QuerySplat Decouples Geometry and Appearance in Feed-Forward 3D Gaussian Splatting

Researchers from Beihang University, InSpatio Research, and Zhejiang University have proposed QuerySplat, a feed-forward 3D Gaussian Splatting (3DGS) framework that separates geometric and appearance representations through a dual-branch query-based decoder. The geometry branch draws on a pretrained Vision Geometric Model for spatial understanding and pose-free modeling, while the appearance branch independently recovers high-frequency detail — addressing the blurry, pose-dependent outputs that have plagued earlier query-based approaches.

On the DL3DV benchmark, QuerySplat achieves state-of-the-art novel view synthesis results, recording average PSNR gains of 2.30 dB over the best pose-free baseline and 1.04 dB over the best pose-required baseline. The work is available as an arXiv preprint alongside code and model releases.