A shape-specific diffusion model for filling missing 3D textures achieved the best overall score in three of four reported metrics, according to an arXiv preprint. The system, GLOSS, completes texture in the local render space of one mesh, using local geometry and paired geometry-texture references. It also supports interactive filling with reference patches, including inpainting and blending of different textures. But the paper identifies limited scalability and cross-shape generalization as major limitations.
A model built around local surface patterns
The training setup uses paired local geometry and texture patches generated automatically with off-the-shelf image and language models conditioned on mesh geometry. That setup targets the relationship between what a surface looks like and the local shape around it, which the study examines for texture completion.
Self-attention is the model's way of comparing patches with one another. GLOSS extends it across all patches in a batch, a design intended to capture long-range self-similarity across the mesh.
The comparison covered 10 base meshes
Automatic evaluation covered from-scratch and fine-tuned settings on 10 base meshes, alongside transfer tests on nine other meshes. The four reported measures were LPIPS, FID, DreamSim and CMMD, with lower values treated as better in the reported scoring.
On the per-mesh setting, GLOSS reported 0.302 for LPIPS, 80.258 for FID, 0.344 for DreamSim and 0.478 for CMMD. In the fine-tuned setting, the figures were 0.282, 104.108, 0.352 and 0.694, respectively. The authors summarize the comparison by saying GLOSS had the best overall score in three of the four metrics.
The design choices were visible in the failures
Ablation tests, which remove parts of a system to see what changes, examined the main design choices. Without multi-attention, the report describes complete texture-generation failure; without geometry conditioning, it describes outputs inconsistent with the mesh's geometry.
On transfer, the authors describe GLOSS as competitive on nine other meshes with shorter fine-tuning steps. The result broadens the evaluation beyond the 10 base meshes, but it does not remove the paper's stated limitation that scalability and cross-shape generalization remain limited.
The model was also tried on material channels
Although it was trained only for albedo, the base-color part of a texture, GLOSS was preliminarily applied to metallic and roughness channels used in physically based rendering, or PBR. The authors describe this extension as preliminary because the material channels were not jointly trained, so consistency across channels is not guaranteed.
A small Blender pilot offered early feedback
The practical test involved five participants with prior 3D-texturing experience and general familiarity with Blender. All acknowledged the iterative workflow, four of five reported that it would make their work more efficient, and all reported that they would integrate the feature into their workflow.
Those responses are early user feedback, not a measured productivity result: participants reported expected efficiency gains and intended integration. The pilot therefore cannot establish how the tool would perform for a broader population of 3D artists.
Reported processing time was 3.4 seconds for 1K texture painting and 5.3 seconds for 4K textures. The timings were unoptimized and excluded communication between the GLOSS server and Blender add-on, so they do not demonstrate complete end-to-end real-time performance.
The open boundary is generalization
The central boundary is the link between geometry and appearance. The method is not applicable when local geometry-texture correlation is weak or when appearance is primarily semantic, and its PBR results do not guarantee consistency across material channels. Taken together, the preprint presents GLOSS as a controllable local texture-completion method with the best overall result in three of four reported metrics, while leaving broader shape generalization and jointly consistent material generation unresolved.
Paper data and sources
Original title: GLOSS: Geometric Local Self-Similarity Learning for Faithful Reference-Guided Texture Fill
Authors: Chenyue Cai, Anita Hu, James Lucas et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text