DeClotH: Decomposable 3D Cloth and Human Body Reconstruction from a Single Image

Hyeongjin Nam1,2, Donghwan Kim1, Jeongtaek Oh1, Kyoung Mu Lee1
1Seoul National University, 2KRAFTON
CVPR 2025
arXiv Paper Video

Given a single image, DeClotH reconstructs 3D cloth and human body based on 3D template models. The reconstructed 3D clothes are decomposable from the human body, allowing cloth transfer to a new human avatar.

Abstract

We propose DeClotH, which reconstructs both 3D cloth and human body from a single image.

Unlike recent 3D human reconstruction methods that consider cloth and human body as one unified object, we aim to reconstruct 3D cloth and human body, while being able to separate them. This is a much more challenging problem because cloth and human body are heavily occluded by each other, making it difficult to infer their overall geometry and texture. To address the occlusion issue, there are two core designs in our framework. First, to alleviate the occlusion issue, we leverage 3D template models of both cloth and human body as regularizations, which provide strong priors and prevent erroneous reconstruction by the occlusion. Second, we introduce a cloth diffusion model specifically designed to provide contextual information about cloth appearance, thereby enhancing the reconstruction of 3D cloth. Qualitative and quantitative experiments demonstrate that our proposed approach is highly effective in reconstructing both 3D cloth and the human body.

Intro video

Reconstruction examples

DeClotH produces much accurate 3D human body and cloth reconstruction from a in-the-wild images. Our proposed template-based optimization pipeline effectively infer geometry and texture in the invisible regions of 3D cloth and human body.

3D cloth transfer

We can transfer the reconstructed 3D clothes into a new 3D avatar, by fitting 3D clothes based on the SMPL(+H) human model.

Pose deformation

Additionally, we can deform the reconstructed 3D clothes to a new pose sequence.

Cloth image generation (ClothDiffusion)

The proposed ClothDiffusion is specifically trained to generate only cloth images, ClothDiffusion is highly beneficial for reconstructing 3D cloth, by providing prior image knowledge specialized for cloth geometry and texture. Additionally, ClothDiffusion can be controlled by incorporating 3D template models as regional information for guidance.

Method

Given an input image, DeClotH optimizes 3D cloth and human body, represented by DMTets. For the optimization, we extract normal map, silhouette, and 3D template meshes. Subsequently, the 3D cloth and human body are optimized by three core loss functions: template regularization loss, cloth SDS loss, and human SDS loss.
Algorithm description of DeClotH

Related links

ClothWild: 3D Clothed Human Reconstruction in the Wild

PIXIE: Collaborative Regression of Expressive Bodies

TeCH: Text-guided Reconstruction of Lifelike Clothed Humans

GALA: Generating Animatable Layered Assets from a Single Scan

BibTeX

@article{nam2025decloth,
      author    = {Nam, Hyeongjin and Kim, Donghwan and Oh, Jeongtaek and Lee, Kyoung Mu},
      title     = {{DeClotH}: Decomposable 3D Cloth and Human Body Reconstruction from a Single Image},
      journal   = {CVPR},
      year      = {2025},
    }