Original Paper: https://arxiv.org/abs/1912.01865.
Original implementation can be found here: https://github.com/clovaai/stargan-v2.
Tensorflow variant can be found here: https://github.com/clovaai/stargan-v2-tensorflow.
Mod log:
-
Adversarial loss has been changed to Energy Distance as suggested in https://arxiv.org/abs/1705.10743 for unbiased gradient estimate.
-
Replaces all average pooling layer with strided convolution downsampling and tranposed upsampling counterpart allowing models to learn their own down/upsampling.
-
Down/upsampling are moved out of residual blks to limit distortion in information flows as suggested in https://arxiv.org/abs/1603.05027.
-
Color space has been normalized to avoid unnecessarily large gradients.
NOTE: Effects of mods are not yet tested extensively(WIP), use at your own peril.