Fast Feedforward 3D Gaussian Splatting Compression

• Per-scene optimization enables strong rate-distortion performance but makes existing 3DGS compression slow because of time-consuming finetuning.Source: §1 · p. 2
• FCGS compresses existing 3DGS in a single feed-forward pass without finetuning, preserving parameter values and structural integrity.Source: §1 · p. 2
• The Multi-path Entropy Module assigns Gaussian attributes to different processing paths to balance compressed size and fidelity.Source: §1 · p. 2
• Inter- and intra-Gaussian context models capture structural relationships among unorganized Gaussian attributes to eliminate redundancies.Source: §1 · p. 2
• FCGS achieves a compression ratio exceeding 20× while maintaining excellent fidelity and surpassing most optimization-based methods.Source: §1 · p. 2

• 3DGS uses fully explicit attributed Gaussians for fast rasterized rendering, but this representation substantially increases storage requirements.Source: §2 · p. 3
• Value-based methods prune, cluster, or quantize parameters according to importance or similarity, often altering values and requiring finetuning to recover information.Source: §2 · p. 3
• Structure-based methods exploit relationships among parameters using grids, anchors, or other organized representations to reduce redundancies in sparse, unorganized 3DGS.Source: §2 · p. 3
• Existing 3DGS compression approaches require per-scene finetuning, whereas this work explores optimization-free context models for Gaussian primitives.Source: §2 · p. 3

3 FAST COMPRESSION OF 3D GAUSSIAN SPLATTING

• FCGS uses an autoencoder-based, hyperprior compression pipeline that maps Gaussian attributes into quantized latents, estimates their distributions, and reconstructs them for rendered-image fidelity.Arithmetic coding operates on the latent and hyperprior representations, while MLP transforms implement the analysis, synthesis, and hyperprior networks.Source: §3 · p. 3, Figure 2 · p. 4, §3.2 · p. 4

3.1 PRELIMINARIES AND DISCUSSIONS

• 3DGS represents scenes with many Gaussians whose geometry and color attributes have different sensitivity, with geometry directly controlling rasterization dependencies.The model denotes geometry as 8 dimensions, color as 48 dimensions, and their concatenation as 56-dimensional Gaussian attributes.Source: §3.1 · p. 4

3.2 VALUE-BASED PRINCIPLE: MULTI-PATH ENTROPY MODULE (MEM)

• MEM assigns color attributes to latent or bypass paths while leaving geometry untransformed, preserving sensitive rasterization information while compressing less-sensitive color values.The binary mask selects either transformed quantized color attributes or directly quantized attributes; its rate is learned through bit consumption rather than an additional mask-rate loss.Source: §3.2 · p. 5

3.3 STRUCTURE-BASED PRINCIPLE: AUTOREGRESSIVE CONTEXT MODELS

• FCGS builds inter-Gaussian context by sequentially decoding batches, creating interpolated grids from decoded Gaussians to predict distributions for the next batch without changing Gaussian locations.Weighted interpolation uses nearby decoded latent features, and multi-resolution 3D and 2D grids capture spatial relationships and local detail.Source: §3.3 · pp. 5–6
• Intra-Gaussian context predicts later latent chunks from earlier ones, using four channel chunks for transformed color and three RGB-axis chunks for bypassed color, while omitting geometry.A Gaussian mixture model combines hyperprior, inter-Gaussian, and intra-Gaussian distribution parameters for probability estimation.Source: §3.3 · pp. 6–7

3.4 TRAINING AND CODING PROCESS

• The training objective combines rendered-image fidelity with entropy, weighting MEM path bits by the learned mask while balancing fidelity and compression through λ.Fidelity uses both MSE and SSIM, and decoding reconstructs latent chunks with GMM probabilities before arithmetic decoding.Source: §3.4 · p. 7

4.1 IMPLEMENTATION DETAILS

• The evaluation reports PSNR as the primary fidelity metric relative to compressed size, with SSIM and LPIPS provided elsewhere.Implementation uses variable bitrates by adjusting the rate-distortion parameter λ.Source: §4.1 · pp. 7–8

4.2 EXPERIMENT EVALUATION

• FCGS surpasses most optimization-based methods despite lacking per-scene adaptation, while qualitative results show high fidelity after compression.The comparison uses DL3DV-GS, MipNeRF360, and Tanks&Temples, with baselines that either finetune existing 3DGS or train from scratch.Source: §4.2 · pp. 8–9
• 15× and 5× compression ratios are achieved zero-shot on 3DGS from MVSplat and LGM, respectively, using different fidelity references for reconstruction and generative models.MVSplat fidelity is measured against ground truth, whereas LGM fidelity compares renders before and after compression because ground truth is unavailable.Source: §4.2 · p. 9, Figure 6 · p. 9

4.4 BOOST EXISTING COMPRESSION APPROACHES

• 40% size and 100× compression are achieved when FCGS is applied to pruned representations, respectively relative to MSC at equal fidelity and vanilla 3DGS.The pruning-compatible evaluation is conducted on Deep Blending, and FCGS is applied to the same pruned representations as MSC.Source: Figure 8 · p. 10, §4.4 · p. 10

4.3 ABLATION STUDY

• MEM balances redundancy removal and fidelity by processing selected color attributes through an autoencoder while directly quantizing geometry attributes.Processing all attributes through the autoencoder can prevent rasterization, all-mask-one processing reduces fidelity, and all-mask-zero processing increases storage.Source: §4.3 · pp. 9–10
• 1.5× higher bit consumption occurs without the full inter- and intra-Gaussian context models under similar fidelity conditions.Removing these mutual-dependency models progressively decreases rate-distortion performance.Source: Figure 7 · p. 9, §4.3 · p. 10

4.5 CODING AND RENDERING EFFICIENCY ANALYSIS

• About 1 second is required to encode 100K Gaussians on one GPU, while decoded rendering remains close to the original at 102 versus 91 FPS.FCGS preserves the number and structure of Gaussians, so compression does not alter decoded rendering complexity.Source: §4.5 · p. 10

5 CONCLUSION

• FCGS provides fast compression of existing 3DGS without finetuning through a generalizable optimization-free pipeline.The conclusion presents this pipeline as the paper’s central contribution.Source: §5 · p. 10
• More than 20× compression is achieved through the MEM module and context models while maintaining rate-distortion performance.The conclusion attributes the result to the design of these modules.Source: §5 · p. 10
• FCGS can boost compression performance when combined with pruning-based methods.This extends the pipeline’s use beyond direct compression of unpruned 3DGS.Source: §5 · p. 10

– Appendix –

• The notation appendix defines Gaussian locations, geometry and color attributes, adaptive masks, latent variables, context-model parameters, and rate-distortion terms used throughout FCGS.It includes x as the autoencoder input, m as the adaptive MEM mask, and λ as the bit-fidelity tradeoff parameter.Source: – Appendix – · p. 15

A STATISTICAL DATA FOR DL3DV-GS

• 6,770 3DGS representations are generated from DL3DV, with 100 randomly selected scenes reserved for testing under a 1-test-view/7-train-view split.FCGS is evaluated on test views, whereas training-based methods are trained or optimized on training views before test-view evaluation.Source: A STATISTICAL DATA FOR DL3DV-GS · p. 16

B ANALYSIS OF LIMITATIONS

• All-zero color masks mitigate MEM failures on feed-forward 3DGS, where distribution gaps can otherwise reduce MVSplat PSNR by approximately 0.7dB.The workaround prioritizes fidelity but exposes limited generalization across differing SH and color-value distributions.Source: B ANALYSIS OF LIMITATIONS · p. 18

C EFFECT OF DIFFERENT RANDOM SEEDS ON GAUSSIAN SPLITTING

• Random Gaussian splitting is stable: five seeds produce a 8e−4 MB standard deviation and 3.6e−5 coefficient of variation without affecting reconstruction fidelity.Farthest Point Sampling instead increases average bits by 0.6% on scenes with fewer than 500K Gaussians because it creates non-uniform splits.Source: C EFFECT OF DIFFERENT RANDOM SEEDS ON… · p. 18

D DISCUSSIONS ON SOTA 3DGS COMPRESSION METHODS

• HAC and ContextGS achieve better compression performance than FCGS, but use per-scene optimization and anchor-based representations that are about 5× smaller than vanilla 3DGS.FCGS targets the standard 3DGS structure without per-scene optimization, leaving extension to anchor-based variants as future work.Source: D DISCUSSIONS ON SOTA 3DGS COMPRESSION… · p. 19

E MASK RATIO ANALYSIS IN MEM

• As λ increases, MEM assigns more color attributes to the m = 1 autoencoder path to reduce bitrate at the cost of reduced fidelity.Lower λ produces the opposite allocation pattern.Source: E MASK RATIO ANALYSIS IN MEM · p. 19

F MORE FIDELITY METRICS AND TRAINING TIME

• PSNR, SSIM, LPIPS, and training time are reported, with FCGS measured by encoding time rather than finetuning or training time.Asterisks distinguish finetuning from a common 3DGS and training from scratch.Source: F MORE FIDELITY METRICS AND TRAINING… · p. 20
• The DL3DV-GS and MipNeRF360 comparisons distinguish methods finetuned from a common 3DGS from methods trained from scratch.The tables also identify the training time of vanilla 3DGS using dataset-specific notation.Source: Table C · p. 20, Table D · p. 20
• The Tank&Temples comparison likewise separates finetuning from scratch training and reports vanilla 3DGS training time through the “(814 +)” notation.For FCGS, the comparison uses the same common 3DGS as the finetuned methods.Source: F MORE FIDELITY METRICS AND TRAINING… · p. 20

G ANALYSIS OF STORAGE SIZE OF DIFFERENT COMPONENTS

• As λ increases, more f_col moves to m = 1; the m = 0 path shrinks substantially, while mask size decreases because a more dominant binary value has lower entropy.Coordinates µ_g retain the same size because they use the same GPCC command across λ values.Source: G ANALYSIS OF STORAGE SIZE OF… · p. 21
• f_col (m = 1) is easiest to compress because it is encoded in an autoencoder latent space, whereas f_geo is harder because geometry has greater variability and less mutual information.With float32 originals, a component’s compression ratio is 32 divided by its compressed bits per parameter.Source: G ANALYSIS OF STORAGE SIZE OF… · p. 21
• Table F reports component storage in MB, while Table G reports per-parameter storage in bits and identifies the mask rate as the share of f_col assigned to m = 1.The reported components are µ_g, f_col for both paths, f_geo, and the mask.Source: G ANALYSIS OF STORAGE SIZE OF… · p. 21

H ANALYSIS OF ENCODING/DECODING TIME OF DIFFERENT COMPONENTS

• Decoding is faster than encoding overall because GPCC decoding avoids rate-distortion search, although arithmetic-decoder index search increases decoding complexity.Most encoding time is spent on GPCC, while arithmetic coding remains efficient.Source: H ANALYSIS OF ENCODING/DECODING TIME… · p. 22
• Higher λ lowers entropy and codec time, shifts more f_col to m = 1, and slightly increases neural forward time while reducing m = 0 processing time.The increased m = 1 path time is offset by lower time per parameter under stronger entropy constraints.Source: H ANALYSIS OF ENCODING/DECODING TIME… · p. 22
• Table H measures single-GPU encoding times for DL3DV-GS scenes averaging over 1.57 million Gaussians, while Table I reports corresponding component decoding times.Both tables include component percentage shares and the m = 1 mask rate.Source: H ANALYSIS OF ENCODING/DECODING TIME… · p. 22, Table I · p. 22

I ABLATION ON THE NUMBER OF SPLITS IN CONTEXT MODELS

• Increasing inter-Gaussian splits from 3 to 4 reduces size by 0.227 MB, while additional splits provide negligible reduction; FCGS therefore uses four splits.A 2-split configuration with a 95%, 5% ratio performs poorly, emphasizing the need for rational split strategies.Source: I ABLATION ON THE NUMBER OF SPLITS IN… · p. 23
• Increasing intra-Gaussian chunks does not consistently reduce size: two chunks increase size through insufficient context, while 8 or 16 chunks complicate convergence.Source: I ABLATION ON THE NUMBER OF SPLITS IN… · p. 23

J VISUALIZATION OF BIT ALLOCATION USING CONTEXT MODELS

• Inter-Gaussian bits per parameter decrease from B1 to B4, with the largest reduction from B1 to B2 because later batches gain progressively richer context.Random splitting gives batches similar average information, while B1 lacks inter-Gaussian context and later batches benefit from accumulated context.Source: J VISUALIZATION OF BIT ALLOCATION… · p. 24, Figure C · p. 24
• Intra-Gaussian allocation assigns minimal information to C1, most information to C2, and decreasing bits from C2 to C4 as context becomes richer.C1 lacks intra-Gaussian context, whereas C2 helps predict later chunks and benefits from C1; color components also receive different bit allocations.Source: J VISUALIZATION OF BIT ALLOCATION… · p. 24, Figure C · p. 24

K ABLATION STUDIES ON THE HYPERPRIOR

• Removing hyperprior information by setting ẑ to zero does not significantly affect compression performance.The experiment retains a Gaussian probability for entropy estimation, predicted from the zero hyperprior.Source: K ABLATION STUDIES ON THE HYPERPRIOR · p. 25
• The limited hyperprior contribution suggests that 3DGS has less redundancy within individual Gaussian attributes than images, making inter- and intra-Gaussian contexts more effective.The comparison is made with the image-compression hyperprior of Ballé et al. (2018).Source: K ABLATION STUDIES ON THE HYPERPRIOR · p. 25

L TRAINING WITH LESS DATA

• A 100-scene training set yields comparable fidelity with only a 12.9% size increase, while FCGS remains under 10 MB.The full training setup uses 6670 DL3DV scenes, and compression after training requires no further optimization.Source: L TRAINING WITH LESS DATA · p. 25

M BOOST COMPRESSION PERFORMANCE OF PRUNING-BASED APPROACHES

• Applying FCGS to pruned 3DGS from Mini-Splatting and Trimming achieves superior compression performance across DL3DV-GS, MipNeRF360, and Tank&Temples.The pruning methods first remove trivial Gaussians, after which FCGS further compresses the pruned representations.Source: M BOOST COMPRESSION PERFORMANCE OF… · p. 26, Figure D · p. 26

N ADDITIONAL QUALITATIVE COMPARISONS

• Qualitative comparisons show substantial size reduction while preserving high fidelity, with PSNR and size reported for each image.Source: N ADDITIONAL QUALITATIVE COMPARISONS · p. 27