Abstract
We present MatAtlas, a method for consistent text-guided 3D model texturing.Following recent progress we leverage a large scale text-to-image generationmodel (e.g., Stable Diffusion) as a prior to texture a 3D model. We carefullydesign an RGB texturing pipeline that leverages a grid pattern diffusion,driven by depth and edges. By proposing a multi-step texture refinementprocess, we significantly improve the quality and 3D consistency of thetexturing output. To further address the problem of baked-in lighting, we movebeyond RGB colors and pursue assigning parametric materials to the assets.Given the high-quality initial RGB texture, we propose a novel materialretrieval method capitalized on Large Language Models (LLM), enablingeditabiliy and relightability. We evaluate our method on a wide variety ofgeometries and show that our method significantly outperform prior arts. Wealso analyze the role of each component through a detailed ablation study.