Skip to content
Tensorflow implement of "Eye In-Painting with Exemplar Generative Adversarial Networks"
Branch: master
Clone or download
Type Name Latest commit message Commit time
Failed to load latest commit information.
images update generated result Jul 27, 2018 Update Jan 18, 2019 Update Jan 5, 2019 Update Jan 18, 2019 random_normal_initializer Nov 27, 2018 Code for different reference images in test Jul 27, 2018


The tensorflow implement of Eye In-Painting with Exemplar Generative Adversarial Networks


Paper Introduction.

Network Architecture

Paper result

Noted: Differences with the original paper.

  • Because the origin paper does't provide the details of model, this implement uses the architecture and hyperparamters from SG-GAN(Using adapted residual image learning for G and spectral norm for D)

  • Just use refernece image as the exemplar, not code.

  • Our model trained using 256x256 pixels, not 128x128 mentioned in the original paper.



  • Clone this repo:
git clone
  • Download the CelebA-ID dataset

You can download CelebA-ID Benchmark dataset according to the Dataset

and unzip CelebA-ID into a directory.

  • Train the model using the default parameter
python --OPER_FLAG=0 --path your_path
  • Test the model
python --OPER_FLAG=1 --path your_path --test_step= your model_name

Our results

(1)Input image; (2) Reference image; (3) image to in-paint; (4) in-painted image; (5) local real eye region; (6) generated eye region;

Different reference images

Reference code


DCGAN tensorflow

Spectral Norm tensorflow

Similar project


You can’t perform that action at this time.