Neural Fluid Fields

This is a small library for doing fluid simulation with neural fields. Check out our review paper, Neural Fields in Visual Computing and Beyond if you want to learn more about neural fields!

Code Organization

neuralff contains the bulk of the library. The library can be installed as a usual Python module, by doing python setup.py develop or any equivalent.

The library contains of several components:

The neuralff.ops module contains the core utility functions. In particular, neuralff/ops/fluid_physics_ops.py contains PDE loss functions, neuralff/ops/vector_ops.py contains differential operators, and neuralff/ops/fluid_ops.py contains functions for performing advection.

These functions generally take as input a neuralff.Field class, which can be a grid-based vector field, or a neural field so long as it can sample vectors on continuous coordinates with some mechanism.

Running the Demo

The demo is located in app. These are standalone demos which use the neuralff library to do things like real-time fluid simulation using neural fields.

The demo runs on glumpy and pycuda, which can be annoying to install. To install:

git clone https://github.com/inducer/pycuda
git submodule update --recursive --init
python configure.py --cuda-root=$CUDA_HOME --cuda-enable-gl
python setup.py develop
pip install pyopengl
pip install glumpy

To run the demo, simply run python3 app/interactive_app.py.

To change the inital image of the animation, set an arbitrary image named test.png in ./data.