Research map / Point clouds / Shape and surface completion
TexOct: Generating Textures of 3D Models with Octree-based 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.
Proposes TexOct, an octree-based DDPM diffusion model that generates dense textures directly on 3D point clouds sampled from ShapeNet meshes, supporting unconditional and text/image-conditional generation.
Problem
Generating complete, high-quality textures directly on 3D model surfaces avoids UV-mapping variability and self-occlusion errors of multi-view methods, but dense point clouds are memory-prohibitive for prior point-based texture generation.
Approach
Represent a densely sampled textured point cloud (100K points) as an octree and run a U-Net of octree convolutions as a DDPM denoiser over the octree node texture values, with octree-based cross-attention for CLIP text/image conditioning.
Key results
- Unconditional generation on ShapeNet: average FID 14.75 and KID 0.13 (x10^2), beating Point-UV 2-stage (FID 17.31) and Text2Tex*(FID 53.22) under the paper's protocol.
- User study: TexOct textures preferred over Point-UV in 80.7% of 2000 responses.
- Octree depth 12 with 100K points selected as best trade-off between training time, reconstruction error, and FID/KID on bench.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TexOct instead of Point-UV (2-Stage) | 3D texture generation · ShapeNet average | FID: 17.31 -> 14.75; average improvement of 2.56 | Helped |
| use TexOct instead of Point-UV (2-Stage) | 3D texture generation · ShapeNet average | KID: 0.30 -> 0.13; average improvement of 0.17 | Helped |
| increase octree depth from 10 to 11 (ablation) | TexOct · ShapeNet bench | FID: 26.06 -> 23.21 | Helped |
| increase octree depth from 10 to 11 (ablation) | TexOct · ShapeNet bench | KID: 0.30 -> 0.21 | Helped |
| increase octree depth from 11 to 12 (ablation) | TexOct · ShapeNet bench | FID: 23.21 -> 23.67 | Hurt |
| increase octree depth from 11 to 12 (ablation) | TexOct · ShapeNet bench | KID: 0.21 -> 0.13 | Helped |
| increase octree depth from 12 to 13 (ablation) | TexOct · ShapeNet bench | FID: 23.67 -> 28.19 | Hurt |
| increase octree depth from 12 to 13 (ablation) | TexOct · ShapeNet bench | KID: 0.13 -> 0.48 | Hurt |
| increase octree depth from 10 to 13 (ablation) | TexOct · ShapeNet bench | training time (second/epoch): 5 -> 57 | Hurt |
| increase octree depth from 10 to 13 (ablation) | TexOct · ShapeNet bench | reconstruction error: 0.020 -> 0.002 | Helped |
| increase sampled point count from 10K toward 200K (ablation) | TexOct · ShapeNet bench | FID | Helped |
| increase sampled point count from 10K toward 200K (ablation) | TexOct · ShapeNet bench | KID | Helped |
Surprising findings
- FID and KID first decrease then increase with octree depth, suggesting over-fitting at very deep octrees.
Limitations
- Uses only xyz and color; additional geometric cues such as normals or curvature are not exploited.
- Evaluated only on four ShapeNet categories, with Text2Tex evaluated on a reduced sample subset.
- Conditional generation limited to chair and table categories.
Experimental setup
Datasets: ShapeNet text-shape pairs from Chen et al. (ref 9)
Benchmarks: ShapeNet
Metrics: FID KID user preference rate reconstruction error
Baselines: Texture Fields Texturify Point-UV Text2Tex
Builds on: DDPM OCNN Point-UV CLIP
Closest papers by meaning
- UV-free Texture Generation with Denoising and Geodesic Heat Diffusion (NeurIPS 2024 · similarity 0.88)
- TexGaussian: Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Single Mesh Diffusion Models with Field Latents for Texture Generation (CVPR 2024 · similarity 0.87)
- TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process (CVPR 2024 · similarity 0.87)
- Sin3DM: Learning a Diffusion Model from a Single 3D Textured Shape (ICLR 2024 · similarity 0.86)
- TexGarment: Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer (CVPR 2025 · similarity 0.86)
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.