SD 2.0 at 768 has made ‘restore faces’ almost completely unneeded

[eadnams22]

Just testing things out generating headshots for communications and marketing persona documents, and noticed I really don’t need to run restore faces anymore, unlike in 1.4 and 1.5.

Eyes and teeth are way more consistent, even with simple prompts. Not perfect, but often very usable, and more consistently so.

nice to not have the over-blurry face restoration needed every time.

[eadnams22]

I’ll add, that this makes sense given the model is trained at a high resolution, allowing clearer small details to be visible.

[Polyphron]

Well, have you tried anything that isn't a closeup?
My hope is that we can finally do zoomed-out shots with the higher resolution supported by codeformer.

[eadnams22]

I have, and yeah, even those look better more often, but still not quite there.

[mykeehu]

The distant faces remain distorted, so we need that face restoration