First let me make the short tangent about what I want to talk about:
Upscaling of images refers to computer programs that increase the image quality of photos. They generally come in roughly 3 flavors:
- Resolution increases: Just cutting all pixels in 4. A worthless technology from 20 years ago
- Vector based upscaling: besides increasing the resolution these upscaling methods detected vectors of hard lines/edges or color transitions which sometimes even deals with some smaller artifacts
- AI based upscaling which invents additional detail out of thin air into images
I have been interested in the later since roughly 2016. We digitized old family photos made on film which look wonderful. But unfortunately in the early 2000s we switched to early digital cameras that look like potatoe. But I'm also interested in making my current iPhone photos even better. I'm sure, you can think of a reason why you'd be interested in this technology as well.
The current situation
One of the great technologies in this field are DLSS by Nvidia and FSR by AMD. They're more targeted towards live footage of gaming instead of single images.
There are consumer products like waifu2x or upscayle.
Then there are tools like real_esrgan or python ISR.
All of them have in common that they do blow up the file size tremendously, do increase the quality a bit but don't really invent much detail and do give everything the cartoonish AI look.
What now?
Sure, maybe it was too much to ask for to turn my 4mb iPhone photos into 30mb full frame professional camera photo. But a 150mb waifu cartoon image seems weak
Am I the only one frustrated by this? While generative AI for text or brand new art makes leaps the upscaling lacks behind.
If there is no demand in the world for this, it will never improve.