This code accompnies the paper "Constructive Solid Geometry on Neural Signed Distance Fields" by Zoë Marschner, Silvia Sellán, Derek Liu, and Alec Jacobson.

Create a virtual enviroment if desired, then install the dependencies (matplotlib and pytorch):

pip3 install torch
pip3 install matplotlib

Matplotlib is only needed for visualization, so if you want to train the model on a seperate device with a GPU, you will only need to install pytorch there.

To train networks for example CSG and SV problems, run either ex_csg.py or ex_sv.py—you will likely want to train these on a GPU. The results of these are also precomputed, and can be found in the models/ and data/ folders. Results can be visualized using viz_ex.py. For example, run either:

python3 viz_ex.py swept_star -1.05 2.05 -0.9 2.2
python3 viz_ex.py circle_square_union

to visualize the precomputed results! (adding -s as an argument at the end of these commands will save the resulting images).

The core contribution of our paper is the losses defined in nsdf_csg_losses.py, where you will find the closest point loss and the editing losses.

Check back soon for additional examples, including 3D examples!