zhujiapeng

No, the J^TJ on the background needs to be full rank since it is the denominator.

Please refer [here](https://github.com/bytedance/Hammer/blob/main/docs/model_conversion.md)

We do not do experiments on truncated latent code since we focus on W space instead of W-plus space. I guess the truncation trick barely influences the results since we...

Scale images to [0, 1], and then compute the mean square error only in/out of the mask region, and the mask is the one we show on the images in...

Yes, it is normal. It will take about one hour when using a 1024x1024 model on 2080 Ti GPU. StyleGAN3 will take a much longer time. We will release the...