zhujiapeng

See this [repo](https://github.com/genforce/genforce).

Please refer to the `README.md` of this [repo](https://github.com/genforce/idinvert)

This is the first version of our encoder structure, the weights give here are exactly match the encoder's structure here. We updated the TensorFlow version a little from the first...

You can try on the images with the resolution of 256x256 and see if the problem still happens.

Your environment may cause it. I find some solutions, such as [here](https://blog.csdn.net/Leo_Xu06/article/details/82023330) and [here](https://stackoverflow.com/questions/43990046/tensorflow-blas-gemm-launch-failed), and see if these can help.

Yes, the provided boundaries are obtained using InterFaceGAN. For other classes, you also can use InterFaceGAN or other methods such as the one in [here](https://github.com/zhujiapeng/resefa).

Yes, InterfaceGAN can be used to find the editing directions. However, you need some classifiers trained on the scene (*e.g.,* Places) to obtain your expected direction.

For your medical dataset, is the generator trained by yourself?

Conditional StyleGAN? Do you mean you are using one-hot labels?

1. The initial reconstruction is still a realistic image? After getting the encoded latent codes, did you feed them to the synthesized network directly or fed them to the mapping...