Not working on middle eastern - Muslim faces #141

New Issue

claunia · 2026-01-29T21:43:38Z

claunia commented

2026-01-29 21:43:38 +00:00

Originally created by @mmdrahmani on GitHub (Jan 4, 2022).

Hi
I used gfp-gan on many different old photos and it works just fine. Except when women have scarf, and the photo is a bit low quality (e.g. eyes are in shadow), for these typically brown eyed people, the eye are blue-ish, or the relatively skin tone becomes mustaches on women faces. Really weird!
So I am wondering how I can work this out. I can send you a few photos if needed.

BTW, I was using colab-demo files and their default settings.

Bests
Mohammad

Originally created by @mmdrahmani on GitHub (Jan 4, 2022). Hi I used gfp-gan on many different old photos and it works just fine. Except when women have scarf, and the photo is a bit low quality (e.g. eyes are in shadow), for these typically brown eyed people, the eye are blue-ish, or the relatively skin tone becomes mustaches on women faces. Really weird! So I am wondering how I can work this out. I can send you a few photos if needed. BTW, I was using colab-demo files and their default settings. Bests Mohammad

claunia commented

2026-01-29 21:43:39 +00:00

@xinntao commented on GitHub (Jan 8, 2022):

Hi Mohammad，

Sorry that this algorithm could not well process all cases. It cannot deal with faces covered with other objects, such as scarves; it also does not perform well on photos with low quality and poor light conditions. And it may also have a bias (such as the eye color) due to the bias of training dataset of FFHQ.

Hope those issues could be solved with improved algorithms in the near future.

@xinntao commented on GitHub (Jan 8, 2022): Hi Mohammad， Sorry that this algorithm could not well process all cases. It cannot deal with faces covered with other objects, such as scarves; it also does not perform well on photos with low quality and poor light conditions. And it may also have a bias (such as the eye color) due to the bias of training dataset of FFHQ. Hope those issues could be solved with improved algorithms in the near future.

claunia commented

2026-01-29 21:43:41 +00:00

@mmdrahmani commented on GitHub (Jan 11, 2022):

Thanks a lot. Looking forward for newer versions of this interesting GAN.
Bests

@mmdrahmani commented on GitHub (Jan 11, 2022): Thanks a lot. Looking forward for newer versions of this interesting GAN. Bests

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: TencentARC/GFPGAN#141