RandomArt training code

Training and inference code for my first real Model, following the HF diffusion model course.
It's an unconditional 128×128 DDPM that generates painterly abstract images. The resulting diffusion model is published in the HF repository.

Examples

Inference

Run inference (assuming PyTorch and Diffusers are installed and the HF repo is downloaded to the same parent directory) from this directory with:

python inference.py

Which will generate 3 images at 500 steps each. For faster exploration, use 100 inference steps. Around 250 steps provides a useful speed/quality balance, while 500 steps gives more time to final renders.

Training

The Model was trained on a 36 GB M3 Max Mac in approx. 5 hours. The dataset is a ~7600 Row subset of Wikiart filtered for images under 2400px.

The images are resized to 128x128 and preprocessed, training then takes place over 20 epochs with a batch size of 8. The training script is train.py and can be run with:

python train.py

which assumes the above parameters. Before training the images have to be downloaded and preprocessed with:

python load_dataset.py

which as mentioned above streams the dataset to filter for images under 2400px and resizes them to 128x128. The resulting dataset is saved in ./wikiart_under_2400_resized_dataset. The loading process can be interrupted at any time if the desired amount of images has been downloaded.