Neural Flow Diffusion Models

Neural Flow Diffusion Models (NFDM) is a framework that generalizes conventional diffusion models. The key idea in NFDM is to define the forward process' conditional SDE implicitly via a learnable transformation $F_\varphi(\varepsilon, t, \mathbf{x})$ that defines the marginal distributions. This lets the user define a broad range of continuous time- and data-dependent forward processes, that the reverse process will learn to invert. Importantly, NFDM retains crucial properties of conventional diffusion models, like likelihood-based and simulation-free training.

Check out the paper for more details.

Setup

We provide a requirements.txt file for a pip environment.

pip install -r requirements.txt

Getting started

The best way to understand our method is to look at the notebook which contains a simple implementation of the NFDM model on a toy dataset.

Citation

If you find our work useful, please consider citing our paper:

@article{bartosh2024neural,
  title={Neural flow diffusion models: Learnable forward process for improved diffusion modelling},
  author={Bartosh, Grigory and Vetrov, Dmitry P and Andersson Naesseth, Christian},
  journal={Advances in Neural Information Processing Systems},
  volume={37},
  pages={73952--73985},
  year={2024}
}