Costs time Costs kudos Repairs small faces, but makes them look more generic
Small faces are where generation falls apart first: at a distance there are not enough pixels for the model to place eyes and a mouth convincingly.GFPGAN and CodeFormers rebuild them from what a face is generally supposed to look like. That is also the catch, because what they know is faces in general rather than the one in your picture.It does nothing for a face that is already large and correct.
Where the difference is
The same picture before and after, close up The same patch cut from both pictures, drawn one picture pixel to one screen pixel.
