Tex-Zero trains on roughly 11.1 million 2D images and generates textures on real 3D assets, beating baselines trained on 3D data
Tex-Zero introduces a data-construction pipeline that treats each 2D image as a plane in 3D space and applies patch-wise random rotations and aggregation, training a native 3D texture generation VAE and DiT without any real 3D assets and achieving high-fidelity texture reconstruction and generation on real 3D assets, outperforming baselines such as NaTex and TRELLIS.2 in six-view and front-view evaluations.