ExGS: Extreme 3D Gaussian Compression with Diffusion Priors (2509.24758v1)

Abstract: Neural scene representations, such as 3D Gaussian Splatting (3DGS), have enabled high-quality neural rendering; however, their large storage and transmission costs hinder deployment in resource-constrained environments. Existing compression methods either rely on costly optimization, which is slow and scene-specific, or adopt training-free pruning and quantization, which degrade rendering quality under high compression ratios. In contrast, recent data-driven approaches provide a promising direction to overcome this trade-off, enabling efficient compression while preserving high rendering quality. We introduce \textbf{ExGS}, a novel feed-forward framework that unifies \textbf{Universal Gaussian Compression} (UGC) with \textbf{GaussPainter} for \textbf{Ex}treme 3D\textbf{GS} compression. \textbf{UGC} performs re-optimization-free pruning to aggressively reduce Gaussian primitives while retaining only essential information, whereas \textbf{GaussPainter} leverages powerful diffusion priors with mask-guided refinement to restore high-quality renderings from heavily pruned Gaussian scenes. Unlike conventional inpainting, GaussPainter not only fills in missing regions but also enhances visible pixels, yielding substantial improvements in degraded renderings. To ensure practicality, it adopts a lightweight VAE and a one-step diffusion design, enabling real-time restoration. Our framework can even achieve over $100\times$ compression (reducing a typical 354.77 MB model to about 3.31 MB) while preserving fidelity and significantly improving image quality under challenging conditions. These results highlight the central role of diffusion priors in bridging the gap between extreme compression and high-quality neural rendering. Our code repository will be released at \href{https://github.com/chenttt2001/ExGS}{here}.

• The paper introduces ExGS, a framework that integrates UGC and GaussPainter to achieve over 100x compression of 3D Gaussian Splatting representations while maintaining rendering quality.
• The methodology employs voxel-based pruning, adaptive amplification, and LZMA lossless compression to retain essential Gaussian primitives effectively.
• Diffusion priors in GaussPainter restore detail lost during aggressive pruning, enabling near real-time, high-fidelity 3D scene renderings.

Extreme 3D Gaussian Compression with Diffusion Priors

The paper "ExGS: Extreme 3D Gaussian Compression with Diffusion Priors" presents a novel approach for compressing 3D Gaussian Splatting (3DGS) neural scene representations using data-driven methods. This compression framework, named ExGS, integrates Universal Gaussian Compression (UGC) and GaussPainter to achieve high compression rates while preserving rendering quality through powerful diffusion priors.

Introduction

3D Gaussian Splatting (3DGS) has gained popularity for its efficient representation of 3D scenes, allowing real-time rendering. However, the large storage and transmission requirements restrict its deployment in resource-constrained environments, such as mobile devices and settings with limited bandwidth. Existing methods for compressing 3DGS involve either slow, scene-specific optimization or training-free pruning and quantization, both of which struggle with maintaining quality under extreme compression.

ExGS addresses this challenge by introducing a feed-forward framework that unifies aggressive compression and high-quality rendering. UGC performs re-optimization-free pruning to retain only essential Gaussian primitives, while GaussPainter uses diffusion priors to restore high-quality renderings from these compressed scenes. This approach enables compression of over $100\times$, as shown by reducing a typical 354.77 MB model to about 3.31 MB.

Methodology

Universal Gaussian Compression (UGC)

UGC compresses Gaussian scene representations by evaluating the global significance of Gaussian primitives using a significance score that incorporates visibility, opacity, and ray transmittance. The compression process involves:

• Global Significance Score: A score calculated for each Gaussian to assess its importance based on visibility and contribution to the scene.
• Voxel-based Pruning: Ensuring spatial uniformity by selecting significant Gaussians within divided voxels. This balances overall importance with spatial coverage, addressing sparse region issues.
• Adaptive Amplification: Predicting amplification factors for selected Gaussians to enhance spatial distribution, mitigating sparse region artifacts.
• Lossless Compression: Further reducing the size using LZMA-based compression, achieving extreme compression ratios without sacrificing numerical precision.

  Figure 1: The overall pipeline of ExGS, illustrating the integration of UGC and GaussPainter for efficient compression and high-fidelity rendering.

GaussPainter

GaussPainter leverages diffusion priors to fill in details lost during aggressive pruning:

• Diffusion Models: GaussPainter incorporates diffusion models for refining Gaussian scenes, addressing information loss.
• Latent Supervision: Introduces latent space supervision to guide the restoration process, improving the fidelity of reconstructed images.
• Mask Guidance: Utilizes visibility masks from 3DGS rendering to guide diffusion, enhancing areas likely to be missing or corrupted.
• Real-Time Rendering: The pipeline is optimized for speed, replacing heavy computational processes with a lightweight autoencoder and one-step diffusion, ensuring near real-time restoration suitable for interactive applications.

  Figure 2: Qualitative results at a 10% compression ratio on ScanNet++, demonstrating superior fidelity and completeness compared to baseline methods.

Experimental Results

The ExGS framework demonstrates state-of-the-art performance across multiple datasets, including ScanNet++, Replica, and Mip-NeRF360. It achieves significant gains in PSNR, SSIM, and LPIPS across various compression ratios compared to existing methods.

Tables and comparisons indicate that ExGS consistently outperforms both compression-only and generative approaches, especially under aggressive compression conditions. The integration of UGC and GaussPainter facilitates robust preservation of structural details and perceptual quality, even at extreme compression levels.

Conclusion

ExGS offers a robust solution for compressing 3DGS representations, balancing extreme compression and high rendering fidelity. The integration of novel data-driven modules enables real-time applications without significant quality loss. Future work could explore extensions of ExGS to other complex scene representations or real-time VR/AR applications.

This framework positions itself as a pivotal advancement for deploying neural representations in bandwidth-constrained and resource-limited environments.