Generative Sparse-View Gaussian Splatting

a) Qualitative comparisons with 3 training views: the vanilla 3D/4DGS v.s. ours. 122 a) Qualitative comparisons with 3 training views: the vanilla 3D/4DGS v.s. ours. b) Comparisons with SOTA methods on the LLFF dataset. Figure 1. Our proposed Generative Sparse-view Gaussian Splatting (GS-GS) achieves high-fidelity quality with only three training views. 1) GS-GS is a general pipeline for static and dynamic scene reconstruction with sparse camera views (left: vanilla GS model, right: ours). 2) Quantitative comparisons with other state-of-the-art methods on the LLFF [22] dataset.