Open Arthurzhangsheng opened 2 years ago
For bad examples, HieRFE may predict bad latent codes for FTM to swap faces. These bad latent codes may not be used even for reconstruction. So, you could solve this by refining those latent codes through optimized GAN Inversion methods.
Or, you may want to fine-tune HieRFE and FTM with more data, which should be helpful either.
Exactly, some other similar cases were found when swapping faces from CelebAMask-HQ.
29230_to_29714, 29198_to_28541, 29387_to_28689 (in
Normal cases:
when I pick the celabAHQ img id randomly to swap, some id pairs seams wrong
badcases:
2-116
3-116
2-120
3-120
25043-1291
25043-1302
25043-1851
normal cases:![image](https://user-images.githubusercontent.com/33758384/139568327-16141cc8-0637-47da-bc93-7d194620664e.png)
anyone know why?