VEDAL: Variational Error-Driven Asynchronous Learning for 3D Gaussian Splatting Pruning

2026-06-01

AI Summary

This work addresses the high memory footprint of 3D Gaussian Splatting, which limits its practical deployment. Existing pruning methods rely on heuristic scoring or synchronous updates, struggling to balance compression ratio and reconstruction fidelity. To overcome this, the paper formulates Gaussian pruning as a variational free energy minimization problem and introduces an asynchronous pruning mechanism guided by reconstruction uncertainty, along with a learnable prior-based variational uncertainty head. This approach uniquely unifies the trade-off between model complexity and reconstruction accuracy from an information-theoretic perspective. Experiments demonstrate a 5.2× compression rate with only a 0.31 dB PSNR drop across multiple datasets, significantly outperforming PUP 3D-GS and LightGaussian while maintaining real-time rendering at 185 FPS.

Abstract

3D Gaussian Splatting (3DGS) achieves remarkable novel view synthesis quality with real-time rendering, yet suffers from excessive memory consumption due to millions of Gaussian primitives. Existing pruning methods rely on heuristic importance scores or synchronous batch updates, leading to suboptimal compression and training instability. We propose VEDAL, a principled framework that formulates Gaussian pruning as variational free energy minimization. Our approach introduces (1) a prediction-error gating mechanism that asynchronously activates pruning based on per-Gaussian reconstruction uncertainty, and (2) a variational uncertainty head that models pruning decisions as latent variables with learnable priors. The free energy objective naturally balances reconstruction fidelity against model complexity through an information-theoretic lens. Extensive experiments on Mip-NeRF 360, Tanks&Temples, and Deep Blending demonstrate that VEDAL achieves 5.2x compression with only 0.31 dB PSNR drop, outperforming PUP 3D-GS by +0.05 dB at a higher compression ratio and LightGaussian by +0.35 dB at comparable quality, while maintaining real-time rendering at 185 FPS.

Problem

• 3D Gaussian Splatting
• pruning
• memory consumption
• compression
• training instability

Innovation

• Variational Free Energy
• Asynchronous Pruning
• Prediction-Error Gating
• Uncertainty Modeling
• 3D Gaussian Splatting