This repo uses the denoising-diffusion-pytorch package and trains the network with the GalaxyMNIST images. The resulting network can be used to generate "fake" galaxy images the have the same distribution as the training set.
There are scripts for training with both the low-res and high-res GalaxyMNIST data sets, and scripts for training with or without the class labels.
The SLURM submission scripts are also included that I used with the SCIAMA HPC at the ICG.
For fun, there is also a gz_font.py file that creates fake galaxies that look like letters.
NOTE: The submissions scripts and the gz_font.py files have hard coded file paths, adjust as needed.