yuval-alaluf

I have a feeling that there is something wrong with the inversion since the reconstruction is quite poor, perhaps leading to the poor editing you got on the right. If...

I haven't encountered such an issue before, but things like this usually happen because the CUDA version and Pytorch version are incompatible. I would check that you are able to...

If you want to change hairstyle, you can edit the images with StyleCLIP, for example. Details are provided in the README.

We don't have such a model available. SG2 generators trained on FFHQ at a resolution of 256 are available on the official SG2 repo from nvidia.

No since the latent spaces are different. You will need to retrained your encoder, although you could try starting from the pretrained encoder rather than training from scratch.

The StyleGAN generator we use here is the official generator that was converted from tensorflow to pytorch. We do not train our own generator.

Correct, we don't retrain it. We used rosinality's implementation of StyleGAN2 rather than StyleGAN2-ada-pytorch, but you can convert the models by following the instructions [here](https://github.com/yuval-alaluf/restyle-encoder#preparing-your-generator).

This is a strange behavior. What is the exact command you tried running?

It seems like your changes are correct. The entire pSp repo is based on SG2 which has latent codes of size 512. If you changed your generator to output latents...

Yea, I believe you'll need to use a batch size of at least 2 to make all the dimensions work out. In terms of reducing the number of parameters in...