GTR: Improving Large 3D Reconstruction Models through Geometry and Texture Refinement

Comparison with the baseline methods on the sparse-view reconstruction task. We visualize the novel view results generated from the GSO dataset (Downs et al., 2022) (row 1-3), or a generated image (row 4). Note that for the generated image, we use Zero123++ (Shi et al., 2023) to generate 6 views, which are used as input for InstantMesh (Xu et al., 2024a) and GTR (Ours).