Research map / 3D vision / Face and shape
GOATex: Geometry & Occlusion-Aware Texturing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces GOATex, a training-free, diffusion-based 3D mesh texturing method that uses ray-cast hit levels and normal flipping with backface culling to progressively texture occluded interior surfaces, merging per-layer textures via soft UV blending.
Problem
Existing text-to-texture methods only texture visible exterior surfaces and produce incomplete textures and seams on occluded interior mesh regions.
Approach
Partition mesh faces into ordered visibility layers via multi-view ray casting (hit levels), progressively expose inner layers using residual face clustering plus normal flipping with backface culling, texture each layer with a pretrained depth-ControlNet MVD pipeline, and blend layers in UV space by view-dependent visibility confidence.
Key results
- Ablation: full GOATex reaches 91.16% average GPT win rate over SyncMVD, up from 75.68% with hit-level assignment alone.
- GPT-4.1 and GPT-o3 agree more with human raters (Pearson r 0.43 and 0.34) than GPT-4o-mini (0.22).
- Human and GPT preference tests show strong preference for GOATex over baselines on interior texturing, though GPT judges rate its advantage lower versus TEXTure and SyncMVD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add hit level assignment (ablation) | SyncMVD · A/B preference vs SyncMVD (GPT-based evaluators) | win rate (%): 75.68 | Helped |
| add superface construction (ablation) | SyncMVD with hit level assignment · A/B preference vs SyncMVD (GPT-based evaluators) | win rate (%): 80.27 | Helped |
| add soft UV merging (ablation) | SyncMVD with hit level assignment and superface construction · A/B preference vs SyncMVD (GPT-based evaluators) | win rate (%): 86.49 | Helped |
| add residual face clustering without normal flipping and backface culling (ablation) | SyncMVD with previously added GOATex components · A/B preference vs SyncMVD (GPT-based evaluators) | win rate (%): 86.49 -> 81.17 | Hurt |
| add normal flipping and backface culling (ablation) | SyncMVD with the cumulatively added GOATex components · A/B preference vs SyncMVD (GPT-based evaluators) | win rate (%): 81.17 -> 91.16 | Helped |
Surprising findings
- GPT judges favored baselines that leave interiors unpainted or extend exterior textures inward, suggesting GPTs weight surface smoothness over interior completeness.
- Residual face clustering alone slightly lowered win rate versus the previous configuration (81.17 vs 86.49 average) and only helped once normal flipping was added.
Limitations
- Relies on Stable Diffusion 1.5 backbone, which can reduce prompt adherence.
- Hit-level assignment is purely geometric and may split semantically unified regions in complex geometry such as thin openings or nested cavities.
- Evaluation relies on a curated set of 226 meshes from 12 categories selected for interior geometry.
Experimental setup
Datasets: Objaverse Objaverse-XL
Benchmarks: 226 curated Objaverse/Objaverse-XL meshes with interior geometry
Metrics: human preference win rate GPT A/B preference win rate Pearson correlation Cohen's kappa
Baselines: TEXTure SyncMVD Paint3D TEXGen
Builds on: SyncMVD TEXTure Text2Tex ControlNet
Closest papers by meaning
- FruitNinja: 3D Object Interior Texture Generation with Gaussian Splatting (CVPR 2025 · similarity 0.87)
- MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (CVPR 2025 · similarity 0.86)
- Single Mesh Diffusion Models with Field Latents for Texture Generation (CVPR 2024 · similarity 0.86)
- DreamCraft3D: Hierarchical 3D Generation with Bootstrapped Diffusion Prior (ICLR 2024 · similarity 0.85)
- 3DTopia-XL: Scaling High-quality 3D Asset Generation via Primitive Diffusion (CVPR 2025 · similarity 0.85)
- Volumetric Surfaces: Representing Fuzzy Geometries with Layered Meshes (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.