Research map / Image generation / Autoregressive and latent generation
UDiFF: Generating Conditional Unsigned Distance Fields with Optimal Wavelet 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 UDiFF, a 3D diffusion model that generates textured shapes with open and closed surfaces by diffusing learned-optimal wavelet coefficients of unsigned distance fields.
Problem
Existing 3D implicit diffusion models use SDF or occupancy fields, which only represent closed shapes, and wavelet-based compression of fields loses information when wavelet types are hand-picked.
Approach
Learn a biorthogonal wavelet filter (decomposition and inversion) end-to-end by minimizing near-surface UDF reconstruction error, then train a text-conditional diffusion model on the resulting coarse coefficient volumes with a fine predictor for detail, enabling UDF generation of open surfaces.
Key results
- On DeepFashion3D, UDiFF reaches COV 69.62 (CD) and 1-NNA 81.83 (CD), beating PointDiff, WaveGen, Diffusion-SDF and LAS-Diffusion.
- Learned wavelet filters reduce L2 Chamfer Distance to 28.51 (x1e-5) when both filters are optimized, versus 42.92 for Biorthogonal 6-8 and 264.8 for Haar.
- On ShapeNet chairs and airplanes, UDiFF is comparable to SALAD and outperforms WaveGen on most metrics (e.g., airplane MMD-CD 3.151 vs 3.528).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UDiFF instead of PointDiff | 3D shape generation · DeepFashion3D | COV-EMD: 64.56 -> 67.72 | Helped |
| use UDiFF instead of PointDiff | 3D shape generation · DeepFashion3D | MMD-CD: 11.01 -> 11.60 | Hurt |
| use UDiFF instead of SALAD | 3D shape generation · ShapeNet Chair | COV-CD: 56.42 -> 52.58 | Hurt |
| use UDiFF instead of SALAD | 3D shape generation · ShapeNet Chair | MMD-CD: 11.69 -> 11.67 | No clear effect |
| use UDiFF instead of SALAD | 3D shape generation · ShapeNet Airplane | MMD-CD: 3.636 -> 3.151 | Helped |
| use UDiFF instead of SALAD | 3D shape generation · ShapeNet Airplane | 1-NNA-EMD: 71.08 -> 78.99 | Hurt |
| add the learned wavelet instead of using Biorthogonal 6-8 (ablation) | UDiFF · DeepFashion3D | COV-CD: 64.52 -> 69.62 | Helped |
| add the fine predictor (ablation) | UDiFF · DeepFashion3D | COV-CD: 66.36 -> 69.62 | Helped |
| add the fine predictor (ablation) | UDiFF · DeepFashion3D | MMD-CD: 12.37 -> 11.60 | Helped |
| optimize both decomposition and inversion wavelet filters (ablation) | UDF wavelet reconstruction · DeepFashion3D | L2 Chamfer Distance: Haar 264.8; Biorthogonal3-3 46.04; Biorthogonal6-8 42.92; Both 28.51 | Helped |
Limitations
- Relies on a small DeepFashion3D dataset of 1,798 models for open-surface training.
- Text conditioning on ShapeNet is shown only qualitatively against AutoSDF.
Experimental setup
Datasets: DeepFashion3D ShapeNet
Benchmarks: DeepFashion3D ShapeNet
Metrics: MMD-CD MMD-EMD COV 1-NNA Chamfer Distance
Baselines: PointDiff WaveGen Diffusion-SDF LAS-Diffusion Shap·E IM-GAN Voxel-GAN SPAGHETTI SALAD AutoSDF
Builds on: WaveGen DDPM Stable Diffusion Text2Tex DCUDF
Closest papers by meaning
- DUDF: Differentiable Unsigned Distance Fields with Hyperbolic Scaling (CVPR 2024 · similarity 0.88)
- High Resolution UDF Meshing via Iterative Networks (NeurIPS 2025 · similarity 0.86)
- Text-to-3D Generation with Bidirectional Diffusion using both 2D and 3D priors (CVPR 2024 · similarity 0.86)
- Single Mesh Diffusion Models with Field Latents for Texture Generation (CVPR 2024 · similarity 0.86)
- Sin3DM: Learning a Diffusion Model from a Single 3D Textured Shape (ICLR 2024 · similarity 0.86)
- TUVF: Learning Generalizable Texture UV Radiance Fields (ICLR 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.