/jp/ - Otaku Culture

>>43722100
There are certainly more efficient upscaler models by now, but the main thing is they work differently, rather than "better." Some of them try to regenerate texture, which can result in JPG noise being treated as such and generating detail where there's not supposed to be any. Others will blur it instead, but may also blur details you may want to preserve.

So, for example, here's a tiny but very smug spooder upscaled 3x with R-ESRGAN 4x+ and 4x-UltraSharp. You'll note the "texture" (actually a scan artefact) on her shirt is flattened in the ESRGAN sample, but preserved and exaggerated by UltraSharp. On the other hand, her eyes look blurry, almost watercolourish in the ESRGAN sample, but retain some of that paper texture in UltraSharp.