Public pytorch implementation for our paper Patronus: Bringing Transparency to Diffusion Models with Prototypes.
If you find this code helpful in your research, please cite the following paper:
@article{weng2025patronus,
title={Patronus: Bringing Transparency to Diffusion Models with Prototypes},
author={Weng, Nina and Feragen, Aasa and Bigdeli, Siavash},
journal={arXiv preprint arXiv:2503.22782},
year={2025}
}
The code is implemented in Python 3.9 using the Pytorch library. We only tested the code with Pytorch 3.9. One way of getting all the requirements is by using virtualenv and the requirements.txt file.
pip install -r requirements.txt
Run the following command:
python ./train/train_patronus.py
configarations (including the dataset choice) are in ./train/train_config.py.
Visualization of the learned prototypes:
Run:
python ./analysis/interpretability/visualize_prototype.py to get the visulized prototypes. Configurations are in ./analysis/interpretability/visualization_config.py.
It will plot the result under ./records/vis_prototype/ (in the form of Fig. 5 in the paper).
Code under analysis/manipulation/manipulation_p.ipynb
For Diagnosis part of experiment, manipualtion on dataset are needed to create mannully controlable bias dataset for analysis, which contains in analysis/diagnosis/manipulate_ds.ipynb. After training the model with the biassed dataset, diagonose it with analysis/diagnosis/diagnosis.ipynb.
Evaluations in the paper could be found under analysis/evaluation. Including: Sec 4.1 Semantic Meaning of the Learned Prototypes (reconstruction/variations, interpolation, visualizing prototypes, consistency, extrapolation) and Sec 4.2 Prototype Quality (under p_quality.ipynb).
For generation quality:
a) Conditional Sampling:
Run python ./train/conditional_sampling.py --dataset dataset_name --version_num version_num to generate conditional samples first. Then compute the FID score by python -m pytorch_fid source_path_1 source_path_2. Notice that you might need to resize the test set images to the same as generated samples.
b) Unconditional Sampling:
For unconditioal sampling, firstly a latent ddim needs to be trained, by running python ./train/train_latent_ddim.py --dataset dataset_name --version_num version_num.
Then run python ./train/unconditional_sampling.py --dataset dataset_name --version_num version_num --latent_version latent_version.
FID score could be computed with the same package metioned above.
./analysis: including codes for interpretability, manipulation, evaluation and diagonosis../models: including codes for patronus model, diffusion model, loss function etc../records: where the results go to../train: training and inference (sampling) scripts.global_config.py: global dir path for the repo.
We use 5 datasets in this work, which could be downloaded from their offical website listed below:
The word Patronus originated from ancient Rome, that a patronus was the protector, sponsor, and benefactor of the client (source: Wikipedia).
It is now more widely known by J.K. Rowling's work, Harry Potter, where a Patronus is a protective charm used to repel Dementors:
The Patronus is a kind of positive force, a projection of the very things that the dementor feeds upon — hope, happiness, the desire to survive — but it cannot feel despair, as real humans can, so the dementors can’t hurt it. -- Harry Potter and the Prisoner of Azkaban
I find this symbolic alignment meaningful: our aim and vision for this work is to build explainability tools for diffusion models that act as protective mechanisms, diagnosing and defending against bias and irresponsible behaviors in generative AI.
Please use the issue tracking tool to report bugs or email directly to ninwe@dtu.dk. If you have further questions regarding the paper, please feel free to email ninwe@dtu.dk as well.