This has been an issue for quite a while now, not just with Topaz’s products but with all the face recovery tools out there. Free or paid ones.
We see similar artifacts with open-source tools like Codeformer or GFPGAN, and even in Topaz’s products. It seems like most face-recovery tech is built on similar foundations, leading to these persistent “beauty-spot” artifacts.
I’d love to see Topaz Labs focus on addressing this and improving how these beauty-spots-looking type of artifacts.
This would set Topaz Labs even more appart from the rest of the free and paid solutions out there. If they find a way to fix this problem that I can see almost everywhere.
It’s strange because I’m sure that Topaz Labs and other companies are training their own models, so the issue might lie in the code base itself. Or maybe it’s a combination of training on faces with freckles and the way the code applies the model, which brings back these artifacts in certain cases.
I’m just speculating here of course I don’t know how this technology works. But I see the same pattern accros the board so that’s why I say that there must be a common base that is creating that.
Topaz face recovery
here with Codeformer-v0.1.0.pth
And here with GFPGanv1.4.pth
It looks like Topaz Labs is using something similar to GFPGan because the beauty-spots have the same pattern.
Note that this problem occurs of most images with compression artifacts or anything that is not originally smooth on the skin. This is not just one extreme example.
Here the original image
