Research map / Point clouds / Shape and surface completion
UV-free Texture Generation with Denoising and Geodesic Heat Diffusion
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.
Generates UV-free textures as coloured point-clouds on mesh surfaces using a DDPM built from heat-diffusion blocks with a farthest-sampled heat-diffusion attention layer and a mixed robust Laplacian.
Problem
Standard UV-based texturing of meshes suffers from seams, distortions, wasted UV space, vertex duplication and varying resolution, especially with automatic UV unwrapping; existing texture generators either rely on UV images or operate in Euclidean space ignoring mesh topology.
Approach
Represent textures as coloured point-clouds sampled on the mesh surface and denoise them with a DDPM whose U-Net blocks use heat diffusion (DiffusionNet-style) on the surface, plus a novel attention layer applied on farthest-point samples and spread back by heat diffusion, a mixed robust Laplacian combining mesh and point-cloud operators, and online Poisson-disk sampling that reuses precomputed sp
Key results
- On ShapeNet chairs, UV3-TeD achieves FID 54.20 and KID 42.17 (x10^-4), beating PointUVDiff (uv) at FID 63.35 and KID 83.19.
- The DiffusionNet DDPM baseline has the highest LPIPS (0.24) but much worse FID (116.58), while UV3-TeD has LPIPS 0.21.
- Ablation on ABO: removing the farthest-sampled attention raises FID from 77.14 to 83.16 and KID from 58.59 to 103.42, the largest drop among ablations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UV3-TeD instead of PointUVDiff (uv) | Texture generation on ShapeNet chairs · ShapeNet chairs test | FID: 63.35 -> 54.20 | Helped |
| Use UV3-TeD instead of PointUVDiff (pcl-texture) | Texture generation on ShapeNet chairs · ShapeNet chairs test | KID × 10⁻⁴: 126.18 -> 42.17 | Helped |
| Use UV3-TeD instead of DiffusionNet DDPM | Texture generation on ShapeNet chairs · ShapeNet chairs test | FID: 116.58 -> 54.20 | Helped |
| Add farthest-sampled attention (ablation) | UV3-TeD · ABO test (100 shapes) | FID: 83.16 -> 77.14 | Helped |
| Add farthest-sampled attention (ablation) | UV3-TeD · ABO test (100 shapes) | KID × 10⁻⁴: 103.42 -> 58.59 | Helped |
| Add farthest-sampled attention (ablation) | UV3-TeD · ABO test (100 shapes) | LPIPS: 0.10 -> 0.14 | Helped |
| Add Λ′ conditioning (ablation) | UV3-TeD · ABO test (100 shapes) | FID: 79.96 -> 77.14 | Helped |
| Add Λ′ conditioning (ablation) | UV3-TeD · ABO test (100 shapes) | LPIPS: 0.15 -> 0.14 | Hurt |
| Add sihks conditioning (ablation) | UV3-TeD · ABO test (100 shapes) | KID × 10⁻⁴: 70.90 -> 58.59 | Helped |
| Use sihks instead of Φp conditioning (ablation) | UV3-TeD · ABO test (100 shapes) | FID: 79.62 -> 77.14 | Helped |
| Use the mixed LBO instead of the mesh LBO (ablation) | UV3-TeD · ABO test (100 shapes) | LPIPS: 0.15 -> 0.14 | Hurt |
| Use the mixed LBO instead of the mesh LBO (ablation) | UV3-TeD · ABO test (100 shapes) | FID: 78.07 -> 77.14 | Helped |
Surprising findings
- Removing the global shape descriptor Lambda' slightly improves diversity (LPIPS 0.15) while lowering sample quality.
- Point-UV Diffusion's texture quality drops when its point-clouds are rendered with the proposed method versus UV projection, with mixed results across metrics (FID 65.09 point-cloud vs 63.35 UV).
Limitations
- Learning high-frequency texture details requires thousands of epochs and the current architecture has no pooling, limiting texture resolution.
- Generated textures can show blotchy patterns, part-recognition errors and long-range inconsistencies.
- Only albedo colours are generated; BRDF generation is left to future work.
Experimental setup
Datasets: ShapeNet chairs Amazon Berkeley Objects (ABO) CelebA
Benchmarks: ShapeNet chairs ABO
Metrics: FID KID LPIPS
Baselines: PointUVDiff (uv) PointUVDiff (pcl-texture) DiffusionNet DDPM
Builds on: DDPM DiffusionNet Point-UV Diffusion Texturify Mitsuba3
Closest papers by meaning
- Single Mesh Diffusion Models with Field Latents for Texture Generation (CVPR 2024 · similarity 0.89)
- Paint3D: Paint Anything 3D with Lighting-Less Texture Diffusion Models (CVPR 2024 · similarity 0.88)
- TexOct: Generating Textures of 3D Models with Octree-based Diffusion (CVPR 2024 · similarity 0.88)
- MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (CVPR 2025 · similarity 0.87)
- Material Anything: Generating Materials for Any 3D Object via Diffusion (CVPR 2025 · similarity 0.86)
- One-2-3-45++: Fast Single Image to 3D Objects with Consistent Multi-View Generation and 3D Diffusion (CVPR 2024 · 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.