Awesome

FreezeD: a Simple Baseline for Fine-tuning GANs

Update (2020/10/28)

Release checkpoints of StyleGAN fine-tuned on cat and dog datasets.

Update (2020/04/06)

Current code evaluates FID scores with inception.train() mode. Fixing it to inception.eval() may degrade the overall scores (both competitors and ours; hence the trend does not change). Thanks to @jychoi118 (Issue #3) for reporting this.

Official code for "Freeze the Discriminator: a Simple Baseline for Fine-Tuning GANs" (CVPRW 2020).

The code is heavily based on the StyleGAN-pytorch and SNGAN-projection-chainer codes.

See stylegan and projection directory for StyleGAN and SNGAN-projection experiments, respectively.

Note: There is a bug in PyTorch 1.4.0, hence one should use torch>=1.5.0 or torch<=1.3.0. See Issue #1.

Generated samples

Generated samples over fine-tuning FFHQ-pretrained StyleGAN

<img src="./resources/cat_trend.gif" width="384">   <img src="./resources/dog_trend.gif" width="384">

More generated samples (StyleGAN)

Generated samples under Animal Face and Anime Face datasets

<img src="./resources/stylegan/original.png" width="256">   <img src="./resources/stylegan/bear.png" width="256">   <img src="./resources/stylegan/cat.png" width="256">

<img src="./resources/stylegan/chicken.png" width="256">   <img src="./resources/stylegan/cow.png" width="256">   <img src="./resources/stylegan/deer.png" width="256">

<img src="./resources/stylegan/dog.png" width="256">   <img src="./resources/stylegan/duck.png" width="256">   <img src="./resources/stylegan/eagle.png" width="256">

<img src="./resources/stylegan/elephant.png" width="256">   <img src="./resources/stylegan/human.png" width="256">   <img src="./resources/stylegan/lion.png" width="256">

<img src="./resources/stylegan/monkey.png" width="256">   <img src="./resources/stylegan/mouse.png" width="256">   <img src="./resources/stylegan/panda.png" width="256">

<img src="./resources/stylegan/pigeon.png" width="256">   <img src="./resources/stylegan/pig.png" width="256">   <img src="./resources/stylegan/rabbit.png" width="256">

<img src="./resources/stylegan/sheep.png" width="256">   <img src="./resources/stylegan/tiger.png" width="256">   <img src="./resources/stylegan/wolf.png" width="256">

<img src="./resources/stylegan/miku.png" width="256">   <img src="./resources/stylegan/sakura.png" width="256">   <img src="./resources/stylegan/haruhi.png" width="256">

<img src="./resources/stylegan/fate.png" width="256">   <img src="./resources/stylegan/nanoha.png" width="256">   <img src="./resources/stylegan/lelouch.png" width="256">

<img src="./resources/stylegan/mio.png" width="256">   <img src="./resources/stylegan/yuki.png" width="256">   <img src="./resources/stylegan/shana.png" width="256">

More generated samples (SNGAN-projection)

Comparison of fine-tuning (left) and freeze D (right) under Oxford Flower, CUB-200-2011, and Caltech-256 datasets

Freeze D generates more class-consistent results (see row 2, 8 of Oxford Flower)

<img src="./resources/projection/flower_base.png" width="384">   <img src="./resources/projection/flower_freeze.png" width="384">

<img src="./resources/projection/cub_base.png" width="384">   <img src="./resources/projection/cub_freeze.png" width="384">

<img src="./resources/projection/caltech_base.png" width="384">   <img src="./resources/projection/caltech_freeze.png" width="384">

Citation

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

@inproceedings{
    mo2020freeze,
    title={Freeze the Discriminator: a Simple Baseline for Fine-Tuning GANs},
    author={Mo, Sangwoo and Cho, Minsu and Shin, Jinwoo},
    booktitle = {CVPR AI for Content Creation Workshop},
    year={2020},
}