September 29, 2019

2228 words 11 mins read



StyleGAN Encoder - converts real images to latent space

repo name Puzer/stylegan-encoder
repo link
language Jupyter Notebook
size (curr.) 10603 kB
stars (curr.) 674
created 2019-02-11
license Other

StyleGAN — Encoder for Official TensorFlow Implementation

Python 3.6 TensorFlow 1.10 cuDNN 7.3.1 License CC BY-NC

Teaser image

These people are real – latent representation of them was found by using perceptual loss trick. Then this representations were moved along “smiling direction” and transformed back into images

Short explanation of encoding approach: 0) Original pre-trained StyleGAN generator is used for generating images

  1. Pre-trained VGG16 network is used for transforming a reference image and generated image into high-level features space
  2. Loss is calculated as a difference between them in the features space
  3. Optimization is performed only for latent representation which we want to obtain.
  4. Upon completion of optimization you are able to transform your latent vector as you wish. For example you can find a “smiling direction” in your latent space, move your latent vector in this direction and transform it back to image using the generator.

New scripts for finding your own directions will be realised soon. For now you can play with existing ones: smiling, age, gender. More examples you can find in the Jupyter notebook

Generating latent representation of your images

You can generate latent representations of your own images using two scripts:

  1. Extract and align faces from images

python raw_images/ aligned_images/

  1. Find latent representation of aligned images

python aligned_images/ generated_images/ latent_representations/

  1. Then you can play with Jupyter notebook

Feel free to join the research. There is still much room for improvement:

  1. Better model for perceptual loss
  2. Is it possible to generate latent representations by using other model instead of direct optimization ? (WIP)

Stay tuned!

Original Readme:

This repository contains (no longer) official TensorFlow implementation of the following paper:

A Style-Based Generator Architecture for Generative Adversarial Networks Tero Karras (NVIDIA), Samuli Laine (NVIDIA), Timo Aila (NVIDIA)

Abstract: We propose an alternative generator architecture for generative adversarial networks, borrowing from style transfer literature. The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis. The new generator improves the state-of-the-art in terms of traditional distribution quality metrics, leads to demonstrably better interpolation properties, and also better disentangles the latent factors of variation. To quantify interpolation quality and disentanglement, we propose two new, automated methods that are applicable to any generator architecture. Finally, we introduce a new, highly varied and high-quality dataset of human faces.

For business inquiries, please contact

For press and other inquiries, please contact Hector Marinez at


All material related to our paper is available via the following links:

Link Description Paper PDF. Result video. Source code. Flickr-Faces-HQ dataset. Google Drive folder.

Additional material can be found in Google Drive folder:

Path Description
StyleGAN Main folder.
├  stylegan-paper.pdf High-quality version of the paper PDF.
├  stylegan-video.mp4 High-quality version of the result video.
├  images Example images produced by our generator.
│  ├  representative-images High-quality images to be used in articles, blog posts, etc.
│  └  100k-generated-images 100,000 generated images for different amounts of truncation.
│     ├  ffhq-1024x1024 Generated using Flickr-Faces-HQ at 1024×1024.
│     ├  bedrooms-256x256 Generated using LSUN Bedroom at 256×256.
│     ├  cars-512x384 Generated using LSUN Car at 512×384.
│     └  cats-256x256 Generated using LSUN Cat at 256×256.
├  videos Example videos produced by our generator.
│  └  high-quality-video-clips Individual segments of the result video as high-quality MP4.
├  ffhq-dataset Raw data for the Flickr-Faces-HQ dataset.
└  networks Pre-trained networks as pickled instances of dnnlib.tflib.Network.
   ├  stylegan-ffhq-1024x1024.pkl StyleGAN trained with Flickr-Faces-HQ dataset at 1024×1024.
   ├  stylegan-celebahq-1024x1024.pkl StyleGAN trained with CelebA-HQ dataset at 1024×1024.
   ├  stylegan-bedrooms-256x256.pkl StyleGAN trained with LSUN Bedroom dataset at 256×256.
   ├  stylegan-cars-512x384.pkl StyleGAN trained with LSUN Car dataset at 512×384.
   ├  stylegan-cats-256x256.pkl StyleGAN trained with LSUN Cat dataset at 256×256.
   └  metrics Auxiliary networks for the quality and disentanglement metrics.
      ├  inception_v3_features.pkl Standard Inception-v3 classifier that outputs a raw feature vector.
      ├  vgg16_zhang_perceptual.pkl Standard LPIPS metric to estimate perceptual similarity.
      ├  celebahq-classifier-00-male.pkl Binary classifier trained to detect a single attribute of CelebA-HQ.
      └ ⋯ Please see the file listing for remaining networks.


All material, excluding the Flickr-Faces-HQ dataset, is made available under Creative Commons BY-NC 4.0 license by NVIDIA Corporation. You can use, redistribute, and adapt the material for non-commercial purposes, as long as you give appropriate credit by citing our paper and indicating any changes that you’ve made.

For license information regarding the FFHQ dataset, please refer to the Flickr-Faces-HQ repository.

inception_v3_features.pkl and inception_v3_softmax.pkl are derived from the pre-trained Inception-v3 network by Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. The network was originally shared under Apache 2.0 license on the TensorFlow Models repository.

vgg16.pkl and vgg16_zhang_perceptual.pkl are derived from the pre-trained VGG-16 network by Karen Simonyan and Andrew Zisserman. The network was originally shared under Creative Commons BY 4.0 license on the Very Deep Convolutional Networks for Large-Scale Visual Recognition project page.

vgg16_zhang_perceptual.pkl is further derived from the pre-trained LPIPS weights by Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The weights were originally shared under BSD 2-Clause “Simplified” License on the PerceptualSimilarity repository.

System requirements

  • Both Linux and Windows are supported, but we strongly recommend Linux for performance and compatibility reasons.
  • 64-bit Python 3.6 installation. We recommend Anaconda3 with numpy 1.14.3 or newer.
  • TensorFlow 1.10.0 or newer with GPU support.
  • One or more high-end NVIDIA GPUs with at least 11GB of DRAM. We recommend NVIDIA DGX-1 with 8 Tesla V100 GPUs.
  • NVIDIA driver 391.35 or newer, CUDA toolkit 9.0 or newer, cuDNN 7.3.1 or newer.

Using pre-trained networks

A minimal example of using a pre-trained StyleGAN generator is given in When executed, the script downloads a pre-trained StyleGAN generator from Google Drive and uses it to generate an image:

> python
Downloading .... done

Gs                              Params    OutputShape          WeightShape
---                             ---       ---                  ---
latents_in                      -         (?, 512)             -
images_out                      -         (?, 3, 1024, 1024)   -
---                             ---       ---                  ---
Total                           26219627

> ls results
example.png #

A more advanced example is given in The script reproduces the figures from our paper in order to illustrate style mixing, noise inputs, and truncation:

> python
results/figure02-uncurated-ffhq.png     #
results/figure03-style-mixing.png       #
results/figure04-noise-detail.png       #
results/figure05-noise-components.png   #
results/figure08-truncation-trick.png   #
results/figure10-uncurated-bedrooms.png #
results/figure11-uncurated-cars.png     #
results/figure12-uncurated-cats.png     #

The pre-trained networks are stored as standard pickle files on Google Drive:

# Load pre-trained network.
url = '' # karras2019stylegan-ffhq-1024x1024.pkl
with dnnlib.util.open_url(url, cache_dir=config.cache_dir) as f:
    _G, _D, Gs = pickle.load(f)
    # _G = Instantaneous snapshot of the generator. Mainly useful for resuming a previous training run.
    # _D = Instantaneous snapshot of the discriminator. Mainly useful for resuming a previous training run.
    # Gs = Long-term average of the generator. Yields higher-quality results than the instantaneous snapshot.

The above code downloads the file and unpickles it to yield 3 instances of dnnlib.tflib.Network. To generate images, you will typically want to use Gs – the other two networks are provided for completeness. In order for pickle.load() to work, you will need to have the dnnlib source directory in your PYTHONPATH and a tf.Session set as default. The session can initialized by calling dnnlib.tflib.init_tf().

There are three ways to use the pre-trained generator:

  1. Use for immediate-mode operation where the inputs and outputs are numpy arrays:

    # Pick latent vector.
    rnd = np.random.RandomState(5)
    latents = rnd.randn(1, Gs.input_shape[1])
    # Generate image.
    fmt = dict(func=tflib.convert_images_to_uint8, nchw_to_nhwc=True)
    images =, None, truncation_psi=0.7, randomize_noise=True, output_transform=fmt)

    The first argument is a batch of latent vectors of shape [num, 512]. The second argument is reserved for class labels (not used by StyleGAN). The remaining keyword arguments are optional and can be used to further modify the operation (see below). The output is a batch of images, whose format is dictated by the output_transform argument.

  2. Use Gs.get_output_for() to incorporate the generator as a part of a larger TensorFlow expression:

    latents = tf.random_normal([self.minibatch_per_gpu] + Gs_clone.input_shape[1:])
    images = Gs_clone.get_output_for(latents, None, is_validation=True, randomize_noise=True)
    images = tflib.convert_images_to_uint8(images)

    The above code is from metrics/ It generates a batch of random images and feeds them directly to the Inception-v3 network without having to convert the data to numpy arrays in between.

  3. Look up Gs.components.mapping and Gs.components.synthesis to access individual sub-networks of the generator. Similar to Gs, the sub-networks are represented as independent instances of dnnlib.tflib.Network:

    src_latents = np.stack(np.random.RandomState(seed).randn(Gs.input_shape[1]) for seed in src_seeds)
    src_dlatents =, None) # [seed, layer, component]
    src_images =, randomize_noise=False, **synthesis_kwargs)

    The above code is from It first transforms a batch of latent vectors into the intermediate W space using the mapping network and then turns these vectors into a batch of images using the synthesis network. The dlatents array stores a separate copy of the same w vector for each layer of the synthesis network to facilitate style mixing.

The exact details of the generator are defined in training/ (see G_style, G_mapping, and G_synthesis). The following keyword arguments can be specified to modify the behavior when calling run() and get_output_for():

  • truncation_psi and truncation_cutoff control the truncation trick that that is performed by default when using Gs (ψ=0.7, cutoff=8). It can be disabled by setting truncation_psi=1 or is_validation=True, and the image quality can be further improved at the cost of variation by setting e.g. truncation_psi=0.5. Note that truncation is always disabled when using the sub-networks directly. The average w needed to manually perform the truncation trick can be looked up using Gs.get_var('dlatent_avg').

  • randomize_noise determines whether to use re-randomize the noise inputs for each generated image (True, default) or whether to use specific noise values for the entire minibatch (False). The specific values can be accessed via the tf.Variable instances that are found using [var for name, var in Gs.components.synthesis.vars.items() if name.startswith('noise')].

  • When using the mapping network directly, you can specify dlatent_broadcast=None to disable the automatic duplication of dlatents over the layers of the synthesis network.

  • Runtime performance can be fine-tuned via structure='fixed' and dtype='float16'. The former disables support for progressive growing, which is not needed for a fully-trained generator, and the latter performs all computation using half-precision floating point arithmetic.

Preparing datasets for training

The training and evaluation scripts operate on datasets stored as multi-resolution TFRecords. Each dataset is represented by a directory containing the same image data in several resolutions to enable efficient streaming. There is a separate *.tfrecords file for each resolution, and if the dataset contains labels, they are stored in a separate file as well. By default, the scripts expect to find the datasets at datasets/<NAME>/<NAME>-<RESOLUTION>.tfrecords. The directory can be changed by editing

result_dir = 'results'
data_dir = 'datasets'
cache_dir = 'cache'

To obtain the FFHQ dataset (datasets/ffhq), please refer to the Flickr-Faces-HQ repository.

To obtain the CelebA-HQ dataset (datasets/celebahq), please refer to the Progressive GAN repository.

To obtain other datasets, including LSUN, please consult their corresponding project pages. The datasets can be converted to multi-resolution TFRecords using the provided

> python create_lsun datasets/lsun-bedroom-full ~/lsun/bedroom_lmdb --resolution 256
> python create_lsun_wide datasets/lsun-car-512x384 ~/lsun/car_lmdb --width 512 --height 384
> python create_lsun datasets/lsun-cat-full ~/lsun/cat_lmdb --resolution 256
> python create_cifar10 datasets/cifar10 ~/cifar10
> python create_from_images datasets/custom-dataset ~/custom-images

Training networks

Once the datasets are set up, you can train your own StyleGAN networks as follows:

  1. Edit to specify the dataset and training configuration by uncommenting or editing specific lines.
  2. Run the training script with python
  3. The results are written to a newly created directory results/<ID>-<DESCRIPTION>.
  4. The training may take several days (or weeks) to complete, depending on the configuration.

By default, is configured to train the highest-quality StyleGAN (configuration F in Table 1) for the FFHQ dataset at 1024×1024 resolution using 8 GPUs. Please note that we have used 8 GPUs in all of our experiments. Training with fewer GPUs may not produce identical results – if you wish to compare against our technique, we strongly recommend using the same number of GPUs.

Expected training time for 1024×1024 resolution using Tesla V100 GPUs:

GPUs Training time
1 5 weeks
2 3 weeks
4 2 weeks
8 1 week

Evaluating quality and disentanglement

The quality and disentanglement metrics used in our paper can be evaluated using By default, the script will evaluate the Fréchet Inception Distance (fid50k) for the pre-trained FFHQ generator and write the results into a newly created directory under results. The exact behavior can be changed by uncommenting or editing specific lines in

Expected evaluation time and results for the pre-trained FFHQ generator using one Tesla V100 GPU:

Metric Time Result Description
fid50k 16 min 4.4159 Fréchet Inception Distance using 50,000 images.
ppl_zfull 55 min 664.8854 Perceptual Path Length for full paths in Z.
ppl_wfull 55 min 233.3059 Perceptual Path Length for full paths in W.
ppl_zend 55 min 666.1057 Perceptual Path Length for path endpoints in Z.
ppl_wend 55 min 197.2266 Perceptual Path Length for path endpoints in W.
ls 10 hours z: 165.0106w: 3.7447 Linear Separability in Z and W.

Please note that the exact results may vary from run to run due to the non-deterministic nature of TensorFlow.


We thank Jaakko Lehtinen, David Luebke, and Tuomas Kynkäänniemi for in-depth discussions and helpful comments; Janne Hellsten, Tero Kuosmanen, and Pekka Jänis for compute infrastructure and help with the code release.

comments powered by Disqus