This repository contains training and evaluation code of Product of Gaussian Mixture Diffusion Models. It depends on the cuda implementation of the LogSumExp potential and activation function.
The following table establishes how to recreate the plots in the paper.
To reproduce any of the figures, run python [filename].py.
The framework assumes that the environment variable DATASETS_ROOT is set and contains the relevant datasets.
As an example, the directory $DATASETS_ROOT/bsds500/train/ should exist and contain the BSDS500 training dataset.
| Figure(s) | File(s) |
|---|---|
| 1 | dirac |
| 2 | --- |
| 3 and 14 | draw |
| 4 | draw_wavelets |
| 5 and 6 | draw_shearlets |
| 7 | compute_shearlet_approx |
| 8 and 9 and Tab. 1 | denoise |
| 10 | noise_estimation |
| 11 | noise_estimation_image |
| 12 | generate-analytical and draw_generated |
| 13 | studentt_diffusion_approx |
If you find the work interesting and use it in any of your research, please consider citing
@Article{Zach2024,
author={Zach, Martin and Kobler, Erich and Chambolle, Antonin and Pock, Thomas},
title={Product of Gaussian Mixture Diffusion Models},
journal={Journal of Mathematical Imaging and Vision},
year={2024},
month={Mar},
day={15},
issn={1573-7683},
doi={10.1007/s10851-024-01180-3},
url={https://doi.org/10.1007/s10851-024-01180-3}
}