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

SG and SG2 issue #44

Closed FeiiYin closed 1 year ago

FeiiYin commented 1 year ago

Hi, thanks for the impressive work, which has raised a great impact. In the original paper, I wondered about the difference between SG, SG w+ and the first step of PTI. Both SG, PTI optimize the original W space, while SG w+ employ the w+ space. SG and SG w+ take more steps to optimize. And all the 3 methods employ noise regularization. Is there any other difference that I missed?

danielroich commented 1 year ago

No, it sounds just fine You can think of PTI as an extended version of SG W optimization. for its vanilla version. In the paper, we also use e4e instead of SG W optimization in order to apply StyleCLIP.

The first step before the pivotal tuning can be whatever you want. You will receive the editing attributes of the pivot latent code you have found.

We mainly use SG W Optimization with fewer steps because it shows excellent results in the popular editing techniques

FeiiYin commented 1 year ago

Thanks! Got it.