Without seeing the original I'm not exactly sure what the problem is. I can see that it isn't correct, but I can't see all the ways that it is messed up without some kind of reference point.
If I had to take a shot in the dark, maybe alpha is being lost or handled incorrectly when upscaling. I work around this in CTT-PS when using ESRGAN or SFTGAN with the "Preserve Alpha Channel" option, which extract the alpha channel, upscales it separately, and merges it with the result. The downside is it doubles the processing time.
If I had to take a shot in the dark, maybe alpha is being lost or handled incorrectly when upscaling. I work around this in CTT-PS when using ESRGAN or SFTGAN with the "Preserve Alpha Channel" option, which extract the alpha channel, upscales it separately, and merges it with the result. The downside is it doubles the processing time.