dataset differences: KISS-GS has a clear quality margin over INRIA after reconstruction and compaction on Tanks and Temples and Mip-NeRF 360.That margin can be traded for bitrate during compression.
  • Dataset differences: Deep Blending lacks this margin, while the gap to HAC++ is already present before SOG-XT encoding.Compressed HAC++ exceeds uncompressed gsplat-MCMC and INRIA 3DGS on Playroom and Dr. Johnson.
  • Dataset differences: The Deep Blending gap is attributed to the underlying representation and its regularization rather than SOG-XT quantization or POPSpa pruning.View inspection links the mean gap to isolated floaters in a small number of views.
  • Dataset differences: POPSpa removes weakly constrained high-capacity splats, keeping 512k models close to HAC++ on full scenes.

6.12 Qualitative Comparison

The Bicycle view comparison shows that KISS-GS preserves major foreground structure at a low primitive budget, while fine background details degrade.

  • Bicycle qualitative comparison: At 128k primitives and 1.9 MB, KISS-GS reconstructs the main foreground structure well while retaining competitive PSNR.
  • Bicycle qualitative comparison: Fine background detail is lost at this operating point, including grass behind the pedals and a gray bicycle-frame label.The differences are visible in the 4× and 8× crops.

6.13 Full-Page Figures and Tables

The figures and tables compare KISS-GS with related methods across multiple scene datasets and evaluation views, highlighting dataset-dependent rate-distortion behavior and qualitative inspection regions.

  • Rate-distortion comparisons: On Tanks and Temples, KISS-GS matches the INRIA baseline at 1.3 MB in PSNR and 5.75 MB in LPIPS.The reconstruction and compaction pipeline substantially surpasses the INRIA 40k baseline, creating headroom for compression.
  • Rate-distortion comparisons: On Mip-NeRF 360, the ordering is similar across PSNR, SSIM, and LPIPS, although gaps are closer for SSIM and LPIPS than for PSNR.Some comparison results are self-reported and may not follow the evaluation protocol in Section 6.7.
  • Rate-distortion comparisons: On Deep Blending, KISS-GS does not substantially improve over INRIA, with the gap to HAC++ appearing before SOG-XT encoding.The figure attributes the limitation to standard-splat reconstruction under these captures rather than image-based compression alone.
  • Rate-distortion comparisons: On Synthetic NeRF scenes, all methods achieve higher absolute quality metrics because the scenes are synthetic, spatially smaller, and free of real-world capture effects.This figure provides dataset context rather than a specific method ranking.
  • Qualitative comparisons: The Bicycle qualitative comparison evaluates INRIA 40k, HAC++-44k, and KISS-GS across five primitive budgets using PSNR, LPIPS, and file size.The listed inspection regions include grass behind the pedals, the bicycle-frame label, and the background behind the wheels at increasing crop factors.
Loading 2608.26948v1…