Jaehoon Hahm*, Junho Lee*, Sunghyun Kim, Joonseok Lee (* equal contribution)
https://arxiv.org/abs/2407.11451
This repository is the official Pytorch implementation for Isometric diffusion.
An illustration of latent traversal between two latents
Setting the environment.
conda create --name isodiff python=3.9
conda activate isodiff
pip install torch==2.0.1 torchvision==0.15.2
pip install -r requirements.txt
Setting the Dataset. Change the 'DATASET_PATH' in 'submit_celeba.sh' and 'submit_celeba_ldm.sh' to CelebA-HQ dataset path. Your directory structure should look like:
$DATASET_PATH/xxx.png
$DATASET_PATH/xxy.png
$DATASET_PATH/[...]/xxz.png
# set accelerate. [https://huggingface.co/docs/accelerate/quicktour]
# DDPM
bash submit_celeba.sh
# LDM
bash submit_celeba_ldm.sh
You can find the pre-trained weights for the Isometric Diffusion model of CelebA-HQ at this link.
@article{hahm2024isometric,
title={Isometric Representation Learning for Disentangled Latent Space of Diffusion Models},
author={Hahm, Jaehoon and Lee, Junho and Kim, Sunghyun and Lee, Joonseok},
journal={arXiv preprint arXiv:2407.11451},
year={2024}
}