Drop-in Perceptual Optimization for 3D Gaussian Splatting

3D Gaussian Splatting (3DGS) ultimately produces results meant to be viewed by humans, yet it often still relies on an ad-hoc combination of pixel-wise losses, which makes rendering prone to blurriness. To address this, various distortion losses were systematically compared, and a large-scale human subjective evaluation was conducted for the first time, including 39,320 pairwise ratings.

The results showed that WD-R, a normalized version of Wasserstein Distortion, emerged as the strongest choice. WD-R better reconstructed fine textures without increasing the number of splats, and was preferred 2.3x more often by raters compared to the existing 3DGS loss, and 1.5x more often compared to Perceptual-GS, which had been rated as the current state of the art.

In quantitative metrics as well, WD-R consistently improved LPIPS, DISTS, and FID across multiple datasets. It was also applied as-is to recent frameworks such as Mip-Splatting and Scaffold-GS, and simply swapping in WD-R for the existing loss improved perceptual quality within a similar resource budget, with human evaluation showing preference increases of 1.8x and 3.6x, respectively.

This approach also extended to 3DGS scene compression. While maintaining similar perceptual metric performance, it enabled roughly 50% bitrate reduction, aiming to achieve both improved rendering quality and compression efficiency simultaneously.