Feed-Forward 3DGS Compression

• Feed-Forward 3DGS Compression is a method that enables optimization-free, rapid encoding of 3D Gaussian Splatting scenes using learned priors and deterministic inference.
• It leverages transform coding, adaptive entropy models, and context prediction to achieve compression ratios ranging from 20× to 300× while maintaining high visual fidelity.
• The framework generalizes across static and dynamic scenes without per-scene retraining, ensuring scalable, real-time 3D content delivery for diverse applications.

Feed-Forward 3DGS Compression Frameworks are a class of methods for compact, generalizable, optimization-free representation of 3D Gaussian Splatting (3DGS) scenes. These frameworks leverage learned priors, transform coding, advanced entropy models, and context prediction to encode scenes rapidly and at extremely high compression ratios—often with minimal visual fidelity loss relative to the original canonical 3DGS. Unlike optimization-based methods that require costly per-scene retraining, feed-forward pipelines execute encoding and decoding entirely through deterministic neural (or geometric) inference, affording high scalability and transferability across scenes, modalities, and use cases.

1. Foundations and Scene Representation

Feed-forward compression frameworks operate on 3DGS representations, each modeled as a set of anisotropic Gaussian “splats” $\{(p_i, \Sigma_i, o_i, C_i)\}_{i=1}^N$, where $p_i \in \mathbb{R}^3$ is position, $\Sigma_i$ is the covariance, $o_i$ is opacity, and $C_i$ contains color coefficients (often spherical harmonics). Datasets such as WideRange4D, MeetRoom, Mip-NeRF360, Tanks & Temples, and DeepBlending are widely used for benchmarking.

A key feature of feed-forward frameworks is the avoidance of scene-specific retraining: all parameters of the compression models are learned once over large multi-scene datasets, and encoding/decoding executes in a single, fully parallel inference pass (Zhang et al., 8 Jul 2025).

Dynamic scenes are handled by modeling a sequence of Gaussian frames $X_t$, and compression exploits inter-frame redundancy through sparse motion representations.

2. Motion Extraction, Prediction, and Transform Coding

Dynamic 3DGS compression (e.g., D-FCGS (Zhang et al., 8 Jul 2025)) employs sparse control-point motion extraction using methods such as Farthest Point Sampling. The positions, orientations, and scales of control points are feature-encoded (using frequency encoding and MLPs) at consecutive frames. Their difference vectors form a low-dimensional motion tensor $m_j$, which succinctly encodes temporal dynamics across frames.

For static scenes, learned transforms such as Karhunen–Loève (KLT) and linear projection are applied to anchor features to decorrelate and sparsify the data (SHTC (Xu et al., 28 May 2025)). Hierarchical structures—base layer (decorrelated/quantized principal components) and enhancement layer (sparse residual code via ISTA unfolding)—further improve the rate-distortion trade-off.

In highly generalizable settings, advanced signal-processing techniques—adaptive voxelization (Wang et al., 30 May 2025), Morton-ordering (Liu et al., 30 Nov 2025), and tri-plane projection (Wang et al., 26 Mar 2025, Zhan et al., 1 Mar 2025)—convert unstructured Gaussians into compact, spatially ordered representations.

3. Entropy Modeling and Bitstream Construction

Accurate entropy modeling is central to feed-forward 3DGS compression:

• Dual Prior Entropy Models fuse hyperpriors (learned latent distributions) with spatial-temporal context priors (multi-level hash grids, local neighborhood features) for precise rate estimation of latent motion tensors or attributes, greatly improving compression relative to simpler factorized models (Zhang et al., 8 Jul 2025).
• Mixture-of-Priors architectures (MoP, (Liu et al., 6 May 2025)) integrate multiple diverse MLP “experts” with a gating mechanism, producing richer prior distributions for element-wise quantization and context-adaptive entropy coding.
• Space-channel auto-regressive models exploit both spatial and channel correlations: by Morton-serializing Gaussians and grouping by context windows (LocoMoco (Liu et al., 30 Nov 2025)), high-dimensional context is leveraged for fine-grained probability estimation.
• Autoregressive and context-adaptive models (CAT-3DGS (Zhan et al., 1 Mar 2025)) encode both inter-anchor (spatial) and intra-anchor (channel) dependencies.
• Feedforward frameworks integrate context queries from hash grids, feature planes, or tri-planes directly as input to entropy models.

The final bitstream typically includes losslessly compressed positions, quantized and entropy-coded attributes, mask vectors for pruning, network weights for decoding, and various auxiliary data (e.g., rate-control metadata) (Zhang et al., 8 Jul 2025, Yang et al., 3 May 2025).

4. Reconstruction, Refinement, and Residual Compensation

Decoding in feed-forward frameworks involves restoring the scene geometry and color:

• Motion compensation is performed via control-point-guided upsampling. Each Gaussian interpolates motion updates from the $K$ nearest control points, weighted by exponential distance. Geometric parameters (position, scale, orientation) are updated accordingly (Zhang et al., 8 Jul 2025).
• Color refinement uses spatial-temporal context to predict residual color adjustments, which are quickly inferred and added to the compensated color coefficients.
• For scenes suffering from aggressive quantization or pruning, restoration networks (learned U-Nets, NAF blocks) denoise rendered images via side information channels (e.g., JPEG-XL compressed residuals), then "pull back" this supervision into Gaussian parameter optimization for improved rate-distortion (Shin et al., 16 Oct 2025).
• Mask-guided diffusion refinement (ExGS (Chen et al., 29 Sep 2025)) lifts compressed renderings through a lightweight VAE and a one-step diffusion model, inpainting and enhancing images in a single forward pass.

These steps are feed-forward, avoiding any iterative optimization or retraining at decode time.

5. Architectural Variants and Rate-Distortion Training

Several architectural paradigms are established in the literature:

• I/P Group-of-Frame coding (D-FCGS (Zhang et al., 8 Jul 2025)) for dynamic scenes, with full I-frame Gaussian coding and P-frame motion encoding.
• Tri-plane and feature-plane approaches (CAT-3DGS (Zhan et al., 1 Mar 2025), TC-GS (Wang et al., 26 Mar 2025, Lee et al., 6 Jan 2025)) replace per-anchor attributes with spatially organized 2D planes, improving entropy coding via structured contexts and standard codecs.
• Prediction-based feed-forward modules leverage hash grids for scene structure prediction and residual compensation to recover fine-grained lost details at low bitrates (Ma et al., 30 Mar 2025).
• Hybrid schemes combine neural and point-cloud codecs, using dual-channel sparse feature representations and standardized GPCC encoding for explicit, interpretable bitstreams (HybridGS (Yang et al., 3 May 2025)).
• Attention-based transform encoders, space-channel auto-regressive entropy networks, and mixture-of-priors feature blocks are jointly trained under rate-distortion losses:

$$\mathcal{L}_{\text{total}} = \lambda D_{\text{SSIM}} + (1-\lambda)\|I_{\text{render}} - I_{\text{gt}}\|_1 + \alpha\,L_{\text{rate}}$$

for image fidelity (SSIM, MSE), plus cross-entropy or sparsity regularizers for coded residuals.

• Masking strategies prune low-impact primitives (anchors rarely seen, low opacity, or with limited rendering effect). View frequency-aware thresholds adapt masking to scene coverage or complexity.

6. Quantitative Performance, Scalability, and Practical Significance

Recent feed-forward 3DGS compression frameworks attain extreme compression ratios (20×–300×) with negligible loss in PSNR/SSIM/LPIPS (Zhang et al., 8 Jul 2025, Ma et al., 30 Mar 2025, Song et al., 11 Jun 2025, Wang et al., 26 Mar 2025, Lee et al., 6 Jan 2025). Representative metrics:

Method	Compression Ratio	PSNR (dB)	Encoding Time (s)	Decoding Time (s)
D-FCGS	>40×	32.02	0.61 (P-frame)	0.72 (P-frame)
FCGS	>20×	28.86–29.26	1–11 (scene)	9–20 (scene)
TinySplat	>100×	26.32–27.43	1.02	0.042
ExGS	>100×	24–27	0.066	0.066

Rate-distortion competitiveness is consistently shown against SOTA optimization-based baselines (K-Planes, HyperReel, HiCoM, QUEEN, 4DGC, HAC++). Ablations reveal the criticality of large-scale context (e.g., Morton-ordering, multi-block attention) and context fusion for high rate-distortion performance (Liu et al., 30 Nov 2025, Wang et al., 29 May 2025).

Scalability is dramatically improved: frameworks such as ZPressor (Wang et al., 29 May 2025) reduce memory and computational cost from $O(K)$ (linear in input views) to $O(N)$ (number of anchor views, $N \ll K$), allowing real-time inference on large-scale, multi-view datasets. Encoding/decoding is fully parallelizable and hardware-efficient—many schemes complete single-scene streaming in <$2$ seconds.

7. Generalization, Limitations, and Future Directions

Feed-forward 3DGS compression schemes are designed for plug-and-play use in immersive media, FVV transmission, and scalable storage. No per-scene fine-tuning is required; a single trained model generalizes across unseen scenes and datasets.

Limitations include mild quality degradation at extremely low bitrates (relative to per-scene optimized baselines) and occasional generalization gaps for atypical SH distributions. Potential future directions focus on further enhancing context models, refining attribute transforms, and integrating pruning and adaptive masking for dynamic environments.

These frameworks are foundational for next-generation neural graphics pipelines, enabling high-fidelity, compact, and generalizable 3D content delivery (Zhang et al., 8 Jul 2025, Ma et al., 30 Mar 2025, Chen et al., 29 Sep 2025, Wang et al., 29 May 2025).