The proposed TexRO method introduces a novel approach for generating realistic textures on 3D geometries. The key innovations include a recursive optimization pipeline and an optimal view selection strategy. The recursive optimization leverages diffusion probabilistic models to synthesize textures at increasing resolutions, focusing on consistent structures and detailed textures. The optimal view selection aims to find the smallest set of viewpoints covering all faces of a mesh, enhancing visual consistency and adaptability.

TexRO employs a depth-driven ControlNet for image restoration to synthesize textures in the adaptive denoising module. The method utilizes PyTorch for most computations and Kaolin for differentiable rendering and texture projection. By setting specific parameters and employing a hybrid Python and C++ implementation, TexRO is capable of generating realistic textures in around one minute. The approach also incorporates prompt augmentation to enhance the performance of stable diffusion models.

An essential aspect of TexRO is the adaptive denoising strategy, which involves interlaced denoising using noise schedulers tailored for overlapping and non-overlapping regions. By gradually decreasing the noise level with increasing steps, the method achieves uniform diffusion across different texture regions. The proposed optimal view selection algorithm aims to find the smallest set of viewpoints covering all faces of a mesh, enhancing the quality and consistency of texture generation.