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

Original: arXiv:2606.02346v1 Announce Type: new 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.

Rewritten:

Abstract: Although 3D Gaussian Splatting (3DGS) delivers high-quality novel view synthesis alongside real-time rendering capabilities, it is hindered by substantial memory demands stemming from the presence of millions of Gaussian primitives. Current pruning techniques, which typically depend on heuristic importance metrics or synchronous batch updates, often result in unstable training and less-than-optimal compression rates. To address these limitations, we introduce VEDAL, a rigorous framework that treats Gaussian pruning as a problem of variational free energy minimization. This methodology incorporates two key innovations: first, a prediction-error gating system that triggers pruning asynchronously according to the reconstruction uncertainty associated with individual Gaussians; and second, a variational uncertainty head that treats pruning choices as latent variables equipped with learnable priors. By leveraging an information-theoretic perspective, the free energy objective inherently manages the trade-off between model complexity and reconstruction accuracy. Comprehensive evaluations across the Mip-NeRF 360, Tanks&Temples, and Deep Blending datasets reveal that VEDAL secures a 5.2x compression ratio with a negligible PSNR decline of just 0.31 dB. In terms of performance, it surpasses PUP 3D-GS by +0.05 dB despite achieving a higher compression rate, and exceeds LightGaussian by +0.35 dB at equivalent quality levels, all while sustaining real-time rendering speeds of 185 FPS.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC