Research map / 3D vision / Face and shape
Diffusion-FOF: Single-View Clothed Human Reconstruction via Diffusion-Based Fourier Occupancy Field
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 Diffusion-FOF, which predicts a back-view image with a style-consistency loss and then generates a Fourier occupancy field in the Haar wavelet domain with a conditional diffusion model plus image-guided refinement, for single-view clothed human reconstruction.
Problem
Single-view clothed human reconstruction struggles with flexible representation of body shapes and poses, complete 3D geometry, consistent texture, and fine-grained details; prior FOF methods lose high-frequency information.
Approach
Combines an image-conditioned diffusion model with the Fourier occupancy field representation, generating FOF coefficients in the Haar wavelet domain (sub-bands at 1/4 spatial size), plus a Siamese back-view predictor trained with a VGG Gram-matrix style-consistency loss.
Key results
- On 2K2K, texture PSNR of 26.362 and LPIPS 0.142 versus 18.352 PSNR for GTA on rendered images.
- Geometry Chamfer on THuman2.0 of 2.131 cm, better than GTA (2.298) and ECON (2.312); FOF baseline 2.474.
- Ablation: wavelet diffusion with back-view and refinement (Ours-M3) reduces 2K2K normal error to 1.502 from FOF-M1 at 2.101.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Diffusion-FOF instead of the strongest reported baseline | Single-view clothed human reconstruction · THuman2.0 | Chamfer: 2.298 -> 2.131 | Helped |
| Use Diffusion-FOF instead of the strongest reported baseline | Single-view clothed human reconstruction · 2K2K | Chamfer: 1.129 -> 0.934 | Helped |
| Use Diffusion-FOF instead of the strongest reported baseline | Single-view clothed human reconstruction · CLOTH4D | Chamfer: 1.102 -> 0.906 | Helped |
| Add style consistency loss to L1 loss (ablation) | Back-view image prediction · 2K2K | PSNR: 18.427 -> 24.476 | Helped |
| Add Siamese network training to L1 plus style consistency loss (ablation) | Back-view image prediction · 2K2K | PSNR: 24.476 -> 26.362 | Helped |
| Add predicted back-view image as a condition (ablation) | FOF-M1 · THuman2.0 | Chamfer: 2.474 -> 2.467 | Helped |
| Add FOF refinement module (ablation) | FOF-M2 · THuman2.0 | Chamfer: 2.467 -> 2.462 | Helped |
| Use diffusion model rather than the FOF baseline (ablation) | Geometric prediction · THuman2.0 | Chamfer: 2.474 -> 2.245 | Helped |
| Add predicted back-view image as a diffusion condition (ablation) | Ours-W1 · THuman2.0 | Chamfer: 2.245 -> 2.243 | Helped |
| Use wavelet diffusion rather than diffusion without wavelet transform (ablation) | Geometric prediction · THuman2.0 | Chamfer: 2.245 -> 2.241 | Helped |
| Add predicted back-view image as a wavelet diffusion condition (ablation) | Ours-M1 · THuman2.0 | Chamfer: 2.241 -> 2.138 | Helped |
| Add FOF refinement module (ablation) | Ours-M2 · THuman2.0 | Chamfer: 2.138 -> 2.131 | Helped |
Limitations
- Geometric accuracy may drop under severe human self-occlusion.
- Evaluation relies on rendered scan datasets rather than real in-the-wild ground truth.
Experimental setup
Datasets: THuman2.0 2K2K CLOTH4D
Benchmarks: THuman2.0 2K2K CLOTH4D
Metrics: PSNR SSIM LPIPS Chamfer distance P2S distance normal L1 error
Baselines: SiCloPe PIFu PIFuHD ICON FOF D-IF ECON 2K2K GTA
Builds on: Fourier occupancy field DDPM DDIM PIFu
Closest papers by meaning
- DeClotH: Decomposable 3D Cloth and Human Body Reconstruction from a Single Image (CVPR 2025 · similarity 0.89)
- R-Cyclic Diffuser: Reductive and Cyclic Latent Diffusion for 3D Clothed Human Digitalization (CVPR 2024 · similarity 0.89)
- MultiGO: Towards Multi-level Geometry Learning for Monocular 3D Textured Human Reconstruction (CVPR 2025 · similarity 0.88)
- SIFU: Side-view Conditioned Implicit Function for Real-world Usable Clothed Human Reconstruction (CVPR 2024 · similarity 0.88)
- D^3-Human: Dynamic Disentangled Digital Human from Monocular Video (CVPR 2025 · similarity 0.88)
- SiTH: Single-view Textured Human Reconstruction with Image-Conditioned Diffusion (CVPR 2024 · similarity 0.87)
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.