Globally and Locally Consistent Image Completion (GLCIC)

This is an implementation of the image completion model proposed in the paper (Globally and Locally Consistent Image Completion) with TensorFlow.

Requirements

Python 3
TensorFlow 1.3
OpenCV

Results

Usage

I. Prepare the training data

Put the images for training the "data/images" directory and convert images to npy format.

$ cd data
$ python to_npy.py

When you use the face images for training, it is recommended that you align the face landmarks with dlib before training. If you have no time for preprocessing, utilize mirror padding.

II. Train the GLCIC model

$ cd src
$ python train.py

Approach

This implementation uses 128x128 images as training data unlike paper. So the both discriminators have 1 conv layer fewer; that is, the local and global discriminator have 4 and 5 conv layers, respectively.

I trained the GLCIC model using 5,434 face images collected from the Internet. The paper says the training should be split into three phases, but I skipped the second step. The completion network is trained with the MSE loss for 100 interatinos; then both the completion network and discriminator are trained to reach the total of 400 iterations. The entire training procedure takes roughly 16 hours on a single machine equipped with a GTX 1070.

Future Work

The result is not good when marginal area is missing.

seokg/glcic