The model is described in https://arxiv.org/abs/2311.05217
Our code is built off of https://github.com/Janspiry/Palette-Image-to-Image-Diffusion-Models.
Requirements: torch>=1.6, numpy, scipy, tqdm
To train the model on (say) GPUs 0 and 1:
python run.py -c 'config/sr-mask-arepo-48px-264px-128p-2000T.json' --gpu '0,1'
To generate a field from a trained model, change the phase to test:
python run.py -c 'config/sr-mask-arepo-48px-264px-128p-2000T.json' -p test
To run the iterative outpainting described in the paper, python run.py is called many times over, each time with a new .json file shifting which data is run on. To automate this, first generate a batch file run_big.bat with
python generate_patch_script.py
Then, run
sh run_big.bat
For iterative outpainting, the low-resolution data that the .json file finds must be properly organized sequentially in the outpanting invervals that you want (we used 24px intervals on 48px cubes). There are scripts to preprocess the data in matter_density_data/.