PTI inversion + DragGAN's quality is not good

[Dae-Hyeon Choi]

Hello, thanks for your great work!

I'm trying to convert real image to latent through PTI inversion as you said and put it in DragGAN.

I'm using pre-trained StyleGAN with FFHQ-512 that you suggested, and I'm reproducing Drag Editing with FFHQ dataset, but the quality is not very good.

Can you present the configuration (e.g., latent space, generator optimization steps, latent inversion steps) in PTI implementation, or present the PTI inversion code that was actually used when showing qualitative results included in the paper?

My questions are specifically as follows.

Question about your PTI implementation

1. After using PTI inversion, do you create additional optimized StyleGANs when you create Edited Images? (I think it would have been, of course.)
2. Which space did you project latent to, w space or w+ space? (In my case, w space)
3. How many steps did you use to inversion latent? (In my case, 1000 steps)
4. How many steps do you do generator optimization? (In my case, 400 steps)
5. Which resolution did you use in real images, and which model did you use in qualitative results? (I'm curious about face dragging results. In section 4 - man with glasses, In section 13 - man)

Now, I implemented it by referring to the implementation of #142, but the quality is not very good. I would appreciate it if you could answer what would be the problem.

Thank you.