Corrects color channels in face restoration; Fixes #1167 #1175
Conversation
psychedelicious
requested review from
blessedcoolant,
hipsterusername and
lstein
|
I guess there's no harm in using cv2 for this if it's already a project dependency, but does Pillow have any modes that work for this built-in? Maybe https://pillow.readthedocs.io/en/stable/handbook/concepts.html#modes |
|
Great find! |
hipsterusername
left a comment
hipsterusername
left a comment
There was a problem hiding this comment.
Lgtm! Seems like a simple whoops😅
Good catch - I tried "BGR" first and it didn't work, shoulda checked the docs. I'll give this a whirl. |
lstein
left a comment
lstein
left a comment
There was a problem hiding this comment.
Congrats on finding and fixing the problem!
|
We already use cv2 so there's no particular advantage to doing this in Pillow, but if you want to, go ahead. |
|
I had blue tint on the final rendered image when using restore face on a M1 Mac. After research and messing around I found this solution. https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/codeformer.pth to /stable-diffusion-webui/models/Codeformer/codeformer-v0.1.0.pth https://github.com/xinntao/facexlib/releases/download/v0.1.0/detection_Resnet50_Final.pth to /stable-diffusion-webui/repositories/CodeFormer/weights/facelib/detection_Resnet50_Final.pth https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/parsing_parsenet.pth to /stable-diffusion-webui/repositories/CodeFormer/weights/facelib/parsing_parsenet.pt |
6 participants
Our Codeformer and GFPGAN implementations are sending RGB image data to the restoration code, but they expect BGR image data (as used by OpenCV).
The fix is simple; use a cv2 method to convert the image from RGB to BGR before processing, and then back to RGB so we can continue using PIL.
This fixes an until-today unknown flaw in our GFPGAN implementation, which was recognized by user JoshuaKimsey on discord.