Figure 1: Sequences of image edits performed using control discovered with our method, applied to three different GANs. The white insets specify the particular edits using notation explained in Section 3.4 ('Layer-wise Edits').
GANSpace: Discovering Interpretable GAN Controls
Erik Härkönen1,2, Aaron Hertzmann2, Jaakko Lehtinen1,3, Sylvain Paris2
1Aalto University, 2Adobe Research, 3NVIDIA
https://arxiv.org/abs/2004.02546Abstract: This paper describes a simple technique to analyze Generative Adversarial Networks (GANs) and create interpretable controls for image synthesis, such as change of viewpoint, aging, lighting, and time of day. We identify important latent directions based on Principal Components Analysis (PCA) applied in activation space. Then, we show that interpretable edits can be defined based on layer-wise application of these edit directions. Moreover, we show that BigGAN can be controlled with layer-wise inputs in a StyleGAN-like manner. A user may identify a large number of interpretable controls with these mechanisms. We demonstrate results on GANs from various datasets.
Video: https://youtu.be/jdTICDa_eAI
See the setup instructions.
This repository includes versions of BigGAN, StyleGAN, and StyleGAN2 modified to support per-layer latent vectors.
Interactive model exploration
# Explore BigGAN-deep husky
python interactive.py --model=BigGAN-512 --class=husky --layer=generator.gen_z -n=1000000
# Explore StyleGAN2 ffhq in W space
python interactive.py --model=StyleGAN2 --class=ffhq --layer=style --use_w -n=1000000 -b=10000
# Explore StyleGAN2 cars in Z space
python interactive.py --model=StyleGAN2 --class=car --layer=style -n=1000000 -b=10000
# Apply previously saved edits interactively
python interactive.py --model=StyleGAN2 --class=ffhq --layer=style --use_w --inputs=out/directions
Visualize principal components
# Visualize StyleGAN2 ffhq W principal components
python visualize.py --model=StyleGAN2 --class=ffhq --use_w --layer=style -b=10000
# Create videos of StyleGAN wikiart components (saved to ./out)
python visualize.py --model=StyleGAN --class=wikiart --use_w --layer=g_mapping -b=10000 --batch --video
Options
Command line paramaters:
--model one of [ProGAN, BigGAN-512, BigGAN-256, BigGAN-128, StyleGAN, StyleGAN2]
--class class name; leave empty to list options
--layer layer at which to perform PCA; leave empty to list options
--use_w treat W as the main latent space (StyleGAN / StyleGAN2)
--inputs load previously exported edits from directory
--sigma number of stdevs to use in visualize.py
-n number of PCA samples
-b override automatic minibatch size detection
-c number of components to keep
All the figures presented in the paper can be recreated using the included Jupyter notebooks:
- Figure 1:
figure_teaser.ipynb
- Figure 2:
figure_pca_illustration.ipynb
- Figure 3:
figure_pca_cleanup.ipynb
- Figure 4:
figure_steerability_comp.ipynb
- Figure 5:
figure_biggan_edit_transferability.ipynb
- Figure 6:
figure_biggan_style_mixing.ipybb
- Figure 7:
figure_biggan_style_resampling.ipynb
- Figure 8:
figure_edit_zoo.ipynb
- The interactive viewer sometimes freezes on startup on Ubuntu 18.04. The freeze is resolved by clicking on the terminal window and pressing the control key. Any insight into the issue would be greatly appreciated!
- Create a wrapper for the model in
models/wrappers.py
using theBaseModel
interface. - Add the model to
get_model()
inmodels/wrappers.py
.
It is possible to import trained StyleGAN and StyleGAN2 weights from TensorFlow into GANSpace.
- Install TensorFlow:
conda install tensorflow-gpu=1.*
. - Modify methods
__init__()
,load_model()
inmodels/wrappers.py
under class StyleGAN.
- Follow instructions in stylegan2-pytorch's instructions.
- Save the converted checkpoint as
checkpoints/stylegan2/<dataset>_<resolution>.pt
. - Modify methods
__init__()
,download_checkpoint()
inmodels/wrappers.py
under class StyleGAN2.
We would like to thank:
- The authors of the PyTorch implementations of BigGAN, StyleGAN, and StyleGAN2:
Thomas Wolf, Piotr Bialecki, Thomas Viehmann, and Kim Seonghyeon. - Joel Simon from ArtBreeder for providing us with the landscape model for StyleGAN.
- David Bau and colleagues for the excellent GAN Dissection project.
- Justin Pinkney for the Awesome Pretrained StyleGAN collection.
- Toumas Kynkäänniemi for giving us a helping hand with the experiments.
- The Aalto Science-IT project for providing computational resources for this project.
This code is released for the purpose of academic reproducibility. No license is granted for derivative works or other uses, besides non-commercial experimentation. We are working on adding a real non-commercial license.
The directory netdissect
is a derivative of the GAN Dissection project, and is provided under the MIT license.
The directory models/biggan
is provided under the MIT license.