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

The more smile, the more black #38

Open tengshaofeng opened 2 years ago

tengshaofeng commented 2 years ago

tbq_0a186d611d3e2493bfb35ec396390806_age tbq_1803151818-00006002_org_smile thanks for your great work. Did you find the phenomenon above?

danielroich commented 2 years ago

Thanks!

Yes, I have, I think it is connected to the fact that the above image is more "out of domain" compared to the second image. I found you need to add more "power" to the editing directions in order to edit OOD images.

This is also a finding from earlier papers, you can see it in our quantitative analysis.

PTI mitigates this pheromone but does not solve it completely

hope this helps

tengshaofeng commented 2 years ago

@danielroich thanks for your reply patiently. Did you means that I should collect more images which covers the OOD images, and
learn the new directions on the images?

danielroich commented 1 year ago

Something like that, use the same edit directions, but add more "power" to the tuning be feeding it with few-shot fine-tuning instead of one-shot as the classic PTI does

tengshaofeng commented 1 year ago

So when fine-tuning, should I fixed some layers?

danielroich commented 1 year ago

No Add more images Try using the multi_id_coach on several images of the same person