Principled understanding and mitigation of WD-induced artifacts in 3D Gaussian Splatting

Establish the root causes of the web-like texture artifacts that arise when optimizing 3D Gaussian Splatting representations using Wasserstein Distortion as the primary training loss without pixel-level regularization, particularly in regions with few training samples and under tight splat-count constraints, and develop a principled remedy that suppresses these artifacts while preserving the perceptual benefits of Wasserstein Distortion.

Background

The paper observes that Wasserstein Distortion (WD), which performs well as a 2D image distortion measure in learned image compression, can introduce web-like artifacts when used alone to optimize 3D Gaussian Splatting (3DGS), especially in sparsely observed regions or when the number of splats is tightly constrained.

The authors propose a practical workaround—WD-Regularized (WD-R), which adds a lightly weighted pixel-level fidelity term (L1+SSIM) to WD—and show that WD-R suppresses the artifacts and improves human-rated perceptual quality. However, the underlying mechanism for why WD alone leads to these artifacts in 3DGS remains unclear, motivating a need for a more principled solution beyond ad hoc regularization.