Figure: Unsupervided Framework CLIP2StyleGAN method
The success of StyleGAN has enabled unprecedented semantic editing capabilities, on both synthesized and real images. However, such editing operations are either trained with semantic supervision or described using human guidance. In another development, the CLIP architecture has been trained with internet-scale image and text pairings and has been shown to be useful in several zero-shot learning settings. In this work, we investigate how to effectively link the pretrained latent spaces of StyleGAN and CLIP, which in turn allows us to automatically extract semantically labeled edit directions from StyleGAN, finding and naming meaningful edit operations without any additional human guidance. Technically, we propose two novel building blocks; one for finding interesting CLIP directions and one for labeling arbitrary directions in CLIP latent space. The setup does not assume any pre-determined labels and hence we do not require any additional supervised text/attributes to build the editing framework. We evaluate the effectiveness of the proposed method and demonstrate that extraction of disentangled labeled StyleGAN edit directions is indeed possible, and reveals interesting and non-trivial edit directions.
CLIP2StyleGAN: Unsupervised Extraction of StyleGAN Edit Directions
Rameen Abdal Peihao Zhu, John Femiani, Niloy Mitra, Peter Wonka
KAUST, Miami University, UCL and Adobe Research
[Paper]
Clone this repo.
git clone https://github.com/RameenAbdal/CLIP2StyleGAN.git
cd CLIP2StyleGAN/Install dependencies.
conda create --n CLIP2StyleGAN python=3.7
conda activate CLIP2StyleGAN
conda install --yes -c pytorch pytorch=1.7.1 torchvision cudatoolkit=11.0
pip install ftfy regex tqdm natsortDownload the StyleGAN2 models, test data and encoded features here.
Compute and save the PCA directions and the projections in a folder.
python compute_pca.py image_path [path to the Dataset (Real/Sampled)] --extracted_features_path [path to the CLIP Image encoder features]Compute the labels given an image folder. For testing, two folders representing first two principal components can be downloaded here.
python text_optimize.py --path [path to the image folder]Disentangle the entangled directions. For testing, two folders representing entangled edits mentioned in the paper can be downloaded here.
python optimize_dir.py --path_image_data [path to the image folder] --path_features [path to the CLIP Image encoder features] --pca_axis [index of the principal axis]Perform the edits on Faces and Cars. StyleGAN2 pretrained models can be downloaded here.
cd edits_sg
python edit_faces.py --edit_type ['to_male', 'to_female', 'glasses','smile', 'kids', 'beard'] --ckpt [path to checkpoint]
python edit_cars.py --edit_type ['scrap_to_car', 'red_car', 'capri', 'race_car', 'blazer'] --ckpt [path to checkpoint]Download the data here to compute the CLIP edit scores (compared with GANSpace) and Disentanglement results.
python test_clip_scores.py
python test_disentanglement.py@article{DBLP:journals/corr/abs-2112-05219,
author = {Rameen Abdal and
Peihao Zhu and
John Femiani and
Niloy J. Mitra and
Peter Wonka},
title = {CLIP2StyleGAN: Unsupervised Extraction of StyleGAN Edit Directions},
journal = {CoRR},
volume = {abs/2112.05219},
year = {2021},
url = {https://arxiv.org/abs/2112.05219},
eprinttype = {arXiv},
eprint = {2112.05219},
timestamp = {Wed, 19 Jan 2022 09:30:45 +0100},
biburl = {https://dblp.org/rec/journals/corr/abs-2112-05219.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
The CLIP and StyleGAN2 codes are taken from rosinality/stylegan2-pytorch and openai/CLIP. This work was supported by Adobe Research and KAUST Office of Sponsored Research (OSR).