Introduction

We propose StructLDM, a structured latent diffusion model that learns 3D human generations from 2D images.

StructLDM generates diverse view-consistent humans, and supports different levels of controllable generation and editing, such as compositional generations by blending the five selected parts from a), and part-aware editing such as identity swapping, local clothing editing, 3D virtual try-on, etc. Note that the generations and editing are clothing-agnostic without clothing types or masks conditioning.

Generations on RenderPeople.

Installation

NVIDIA GPUs are required for this project. We have trained and tested code on NVIDIA V100. We recommend using anaconda to manage the python environments.

conda create --name structldm python=3.9
conda install pytorch==1.10.1 torchvision==0.11.2 cudatoolkit=11.1 -c pytorch
conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install pytorch3d -c pytorch3d
pip install -r requirements.txt

Implementation

Download Models & Assets & Datasets

Download sample data, necessary assets, and pretrained models from OneDrive. Put them in DATA_DIR/result/trained_model and DATA_DIR/asset respectively. DATA_DIR is specified as ./data in default.

Register and download SMPL models here. Put them in the folder smpl_data.

The folder structure should look like

DATA_DIR
├── dataset
    ├──renderpeople/
└── asset/
    ├── smpl_data/
        └── SMPL_NEUTRAL.pkl
    ├── uv_sampler/
    ├── uv_table.npy
    ├── smpl_uv.obj
    ├── smpl_template_sdf.npy
    ├── sample_data.pkl
├── result/
    ├── trained_model/modelname/
        └──decoder_xx, diffusion_xx
        ├──samples/
    ├── test_output

Commands

Generating 3D humans via (e.g., models trained on RenderPeople).

bash scripts/renderpeople.sh gpu_ids

The generation results will be found in DATA_DIR/result/test_output.

The training script of latent diffusion can be found in struct_diffusion.

bash struct_diffusion/scripts/exec.sh "train" gpu_ids

Trained models will be stored in DATA_DIR/result/trained_model/modelname/diffusion_xx.pt.

Refer to the downloaded sample data at ./data/dataset/renderpeople to prepare your own dataset, and modify the corresponding path in the config file.

The inference script of latent diffusion can be found in struct_diffusion.

bash struct_diffusion/scripts/test.sh gpu_ids

Samples will be stored in DATA_DIR/result/trained_model/modelname/samples.

License

Distributed under the S-Lab License. See LICENSE for more information.

SMPL-X related files are subject to the license of SMPL-X.

Citation

If you find our code or paper is useful to your research, please consider citing:

@misc{hu2024structldm,
      title={StructLDM: Structured Latent Diffusion for 3D Human Generation}, 
      author={Tao Hu and Fangzhou Hong and Ziwei Liu},
      year={2024},
      eprint={2404.01241},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
  }

Acknowledgements

The structured diffusion model is implemented on top of the Latent-Diffusion.