Research map / 3D vision / Face and shape
DeClotH: Decomposable 3D Cloth and Human Body Reconstruction from a Single Image
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.
Presents DeClotH, a template-based optimization framework that reconstructs separate 3D cloth and human body meshes from a single image, using 3D template regularization and a new ClothDiffusion model for cloth-specific SDS guidance.
Problem
Existing 3D clothed human reconstruction from a single image treats cloth and body as one unified mesh, and it is hard to decompose them under severe cloth-body occlusion; vanilla SDS guidance from Stable Diffusion mixes cloth and body content.
Approach
Use 3D cloth and body template meshes (ClothWild/ClothNet, PIXIE/BodyNet) as silhouette-based regularizers to handle occlusion, and fine-tune a ControlNet-style Stable Diffusion (ClothDiffusion) conditioned on template silhouettes and skeletons to generate cloth-only images for SDS supervision.
Key results
- On 4D-DRESS (cloth+human), DeClotH reaches CD 3.292 and NC 0.079, versus 3.334 for the best baseline TeCH + GALA.
- On 4D-DRESS (cloth), DeClotH reaches CD 3.902 and PSNR 31.582, versus CD 5.043 and PSNR 29.123 for TeCH + GALA.
- ClothDiffusion achieves CLIP-Score 0.713 and Cloth-IoU 0.548, versus 0.696 and 0.112 for Stable Diffusion.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DeClotH instead of TeCH + GALA | 3D cloth reconstruction · 4D-DRESS (cloth) | CD: 5.043 -> 3.902 | Helped |
| use DeClotH instead of TeCH + GALA | 3D cloth and human reconstruction · 4D-DRESS (cloth + human) | CD: 3.334 -> 3.292 | Helped |
| use DeClotH instead of TeCH + GALA | 3D cloth reconstruction · THuman2.0 (cloth) | CD: 2.112 -> 1.756 | Helped |
| use DeClotH instead of SiTH + GALA | 3D cloth and human reconstruction · THuman2.0 (cloth + human) | CD: 1.828 -> 1.812 | Helped |
| add cloth and body template regularization losses (ablation) | DeClotH · 4D-DRESS (cloth) | CD: 8.245 -> 3.902 | Helped |
| add cloth and body template regularization losses (ablation) | DeClotH · 4D-DRESS (cloth + human) | CD: 3.880 -> 3.292 | Helped |
| use cloth and human SDS losses instead of no SDS losses (ablation) | DeClotH · 4D-DRESS (cloth) | CD: 4.704 -> 3.902 | Helped |
| replace vanilla StableDiffusion SDS with the cloth and human SDS losses (ablation) | DeClotH · 4D-DRESS (cloth) | CD: 4.922 -> 3.902 | Helped |
| use ClothDiffusion instead of StableDiffusion (ablation) | cloth image generation · 4D-DRESS (cloth) | Cloth-IoU: 0.112 -> 0.548 | Helped |
Surprising findings
- Vanilla Stable Diffusion produces images with mixed cloth and body content, yielding a Cloth-IoU of only 0.112 on 4D-DRESS.
Limitations
- Two-stage baselines propagate 3D errors, while DeClotH's one-stage optimization is per-image and relies on template quality from ClothNet and BodyNet.
- Template-based methods are constrained by pre-defined topology, so the optimization must add fine details on top of templates.
Experimental setup
Datasets: DeepFashion SHHQ MSCOCO THuman2.0 4D-DRESS
Benchmarks: 4D-DRESS THuman2.0
Metrics: CD NC PSNR LPIPS CLIP-Score Cloth-IoU
Baselines: BCNet SMPLicit ClothWild + PIXIE GALA* SiTH + GALA TeCH + GALA
Builds on: DreamFusion ControlNet DMTet ECON GALA
Closest papers by meaning
- SiTH: Single-view Textured Human Reconstruction with Image-Conditioned Diffusion (CVPR 2024 · similarity 0.89)
- Diffusion-FOF: Single-View Clothed Human Reconstruction via Diffusion-Based Fourier Occupancy Field (CVPR 2024 · similarity 0.89)
- R-Cyclic Diffuser: Reductive and Cyclic Latent Diffusion for 3D Clothed Human Digitalization (CVPR 2024 · similarity 0.88)
- D^3-Human: Dynamic Disentangled Digital Human from Monocular Video (CVPR 2025 · similarity 0.88)
- VS: Reconstructing Clothed 3D Human from Single Image via Vertex Shift (CVPR 2024 · similarity 0.87)
- MultiGO: Towards Multi-level Geometry Learning for Monocular 3D Textured Human Reconstruction (CVPR 2025 · 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.