Sin3DM: Learning a Diffusion Model from a Single 3D Textured Shape
Rundi Wu, Ruoshi Liu, Carl Vondrick, Changxi Zheng
Columbia University
ICLR 2024

Clone this repo

git clone --recursive https://github.com/Sin3DM/Sin3DM.git

Install required packages

conda create -n sin3dm python=3.10
conda activate sin3dm
pip install -r requirements.txt

Download pretrained models from here and put under src/checkpoints. Then run our gradio demo for generation (~3GB VRAM):

cd src
python app.py

We provide a script on one example for data pre-processing, training, sampling and evaluation. Change the BLENDER_PATH in it if needed.

bash scripts/run_single.sh

We create training data by sampling points from a textured mesh:

cd data
python mesh_sampler.py -s {OBJ_PATH} -d {SAVE_NPZ_PATH} --n_surf 5000000

Add --watertight flag if the mesh is already watertight. Use mesh_sampler_pbr.py for pbr models.

Train the autoencoder and then the diffusion model:

cd src
python train.py --tag {EXP_DIR} --data_path {NPZ_PATH} --gpu_id 0

Generate new textured meshes:

cd src
python sample.py --tag {EXP_DIR} --n_samples 10 --n_faces 50000 --output results10 --gpu_id 0

• --n_faces specifies the desired number of mesh triangles.
• Use --resize for retargeting, e.g., --resize 1 1 1.5.
• By default, it uses DDPM sampling with 1000 steps. Add --use_ddim --timestep_respacing '100' for DDIM sampling.

Geometry quality (SSFID) relies on a pretrained 3D shape classifier. Please download it from here and put Clsshapenet_128.pth under evaluation folder. First, render each model from 8 views:

cd rendering
python mvrender_script.py -s {RESULT_DIR} -g 0 -bl {BLENDER_PATH}

BLENDER_PATH is the path to the Blender program.

Then, run evaluation for all the metrics:

cd evaluation
python eval_full.py -s {RESULT_DIR} -r {NPZ_DATA_DIR} -g 0

NPZ_DATA_DIR is the parent folder of the data path NPZ_PATH.

See rendering for the rendering scripts that are used to produce paper figures.

The code is partly based on guided-diffusion, BlenderToolBox and SinGAN. We thank the authors of these works for sharing their code.

@article{wu2023sin3dm,
  title={Sin3DM: Learning a Diffusion Model from a Single 3D Textured Shape},
  author={Wu, Rundi and Liu, Ruoshi and Vondrick, Carl and Zheng, Changxi},
  journal={arXiv preprint arXiv:2305.15399},
  year={2023}
}