Micro-splatting: Multistage Isotropy-informed Covariance Regularization Optimization for High-Fidelity 3D Gaussian Splatting

  • 2025-09-02 10:05:44
  • Jee Won Lee, Hansol Lim, Sooyeun Yang, Jongseong Brad Choi
  • 0

Abstract

High-fidelity 3D Gaussian Splatting methods excel at capturing fine texturesbut often overlook model compactness, resulting in massive splat counts,bloated memory, long training, and complex post-processing. We presentMicro-Splatting: Two-Stage Adaptive Growth and Refinement, a unified,in-training pipeline that preserves visual detail while drastically reducingmodel complexity without any post-processing or auxiliary neural modules. InStage I (Growth), we introduce a trace-based covariance regularization tomaintain near-isotropic Gaussians, mitigating low-pass filtering inhigh-frequency regions and improving spherical-harmonic color fitting. We thenapply gradient-guided adaptive densification that subdivides splats only invisually complex regions, leaving smooth areas sparse. In Stage II(Refinement), we prune low-impact splats using a simple opacity-scaleimportance score and merge redundant neighbors via lightweight spatial andfeature thresholds, producing a lean yet detail-rich model. On fourobject-centric benchmarks, Micro-Splatting reduces splat count and model sizeby up to 60% and shortens training by 20%, while matching or surpassingstate-of-the-art PSNR, SSIM, and LPIPS in real-time rendering. These resultsdemonstrate that Micro-Splatting delivers both compactness and high fidelity ina single, efficient, end-to-end framework.

 

Quick Read (beta)

loading the full paper ...