TropComplique / bicycle-gan

Multimodal edges to image translation

BicycleGAN in pytorch

This an implementation of Toward Multimodal Image-to-Image Translation.

Requirements

1. pytorch 1.0
2. numpy, Pillow, opencv

How to reproduce my results

1. Download edges2shoes dataset from here.
2. Edit the path in train.py file and run it.

You can also download my trained models from here (run01.tar.gz).
You can try them using inference/test.ipynb.

Notes

1. For the generator I use resnet-like architecture not unet (like they do in the original paper).
2. I insert style information into the generator using AdaIN layers (like in StyleGAN).
3. I feed into the generator not only edges but also a binary mask.
4. I also use the binary mask to mask the outputs of the discriminators.

Examples

Acknowledgments

This code is inspired by

1. junyanz/BicycleGAN
2. eveningglow/BicycleGAN-pytorch
3. NVlabs/MUNIT

Other relevant papers

1. A Style-Based Generator Architecture for Generative Adversarial Networks
2. Multimodal Unsupervised Image-to-Image Translation
3. Semantic Image Synthesis with Spatially-Adaptive Normalization
4. Latent Filter Scaling for Multimodal Unsupervised Image-to-Image Translation