Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

GAN steerability

Project Page | Paper

On the "steerability" of generative adversarial networks.
Ali Jahanian*, Lucy Chai*, Phillip Isola


  • Linux
  • Python 3

Table of Contents:

  1. Setup
  2. Visualizations - plotting image panels, videos, and distributions
  3. Training - pipeline for training your own walks
  4. Notebooks - some jupyter notebooks, good place to start for trying your own transformations
  5. PyTorch/Colab Demo - pytorch implementation in a colab notebook


  • Clone this repo:
git clone
  • Install dependencies:
    • we provide a Conda environment.yml file listing the dependencies. You can create a Conda environment with the dependencies using:
conda env create -f environment.yml
  • Download resources:
    • we provide a script for downloading associated resources (e.g. stylegan). Fetch these by running:
bash resources/


Plotting image panels:

  • Run python -h to list available visualization options. The key things to provide are a model checkpoint and a config yml file. We added some pretrained weights in the ./models_pretrained, but you can also use the models you train yourself.
python \
	models_pretrained/biggan_zoom_linear_lr0.0001_l2/model_20000_final.ckpt \
	models_pretrained/biggan_zoom_linear_lr0.0001_l2/opt.yml \
	--gpu 0 --num_samples 50 --noise_seed 20 --truncation 0.5 --category 207

python \
        models_pretrained/stylegan_color_linear_lr0.0001_l2_cats_w/model_2000_final.ckpt \
        models_pretrained/stylegan_color_linear_lr0.0001_l2_cats_w/opt.yml \
        --gpu 1 --num_samples 10 --noise_seed 20 
  • By default this will save generated images to <output_dir>/images specified in the config yml, unless overwritten with the --output_dir option

To make a videos:

python [CHECKPOINT] [CONFIG] --gpu [GPU] --noise_seed [SEED] --sample [SAMPLE]

python models_pretrained/biggan_color_linear_lr0.001_l2/model_20000_final.ckpt \
	models_pretrained/biggan_color_linear_lr0.001_l2/opt.yml  --gpu 0 --sample 10 \
	--noise_seed 20 --truncation 0.5 --category 538 --min_alpha -1 --max_alpha 0
  • By default this will save output to <output_dir>/video specified in the config yml, unless overwritten with the --output_dir and --filename options

To draw distributions:

To draw distributions, you will need to have downloaded the object detector through resources/ (for objects) or installed dlib through environment.yml (for faces).

python [CHECKPOINT] [CONFIG] --gpu [GPU]

python models_pretrained/biggan_shiftx_linear_lr0.001_l2/model_20000_final.ckpt \
	models_pretrained/biggan_shiftx_linear_lr0.001_l2/opt.yml  --gpu 0
  • By default this will save output to <output_dir>/distribution specified in the config yml, unless overwritten with the --output_dir option

Training walks

  • The current implementation covers these variants:
    • models: biggan, stylegan, pgan
    • transforms: color, colorlab, shiftx, shifty, zoom, rotate2d, rotate3d
    • walk_type: linear, NNz
    • losses: l2, lpips
  • Some examples of commands for training walks:
# train a biggan NN walk for shiftx with lpips loss
python --model biggan --transform shiftx --num_samples 20000 --learning_rate 0.0001 \
	--walk_type NNz --loss lpips --gpu 0 --eps 25 --num_steps 5

# train a stylegan linear walk with l2 loss using the w latent space
python --model stylegan --transform color --num_samples 2000 --learning_rate 0.0001 \
	--walk_type linear --loss l2 --gpu 0 --latent w --model_save_freq 100

# train a pgan linear walk with l2 loss
python --model pgan --transform color --num_samples 60000 --learning_rate 0.0001 \
	--walk_type linear --loss l2 --gpu 0 --dset celebahq --model_save_freq 1000
  • Currently for pgan we only support the celebahq dataset, but you can easily train walks using other pretrained models by downloading the weights to resources/pgan_pretrained and adding the model information to graphs/pgan/

  • Alternatively you can train using a config yml file, for example:

python --config_file config/biggan_color_linear.yml
  • Each training run will save a config file called opt.yml in the same directory as the weights, which can be used for rerunning experiments with the same settings. You may need to use the flag --overwrite_config for overwriting existing weights and config files.

  • Run python -h to list available training options

  • You can use these walks that you train yourself for the above visualizations, by specifying the path to the corresponding model checkpoint and opt.yml config file.


source activate gan_steerability
python -m ipykernel install --user --name gan_steerability



If you use this code for your research, please cite our paper:

  title={On the "steerability" of generative adversarial networks},
  author={Jahanian, Ali and Chai, Lucy and Isola, Phillip},
  booktitle={International Conference on Learning Representations},