Research map / 3D vision / Face and shape
TUVF: Learning Generalizable Texture UV Radiance Fields
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.
Proposes Texture UV Radiance Fields (TUVF), which generate category-level textures on a learnable UV sphere shared across shapes and render them via a point-based radiance field trained adversarially from single-view images.
Problem
Generating high-fidelity 3D textures on given shapes that are controllable and disentangled from geometry, so one texture code gives a consistent appearance style across shapes of a category.
Approach
Generate textures on a canonical UV sphere learned via a canonical surface auto-encoder that gives dense correspondence across instances, then render a point-based radiance field (Poisson-reconstructed indicator as density) with a patch-based GAN trained only on single-view images.
Key results
- On CompCars, TUVF reaches FID 41.79 and KID 2.95 versus 59.55 / 4.97 for Texturify, with LPIPS_g 15.87 vs 9.75.
- On DiffusionCats, TUVF achieves FID 64.13 vs 72.43 for Texturify and LPIPS_t 1.50 vs 2.61 (better transfer consistency).
- On Photoshape, TUVF has slightly higher FID (51.29 vs 45.92 for Texturify) but much higher LPIPS_g (14.93 vs 6.74); user study favored TUVF over Texturify 70-82% of the time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TUVF instead of Texturify | CompCars texture synthesis · CompCars | LPIPS_g: 9.75 -> 15.87 | Helped |
| use TUVF instead of Text2Tex | CompCars texture synthesis · CompCars | FID: 46.91 -> 41.79 | Helped |
| use TUVF instead of Texturify | Photoshape texture synthesis · Photoshape | FID: 45.92 -> 51.29 | Hurt |
| use TUVF instead of Texturify | Photoshape texture synthesis · Photoshape | LPIPS_g: 6.74 -> 14.93 | Helped |
| use TUVF instead of Texturify | DiffusionCats texture synthesis · DiffusionCats | FID: 72.43 -> 64.13 | Helped |
| replace CIPS-UV with CIPS-2D texture mapping (ablation) | TUVF texture generator on CompCars · CompCars | FID: 41.79 -> 148.09 | Hurt |
| replace CIPS-UV with StyleGAN2 texture mapping (ablation) | TUVF texture generator on CompCars · CompCars | FID: 41.79 -> 103.62 | Hurt |
| reduce UV resolution from 2K to 1K (ablation) | TUVF on CompCars · CompCars | FID: 41.79 -> 43.65 | Hurt |
| remove surface normal decoder g_theta and use ground-truth indicator function (ablation) | Canonical Surface Auto-encoder in TUVF · CompCars | FID: 41.79 -> 61.81 | Hurt |
| replace smooth learned correspondence with nearest-neighbor proxy surface points (ablation) | Canonical Surface Auto-encoder in TUVF · CompCars | FID: 41.79 -> 79.43 | Hurt |
| reverse mapping direction from UV-to-surface to Surface-to-UV (ablation) | Canonical Surface Auto-encoder in TUVF · CompCars | FID: 41.79 -> 139.19 | Hurt |
Surprising findings
- Texturify achieves lower FID/KID than TUVF on Photoshape and full Photoshape, a case where the proposed method is slightly worse on fidelity metrics.
Limitations
- Assumes one-to-one dense correspondence between instances, which fails for structurally varied categories such as diverse chair topologies.
- UV sphere constrains representable topologies; results on non-genus-zero shapes are less reliable.
- Photorealism limited relative to data-driven priors such as diffusion models, as noted by the authors.
Experimental setup
Datasets: ShapeNet CompCars Photoshape SMAL DFAUST ShapeNet Airplanes
Benchmarks: CompCars Photoshape DiffusionCats
Metrics: LPIPS_g LPIPS_t FID KID
Baselines: TexFields LTG EG3D-Mesh Text2Tex Texturify EpiGRAF CIPS-2D StyleGAN2
Builds on: EpiGRAF CIPS VolSDF Texturify DGCNN
Closest papers by meaning
- AUV-Net: Learning Aligned UV Maps for Texture Transfer and Synthesis (CVPR 2022 · similarity 0.90)
- TextureDreamer: Image-Guided Texture Synthesis Through Geometry-Aware Diffusion (CVPR 2024 · similarity 0.86)
- Paint3D: Paint Anything 3D with Lighting-Less Texture Diffusion Models (CVPR 2024 · similarity 0.86)
- LAENeRF: Local Appearance Editing for Neural Radiance Fields (CVPR 2024 · similarity 0.86)
- UDiFF: Generating Conditional Unsigned Distance Fields with Optimal Wavelet Diffusion (CVPR 2024 · similarity 0.85)
- TriTex: Learning Texture from a Single Mesh via Triplane Semantic Features (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.