search

danielroich / PTI

Official Implementation for "Pivotal Tuning for Latent-based editing of Real Images" (ACM TOG 2022) https://arxiv.org/abs/2106.05744
MIT License
897 stars 112 forks source link

How many images have been tried at most to train together? #33

Closed laolongboy closed 2 years ago

laolongboy commented 2 years ago

Thans for your contribution. How many images have been tried at most to train together? Ideally, each picture corresponds to one pivotal finetuned stylegan. I have 2w+ images with different IDs and quality of images are different. What if I train them together to get one stylegan?

danielroich commented 2 years ago

Hey @laolongboy, Personally, I have tried 64 at most while the locality regularization is active. Of course you can invert even more identities if you want.

You can use PTI on those 2 identities together and receive only one StyleGAN. Which will incorporate both identities. To do so you need to use the MultiIDCoach by passing True to the use_multi_id_training variable in the run_pti script.

You should receive results as if you have used PTI on each of them separately, but with only one StyleGAN

laolongboy commented 2 years ago

Thanks, I will try and update the results.

What's more, do we need to reseek the InterfaceGan's attrbute boundarys? or we can directly use the boundarys detected in the original stylegan space?

danielroich commented 2 years ago

The best part of the PTI process is that you can directly use the boundaries detected in the original StyleGAN space. Not only InterfaceGAN, but also GANSpace and StyleCLIP, and much more. The PTI process together with the regularization affects only the identity attribute in a very local environment.