TriTex: Learning Texture from a Single Mesh via Triplane Semantic Features

TL;DR

TriTex tackles single-exemplar texture transfer by learning a volumetric texture field encoded as a triplane of semantic features. It leverages pre-trained 3D semantic descriptors from Diff3F, projected onto a triplane and refined with triplane-aware convolutions to produce per-point colors via a color MLP, enabling fast, feed-forward texture transfer that preserves source semantics across related shapes. Evaluations on a curated dataset show TriTex outperforms optimization-free baselines in both automated metrics and human preferences, while offering rapid inference suitable for large scenes. The approach enables coherent texture propagation across related 3D models in game development and simulation, with potential extensions using cross-attention and priors to further enhance detail and generalization.

Abstract

As 3D content creation continues to grow, transferring semantic textures between 3D meshes remains a significant challenge in computer graphics. While recent methods leverage text-to-image diffusion models for texturing, they often struggle to preserve the appearance of the source texture during texture transfer. We present \ourmethod, a novel approach that learns a volumetric texture field from a single textured mesh by mapping semantic features to surface colors. Using an efficient triplane-based architecture, our method enables semantic-aware texture transfer to a novel target mesh. Despite training on just one example, it generalizes effectively to diverse shapes within the same category. Extensive evaluation on our newly created benchmark dataset shows that \ourmethod{} achieves superior texture transfer quality and fast inference times compared to existing methods. Our approach advances single-example texture transfer, providing a practical solution for maintaining visual coherence across related 3D models in applications like game development and simulation.

Figures (10)

• Figure 1: User study interface and instructions. Participants were asked to select which result better preserves the style of the source object while adapting to the target shape.
• Figure 2: Training Pipeline (Top). Given an input textured mesh and its pre-extracted Diff3F features, we project six orthographic views to create an initial triplane. This triplane is processed using triplane-aware convolutional blocks, which, along with the texture MLP, define a coloring neural field. This field, together with the input geometry, is used to render the colored mesh. Appearance losses are applied between the true mesh appearance and the rendered appearance. Inference (Bottom). Given a new mesh (left), our pre-trained model maps its semantic properties to colors (right), transferring the texture from the original textured mesh learned during the training phase.
• Figure 3: Texture Transfer Results. Given a source mesh (left column), our network transfers its appearance onto target meshes of varying shapes (three right columns). Each row demonstrates texture transfer using a different source mesh.
• Figure 4: Additional Qualitative Results. Showing the target geometry and the texture transfer.
• Figure 5: Qualitative Comparisons with baselines. A comparison between TriTex (right-most column) and state-of-the-art approaches for texture transfer. As observed, TriTex produces much more plausible texture transfer results.