SplatAnything wins this comparison. In the authors' own seven-dataset benchmark it ranks first on most datasets and settings against Mirror2-GS and other feed-forward rivals, and its weights and inference code went public on 8 November 2026 ( paper , GitHub ). It is the better pick for prototypes and products built outside the EU, UK and South Korea, provided your platform stays under 1 million monthly active users. Anyone else needs to read the licence before writing a line of integration code.

This is a practical look at SplatAnything 3D gaussian splatting open weights: what changed technically, how it compares, what it needs to run, and the licence detail that the Hugging Face model card does not make obvious.

TL;DR

SplatAnything is a feed-forward 3D Gaussian Splatting model from Zhejiang University and Tencent Hunyuan.…