Texture Space Material Diffusion

arXiv cs.GR details Texture Space Material Diffusion, a material-generation approach that keeps the diffusion process in texture space instead of fighting geometry in image space. The authors fine-tune a video diffusion transformer for text-guided material creation, multi-view generation, and upscaling, then lean on the known projection from image space to texture space to make the system work across arbitrary meshes and UV layouts.

That matters because view consistency is one of the persistent pain points in multi-view and video-based material workflows. By moving the problem into a 2D texture domain, the method can reuse pretrained video diffusion priors while avoiding many of the alignment issues that show up when synthesizing materials from multiple angles.

The paper also targets practical production scenarios: reconstructing high-quality materials from posed photos captured under unknown lighting, generating materials from text or images, and scaling to 8K output with 100+ input views. The authors report state-of-the-art results in both quantitative and qualitative evaluations, which makes this especially relevant for teams building capture, look-dev, or AI-assisted asset pipelines.

For game developers, the interesting part is less the novelty of diffusion itself and more the workflow implication: a texture-space representation could make material generation more robust for real assets, not just curated demos. If the approach holds up outside the paper, it could reduce cleanup time for artists and technical artists while giving tools programmers a clearer path to...