Does Generation Require Memorization? Creative Diffusion Models using Ambient Diffusion

This repository hosts the official PyTorch implementation of the paper.

Installation

The recommended way to run the code is with an Anaconda/Miniconda environment.

Create a new Anaconda environment and install the dependencies using the following command:

conda env create -f environment.yml -n diffusion

Download datasets

You might also need to download dataset and dataset statistics for training and for FID calculation. To do so, follow the instructions provided here.

To create a small dataset with 300, 1000 and 3000 images, we randomly choose the images from the complete dataset.

Training New Models

To train a new model, set the arguments in scripts/train.sh and run the script:

bash scripts/train.sh

Generating images from a model and calculating FID

To generate images from a model, specify the model path CKPT, output directory to store the generated images OUTDIR and a path to reference statistics REF_PATH in scripts/generate_script.sh and then run the script:

bash scripts/generate_script.sh

Calculating similarity scores to measure memorization

To calculate the similarity score, use scripts/memorization_metrics.sh. First, specify the path of of generated images in GEN_DATA and training data in DATA and then run the script:

bash scripts/memorization_metrics.sh

Acknowledgement

This code is adapted from the EDM repository.