Jun-Yan Zhu

Some recent [work](https://arxiv.org/abs/1808.04325) proposed to learn the geometric transformation.

Tensorflow?

2K might not be enough. Researchers often trained the colorization models using millions of RGB images. See the [colorization-pytorch](https://github.com/richzhang/colorization-pytorch) repo for more details.

You can also train the model with a paired dataset. See the [colorization-pytorch](https://github.com/richzhang/colorization-pytorch) and this [paper](https://arxiv.org/pdf/1603.08511.pdf) for more details.

You can create paired data by converting RGB to grayscale images. Please look at the paper for more details.

We want to make sure that the data is [-1, 1] (with mean=0). Also, the generator's last layer is TanH, whose range is also [-1, 1]. You can also do...

You need to train the model on faces with different kinds of skin complexion and eye sizes. These two papers might be related to your applications: [Conditional CycleGAN](https://arxiv.org/pdf/1705.09966.pdf) and [StarGAN](https://arxiv.org/abs/1711.09020).

See this two posts [1](https://github.com/junyanz/CycleGAN/issues/71) and [2](https://github.com/junyanz/CycleGAN/issues/57) for solutions. Also check out our [PyTorch](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix) version, which is more stable and still under active development.

@happsky and @ahmedsalaheldin Could you report your CUDA and Torch version?

See this [issue](https://github.com/junyanz/CycleGAN/issues/28) for potential solutions. Also if you are not familiar with Lua/Torch, you may want to use our PyTorch [version](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix) written in python.