User-Controllable Latent Transformer for StyleGAN Image Layout Editing

This repository contains our implementation of the following paper:

Yuki Endo: "User-Controllable Latent Transformer for StyleGAN Image Layout Editing," Computer Graphpics Forum (Pacific Graphics 2022) [Project] [PDF (preprint)]

Prerequisites

Python 3.8
PyTorch 1.9.0
Flask
Others (see env.yml)

Preparation

Download and decompress our pre-trained models.

Inference with our pre-trained models

We provide an interactive interface based on Flask. This interface can be locally launched with

python interface/flask_app.py --checkpoint_path=pretrained_models/latent_transformer/cat.pt

The interface can be accessed via http://localhost:8000/.

Training

The latent transformer can be trained with

python scripts/train.py --exp_dir=results --stylegan_weights=pretrained_models/stylegan2-cat-config-f.pt

Citation

Please cite our paper if you find the code useful:

@Article{endoPG2022,
Title = {User-Controllable Latent Transformer for StyleGAN Image Layout Editing},
Author = {Yuki Endo},
Journal = {Computer Graphics Forum},
volume = {},
number = {},
pages = {},
doi = {},
Year = {2022}
}

Acknowledgements

This code heavily borrows from the pixel2style2pixel and expansion repositories.