KISS-GS: 3D Gaussian Splatting Compression Kept Simple
Authors: Wieland Morgenstern, Friedrich Elias Branschke, Florian Fleischmann, Adrian Szatmari, Paul Schlack, Florian Barthel, Peter Eisert, Anna Hilsmann
Organizations: Fraunhofer HHI, Germany · Humboldt-Universität zu Berlin, Germany · Technische Universität Berlin, Germany
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/
Figures & tables
Figure 1: Rate distortion on Mip-NeRF 360 for KISS-GS, with quality measured by peak signal-to-noise ratio (PSNR). At the INRIA 3DGS quality reference (INRIA-Q), KISS-GS reduces file size by 228× from 738 MB to 3.2 MB. From the uncompressed INRIA .ply baseline, POPSpa matches INRIA-Q with a 15.7× reduction, our SOG-XT encoding adds 6.6× , and encoding- aware fine-tuning adds a further 2.2× , outperforming HAC++ [ 5 ] .
Figure 2: The KISS-GS pipeline with four distinct stages. (1) A standard 3DGS reconstruction yields a .ply , with primitive count setting the rate-distortion point. (2) POPSpa compacts the scene. (3) SOG-XT encodes the compacted primitives into attribute images. (4) Optional codec-aware fine-tuning improves rate-distortion without changing the output format. Below the dashed separator, we show per-stage alternatives to the KISS-GS stages and contrast them with integrated methods like HAC++ spanning all stages as a single non-modular unit.
Figure 3: Encoding overview for SOG-XT. PLAS sorts primitives into 2D attribute grids using both mean byte planes. View dependent color is reduced with a k-means codebook whose centroids are PLAS sorted again, and assignments are stored as UV indices. Bars show the relative on-disk size contribution of each component.
Figure 4: Runtime breakdown on the Bicycle scene. The top plot shows wall-clock time per KISS-GS stage and the bottom plot breaks down post-hoc SOG-XT encoding. Bicycle belongs to Mip-NeRF 360 , and 512k is the smallest Bicycle point exceeding INRIA-40k PSNR. At that point, reconstruction takes 1215 s, POPSpa compaction takes 277 s, post-hoc SOG-XT encoding takes 34 s, and optional encoding-aware fine-tuning adds 441 s. For context, Fig. 1 reports stage-wise size reductions of 15.7x, 6.6x, and 2.2x on Mip-NeRF 360 at INRIA-Q. Encoding costs are balanced between components for small scenes, while SH codebook clustering dominates at high primitive counts.
Figure 5: Parameter sensitivity of encoding-aware fine-tuning on the Bicycle scene with 512k Gaussians. We vary one parameter at a time around the default setting used in the paper: SH codebook size ( K∈N/{2,4,12,16,32,64} compared with the default N/8 ), PRAS timing, TV regularization (off, stronger), fine-tuning steps from 0.5k to 12k, quaternion quantization range 47–255, and SH-centroid quantization range 31–255. Points are means over seven runs, with error bars showing one standard deviation.
Table 5: Scores from the original HAC++ paper [ 5 ] and our own runs on Mip-NeRF 360 . We list the reported 30k numbers, then our recomputed results with training resolution equal to evaluation resolution, and finally our recomputed results under the full evaluation protocol with 1600 px test resolution. Scene size decreases up to 40k iterations and stabilizes thereafter, indicating convergence.
Figure 7: Rate-distortion performance of KISS-GS and related work on the Mip-NeRF 360 scenes. Methods marked with a dagger † are self-reported by the authors, as collected by the 3DGS.zip survey [ 1 ] , and may not follow our evaluation protocol from Sec. 6.7 .
Figure 8: Rate-distortion performance of KISS-GS on the Tanks and Temples scenes. Our reconstruction and compaction pipeline substantially surpasses the INRIA 40k baseline, which creates headroom for compression. As a result, KISS-GS can match the INRIA baseline at file sizes as low as 1.3 MB in PSNR and 5.75 MB in LPIPS.
Figure 9: Rate-distortion performance of KISS-GS on the Mip-NeRF 360 scenes: In addition to the PSNR plot in Fig. 1 , we report SSIM and LPIPS. The ordering is similar across metrics, although the gaps are closer for SSIM and LPIPS than for PSNR.
Figure 10: Rate-distortion performance of KISS-GS on the Deep Blending scenes. Unlike on the other datasets, our reconstruction and compaction pipeline does not substantially improve over the INRIA baseline. The gap to HAC++ is already present before SOG-XT encoding, which indicates that the limiting factor is the standard-splat reconstruction under these captures rather than the image-based compression stage alone. As a result, none of the operating points of our compressed curve exceeds INRIA quality, while HAC++ [ 5 ] remains above INRIA for part of the range.
Figure 11: Rate-distortion performance of KISS-GS on the Synthetic NeRF scenes. Because these scenes are synthetic, spatially smaller, and free of real-world capture effects, absolute quality metrics are higher for all methods.
Table 7: Qualitative comparison on evaluation view 17 of the Bicycle scene. We compare INRIA 40k, HAC++-44k, and our method (POPSpa + SOG-XT + Finetuning) at five primitive budgets. The Metrics column reports PSNR, LPIPS, and file size for this view. Informative regions to inspect are the grass behind the pedals in the 4x crop, the gray label on the bicycle frame on the left in the 8x crop, and the background behind the wheels in the 12x crop.
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. Note that, 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.
Pengpeng Yu, Runqing Jiang, Qi Zhang +3
Sun Yat-sen University, China · Pengcheng Laboratory, China · Peking University, China
3D Gaussian Splatting (3DGS) enables high-quality novel-view synthesis but requires substantial storage. Existing compression methods often rely on spatial context modeling over irregular 3D representations, increasing the complexity of training and coding. Meanwhile, floating-point context inference can introduce numerical inconsistencies across platforms, causing entropy-decoding failures. To address these practical challenges, we propose COSA-GS, which constructs context without spatial aggregation through anchor-wise causal factorization. Specifically, we use geometry context derived from each anchor's coordinates to model a compact learnable anchor latent. The anchor latent is then fused with the geometry context to form an anchor context for attribute coding. The resulting context model features a simple architecture composed solely of linear transformations and activations. We train COSA-GS using rate--distortion optimization with adaptive Gaussian pruning. Further, we develop quantization-aware training and integer inference for the context model to achieve bit-exact consistency of entropy-decoded symbols across platforms. Experiments demonstrate that COSA-GS achieves state-of-the-art compression performance while retaining fast and consistent cross-platform decoding, providing a simple yet effective framework for practical 3DGS compression. Code is available at https://github.com/pengpeng-yu/COSA-GS.
Pengpeng Yu, Yueru Chen, Fei Song +4
Sun Yat-sen University, China · Pengcheng Laboratory, China · Academy of Broadcasting Science, National Radio and Television Administration, China +1
3D Gaussian Splatting (3DGS) achieves high-quality novel view synthesis with real-time rendering, but its storage cost remains prohibitive for practical deployment. Existing post-training compression methods still rely on many coupled hyperparameters across pruning, transformation, quantization, and entropy coding, making it difficult to control the final compressed size and fully exploit the rate-distortion trade-off. We propose MesonGS++, a size-aware post-training codec for 3D Gaussian compression. On the codec side, MesonGS++ combines joint importance-based pruning, octree geometry coding, attribute transformation, selective vector quantization for higher-degree spherical harmonics, and group-wise mixed-precision quantization with entropy coding. On the configuration side, it treats the reserve ratio and bit-width allocation as the dominant rate-distortion knobs and jointly optimizes them under a target storage budget via discrete sampling and 0--1 integer linear programming. We further propose a linear size estimator and a CUDA parallel quantization operator to accelerate the hyperparameter searching process. Extensive experiments show that MesonGS++ achieves over 34× compression while preserving rendering fidelity, outperforming state-of-the-art post-training methods and accurately meeting target size budgets. Remarkably, without any training, MesonGS++ can even surpass the PSNR of vanilla 3DGS at a 20× compression rate on the Stump scene. Our code is available at https://github.com/mmlab-sigs/mesongs_plus
Shuzhao Xie, Junchen Ge, Weixiang Zhang +10
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