search

KingJamesSong / PDETraversal

ICML23 "Latent Traversals in Generative Models as Potential Flows"
https://arxiv.org/abs/2304.12944
Apache License 2.0
26 stars 2 forks source link
disentanglement-learning generative-adversarial-network variational-autoencoder

readme

PDETraversal

ICML23 paper "Latent Traversals in Generative Models as Potential Flows"
Yue Song1,2, Andy Keller1, Nicu Sebe2, Max Welling1
1University of Amsterdam, the Netherlands
2University of Trento, Italy


Overview of our learned potential PDEs for latent traversal in two different experimental settings.

Pre-trained GAN

Please first run checkpoint2model.py for downloading pre-trained GANs, and run anime.sh and anime_eval.sh for the training potential functions and evaluation.

Pre-trained VAE

Please first run train_vae.py to train VAEs and then run mnist.sh for training potentials.

Training VAE from scratch

Please run mnist_scratch.sh for training VAEs and potentials simultaneously.

Citation

If you think the code is helpful to your research, please consider citing our paper:

@inproceedings{song2023latent,
  title={Latent Traversals in Generative Models as Potential Flows},
  author={Song, Yue and Keller, Andy and Sebe, Nicu and Welling, Max},
  booktitle={ICML},
  year={2023},
  organization={PMLR}
}

The code is built based on WarpedGANSpace and we sincerely thank their contributions. If you have any questions or suggestions, please feel free to contact me via yue.song@unitn.it.