jacktsmi/ganimate

CycleGAN implementation which allows image-to-image translation between human faces and cartoon faces.

GANimate

CycleGAN implementation which allows image-to-image translation between human faces and cartoon faces.

Dependencies

• Python Version 3.8.5+
• PyTorch Version 1.8.0

Instructions for Training

• If you have not yet downloaded the data, do so by uncommenting the "Download data" portion in train.py.
• Run train.py

TODO : add file for testing pre-existing weights.

Results

I will summarize my results here, for a more detailed analysis of my training process, see here

Training Details:

• Trained map from human to cartoon for 15 epochs.
• Continued to train the inverse map and both discriminators for an additional 16 epochs (total of 31 epochs for those networks).
• Learning parameters are the same as those in the train.py file and CycleGAN paper.

Here are some sample outputs after training:

The left column of the images show the original domain image, the center column shows the mapped image, and the right column shows the cycled image.