I hate how AI upscaling looks and I really don’t get why everyone seems to be gaga over it. In addition to the artifacts and other weirdness it can introduce, it just looks generally like someone smeared vaseline over the picture to me.
I’ve tried upscaling with ESRGAN as well and it has similar problems. It messes with the original textures too much. For example, it made carpet look like a solid surface. Skin looks too smooth and shiny. That kind of thing.
It depends a lot on the source picture, but it’s definitely not a general problem inherent to AI upscaling. Otherwise there wouldn’t be so many positive examples of ESRGAN.
DLSS isn’t like all the other upscalers, it’s on a whole different level. FSR is a blur filter. FSR2 is better, but still noticeably upscaled with tonnes of artifacting. Same with XeSS, because that and FSR are just software upscaling.
DLSS on the other hand has actual hardware that is dedicated to it. It actually gives better than native results quite often. It doesn’t at all look like someone smeared Vaseline on the screen.
FSR/XeSS are basic sharpening tools, and yeah they are inherently limited because it’s just an impossible thing to do with 100% accuracy. DLSS is the same thing except NVIDIA tries to circumvent this limitation through some kind of proprietary AI magic, accelerated via their hardware. It’s impossible for it to be “better than native”, it’s using AI to approximate what “native” is. And in doing so, it makes the original image look too different to my liking. In motion the textures definitely look a little muddied to me as things blend into each other since the AI cannot accurately predict how things should look in realtime. At that point I’d rather just use FSR/XeSS as it at least preserves the original art style.
I hate how AI upscaling looks and I really don’t get why everyone seems to be gaga over it. In addition to the artifacts and other weirdness it can introduce, it just looks generally like someone smeared vaseline over the picture to me.
That’s not inherent to “AI upscaling” as a process. ESRGAN for example is pretty good at upscaling pictures while keeping the quality.
I’ve tried upscaling with ESRGAN as well and it has similar problems. It messes with the original textures too much. For example, it made carpet look like a solid surface. Skin looks too smooth and shiny. That kind of thing.
It depends a lot on the source picture, but it’s definitely not a general problem inherent to AI upscaling. Otherwise there wouldn’t be so many positive examples of ESRGAN.
DLSS isn’t like all the other upscalers, it’s on a whole different level. FSR is a blur filter. FSR2 is better, but still noticeably upscaled with tonnes of artifacting. Same with XeSS, because that and FSR are just software upscaling.
DLSS on the other hand has actual hardware that is dedicated to it. It actually gives better than native results quite often. It doesn’t at all look like someone smeared Vaseline on the screen.
FSR/XeSS are basic sharpening tools, and yeah they are inherently limited because it’s just an impossible thing to do with 100% accuracy. DLSS is the same thing except NVIDIA tries to circumvent this limitation through some kind of proprietary AI magic, accelerated via their hardware. It’s impossible for it to be “better than native”, it’s using AI to approximate what “native” is. And in doing so, it makes the original image look too different to my liking. In motion the textures definitely look a little muddied to me as things blend into each other since the AI cannot accurately predict how things should look in realtime. At that point I’d rather just use FSR/XeSS as it at least preserves the original art style.
It’s not impossible for it to be better than native.
https://www.techspot.com/article/2665-dlss-vs-native-rendering/