This is the code that accompanies the paper "Disentangling Mean Embeddings for Better Diagnostics of Image Generators", published at IAI Workshop @ NeurIPS 2024.
We propose a novel approach to disentangle the cosine similarity of image-wise mean embeddings into the product of cosine similarities for individual pixel clusters via central kernel alignment. This allows quantifying the contribution of the cluster-wise performance to the overall image generation performance.
First, download and unzip img_align_celeba.zip into data/celeba/ from, e.g., https://cseweb.ucsd.edu/~weijian/static/datasets/celeba/.
Second, run all cells in preprocess_celeba.ipynb.
Third, download all files from here and put them into the folder data (this includes the DCGAN and DDPM image generations).
Then, run the cells in experiments.ipynb.
The figures are printed as cell outputs and are also stored in the folder plots.
If you found this work or code useful, please cite:
@inproceedings{
gruber2024disentangling,
title={Disentangling Mean Embeddings for Better Diagnostics of Image Generators},
author={Sebastian Gregor Gruber and Pascal Tobias Ziegler and Florian Buettner},
booktitle={Interpretable AI: Past, Present and Future},
year={2024},
url={https://openreview.net/forum?id=mZkShMbCaS}
}
Everything is licensed under the MIT License.