Closed hallwaypzh closed 2 years ago
We trained two models, one for ff++ and one for celeba-hq (you can see more details in the experiment section). The released model is the latter one.
The difference is the training dataset, i.e. the model trained for ff++ used more images from celeba, which gives better generalization on 256px images.
Thank u for the reply, will the other one be released?
Nope
I have tried the inference code on the examples in the paper (Figure 5.), however this one goes horribly wrong ,
this is the picture generated by stylegan
![sg2](https://user-images.githubusercontent.com/29053705/136660928-a6479b13-192d-4d55-a090-f2bb64c1b4a9.png)
I'm using ftm method btw