This repository try to re-produce the idea of Re-cycle GAN [1] which is purposed by CMU. However, since CMU doesn't release any source code and collected dataset, we only extract the simple white and orange flower video to train the model. You should notice that it's not the official implementation. The idea of Re-cycle GAN is very similar to vid2vid. On the other hand, we provide the simple version whose idea can be traced much easily! For simplicity, this repository doesn't provide multi-GPU training or inference.

• master: This branch clearly address the code with detailed comment. You can refer to this branch if you are not realize the content of each part.
• develop: This branch will update the latest version of repo.
• perfect: This branch contains the code, and you can re-produce the result successfully by this branch.
• clear: Since the full comment is lengthy and hard to trace, we also provide the code version with least comment in the recycle_gan.py. Also, some redundant check will be removed to add the readability. You can read with shortest length.

The detail can be found in here. But you should download the dataset from the following link.

https://drive.google.com/drive/folders/1mmWND9ZLK9nZwa8lMQWOVjN5sU_rrWD0?usp=sharing

And you can simply use the following command:

# For training
$ python3 train.py --A <domain_A_path> --B <domain_B_path> --T 3 --resume result.pkl --record_iter 500 --n_iter 30000
# For inference
$ python3 demo.py --in <video_path> --direction a2b --resume result.pkl

The above image shows the result which can be re-produced in perfect branch. The detail explanation can be found in perfect branch.

[1] A. Bansal, S. Ma, D. Ramanan, and Y. Sheikh, "Recycle-gan: Unsupervised video retargeting," arXiv preprint arXiv:1808.05174, 2018.