UNIT: Unsupervised Image-to-Image Translation

Paper | Previous Implementation | Two Minute Papers Video

This is an improved implementation of UNIT. The main differences are:

• Use spectral normalization in the generator and the discriminator.
• Use the two-timescale update rule (TTUR) with the discriminator learning rate 0.0004 and the generator learning rate 0.0001.
• Use hinge loss instead of least square loss.

License

Imaginaire is released under NVIDIA Software license. For commercial use, please consult researchinquiries@nvidia.com

Software Installation

For installation, please checkout INSTALL.md.

Hardware Requirement

We trained our models using an NVIDIA DGX1 with 8 V100 32GB GPUs. You can try to use fewer GPUs or reduce the batch size if it does not fit in your GPU memory, but training stability and image quality cannot be guaranteed.

Training

The procedure for training is very similar to the procedure of training for MUNIT. Please checkout the project page of MUNIT for details.

Inference

• Download a small set of test examples

python scripts/download_test_data.py --model_name unit

• Or arrange your own data into the following format.

projects/munit/test_data/summer2winter_test
└───images_a
    └───0001.jpg
    └───0002.jpg
    └───0003.jpg
    ...
└───images_b
    └───0001.jpg
    └───0002.jpg
    └───0003.jpg
    ...

Translate images from winter to summer

• Inference command

python -m torch.distributed.launch --nproc_per_node=1 inference.py \
--config configs/projects/unit/winter2summer/base48_bs1.yaml \
--output_dir projects/unit/output/winter2summer

Outputs are saved in projects/unit/output/winter2summer:

Input	Translation

• To translate images in the other direction (from summer to winter), change a2b to False.

Input	Translation

Citation

If you use this code for your research, please cite our paper.

@inproceedings{liu2016unsupervised,
	title={Unsupervised Image-to-Image Translation Networks},
	author={Ming-Yu Liu and Thomas Breuel and Jan Kautz},
	booktitle={Conference on Neural Information Processing Systems (NeurIPS)}},
	year={2017}
}