DANTE-W: Diffuse Albedo Neural Texturing in the Wild

DANTE-W is a neural texturing approach for recovering diffuse albedo from real-world image collections, rather than blending photos directly into a texture atlas. That matters because conventional mesh texturing tends to bake in lighting, shadows, and other capture artifacts that become obvious once you try to relight the asset in-engine.

The paper says the method plugs into traditional 3D reconstruction pipelines: you start with a reconstructed mesh and its surface parameterization, then DANTE-W fuses view-space generative albedo priors into a coherent texture space. It also uses physically principled neural rendering to recover finer texture detail, which is the part that should interest anyone trying to keep scan fidelity without inheriting the original lighting conditions.

For game teams, the practical angle is cleaner scan-to-game asset creation. If the method holds up outside the paper, it could reduce cleanup time for environment artists and technical artists working on scanned props, modular kits, or large outdoor scenes where baked shadows are especially painful. It also sounds relevant for relightable content pipelines, where you want the texture to represent surface color rather than a specific moment in time.

The authors also mention a benchmark dataset with both real in-the-wild scenes and synthetic objects, designed to test fine-grained textures. The paper was submitted to arXiv on 26 Jun 2026, and the reported experiments claim improved albedo reconstruction and relighting fidelity. As usual with research like this, the big question is how robust it is on...