IrwGAN (ICCV2021)

Unaligned Image-to-Image Translation by Learning to Reweight

[Update] 12/15/2021 All dataset are released, trained models and generated images of IrwGAN are released

[Update] 11/16/2021 Code is pushed, selfie2anime-danbooru dataset released.

Dataset

selfie2anime-danbooru | selfie-horse2zebra-dog | horse-cat2dog-anime | beetle-tiger2lion-sealion

Trained Models and Generated Images

• selfie2anime-danbooru   IrwGAN | [Baseline] | [CycleGAN] | [MUNIT] | [GcGAN] | [NICE-GAN]
• selfie-horse2zebra-dog   IrwGAN | [Baseline] | [CycleGAN] | [MUNIT] | [GcGAN] | [NICE-GAN]
• horse-cat2dog-anime     IrwGAN | [Baseline] | [CycleGAN] | [MUNIT] | [GcGAN] | [NICE-GAN]
• beetle-tiger2lion-sealion IrwGAN | [Baseline] | [CycleGAN] | [MUNIT] | [GcGAN] | [NICE-GAN]

Basic Usage

• Training:

python main.py --dataroot=datasets/selfie2anime-danbooru 

• Resume:

python main.py --dataroot=datasets/selfie2anime-danbooru --phase=resume

• Test:

python main.py --dataroot=datasets/selfie2anime-danbooru --phase=test

• Beta Mode --beta_mode=A if domain A is unaligned, --beta_mode=B if domain B is unaligned, --beta_mode=AB if two domains are unaligned
• Effective Sample Size lambda_nos_A and lambda_nos_B are used to control how many samples are selected. The higher the weight, more samples are selected. We use 1.0 across all experiments.

Example Results

Citation

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

@inproceedings{xie2021unaligned,
  title={Unaligned Image-to-Image Translation by Learning to Reweight},
  author={Xie, Shaoan and Gong, Mingming and Xu, Yanwu and Zhang, Kun},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={14174--14184},
  year={2021}
}