This installation guide assumes you have Python 3.9+ and recommend using Anaconda to manage the environment.

• Python 3.9+
• Anaconda (recommended)
• NVIDIA GPU with CUDA support (CUDA 11.8+ recommended)
• SMPL model files (download from SMPL website)

Clone the repository with submodules:

If you've already cloned the repository without submodules, initialize them:

git submodule update --init --recursive

conda create -n monoclothrecon python=3.9
conda activate monoclothrecon

Install PyTorch with CUDA support. For CUDA 11.8:

pip install torch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 --index-url https://download.pytorch.org/whl/cu118

Or using conda:

conda install pytorch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 pytorch-cuda=11.8 -c pytorch -c nvidia

First, install the required dependencies:

conda install -c fvcore -c iopath -c conda-forge fvcore iopath

Then install pytorch3d:

pip install "git+https://github.com/facebookresearch/pytorch3d.git"

Build and install nvdiffrast from the submodule:

cd packages/nvdiffrast
pip install .
cd ../..

Install the remaining required Python packages:

pip install trimesh pyyaml easydict tensorboard torchvision numpy scipy

Install the local nvdiffmodeling package:

pip install -e packages/nvdiffmodeling

1. Copy the configuration template:

cp scripts/configs/config_template.yml my_config.yml

2. Edit my_config.yml and update the following paths:
   • smpl_path: Path to SMPL_NEUTRAL.pkl
   • smpl_pkl: Path to SMPL fitting results
   • mesh: Path to template mesh (.obj file)
   • target_images: Path to input images
   • target_diffuse_maps: Path to diffuse maps
   • target_shil_maps: Path to silhouette maps
   • output_path: Path for output directory
   • Other dataset-specific paths

Run the full pipeline (geometry + texture reconstruction):

python main.py --config my_config.yml

Or run specific stages:

# Geometry training only
from scripts.geometry import geometry_training_loop
geometry_training_loop(cfg, device)

# Texture/appearance training
from scripts.appearance import appearance_training_loop
appearance_training_loop(cfg, device)

# Inference
from scripts.inference import Inference
Inference(cfg, texture=True, device=device)

Key configuration parameters in config.yml:

• Model Configuration:

  • model_type: Dataset type (Dress4D or PeopleSnapshot)
  • model: Model architecture (hashgrid, hashgrid_vd, general, diffusion)

• Training Configuration:

  • num_epochs: Number of training epochs
  • warm_ups: Warm-up epochs before deformation
  • remeshing: Enable/disable remeshing

• Texture Configuration:

  • texture_recon: Enable texture reconstruction
  • use_dynamic_texture: Use dynamic texture model
  • texture_map: Texture map resolution [width, height]

See scripts/configs/config_template.yml for all available options.

1. Organize your data in the following structure:

data/
  ├── images/
  ├── masks/
  ├── smpl_fittings.pkl
  └── template_mesh.obj

2. Update the configuration file with your data paths.

Follow similar structure as Dress4D. Ensure SMPL fittings are available for each frame.

MonoClothRecon/
├── main.py                 # Main entry point
├── scripts/
│   ├── geometry.py         # Geometry training loop
│   ├── appearance.py       # Texture/appearance training
│   ├── inference.py        # Inference and evaluation
│   ├── models/            # Model definitions
│   ├── dataloader/        # Data loading utilities
│   ├── losses/            # Loss functions
│   ├── renderer/          # Rendering utilities
│   └── utils/             # Utility functions
├── packages/              # External dependencies (submodules)
│   ├── nvdiffrast/       # NVIDIA differentiable rasterizer
│   ├── nvdiffrec/        # NVIDIA differentiable reconstruction
│   ├── fashion-clip/      # Fashion-CLIP model
│   ├── nvdiffmodeling/   # Local modeling utilities
│   └── fashion_clip/     # Local fashion-clip copy
└── preprocessing/         # Data preprocessing scripts

If you use this code, please cite:

@article{dasgupta2025ngd,
  title={NGD: Neural Gradient Based Deformation for Monocular Garment Reconstruction},
  author={Dasgupta, Soham and Naik, Shanthika and Savalia, Preet and Ingle, Sujay Kumar and Sharma, Avinash},
  journal={arXiv preprint arXiv:2508.17712},
  year={2025}
}

Please check individual package licenses:

• nvdiffrast: NVIDIA Source Code License
• nvdiffrec: NVIDIA Source Code License
• fashion-clip: Check original repository license

• NVIDIA nvdiffrast and nvdiffrec for differentiable rendering
• SMPL model from Max Planck Institute