Comment on YouTube secretly used AI to edit people's videos. The results could bend reality
CeeBee_Eh@lemmy.world 3 weeks agoThey don’t require AI neural networks.
Sharpening and denoising don’t. But upscalers worth anything do require neural nets.
Anything that uses a neural network is the definition of AI.
ccunix@sh.itjust.works 3 weeks ago
Not true
Company I used to work for had excellent upscalers running on FPGAs that they developed 20+ years ago.
The algorithms have been there for years, just AI gives it bit of marketing sprinkle to something that has been a solved problem for years.
Probius@sopuli.xyz 3 weeks ago
Depends on what you’re trying to upscale.
CeeBee_Eh@lemmy.world 3 weeks ago
Well, the algorithms that make up many neural networks have existed for over 60 years. It’s only recently that hardware has been able to make it happen.
Not true and I did say “any upscaler that’s worth anything”. Upscaling tech has existed at least since digital video was a thing. Pixel interpolation is the simplest and computationally easiest method. But it tends to give a slight hazy appearance.
It’s actually far from a solved problem. There’s a constant trade-off beyond processing power and quality. And quality can still be improved by a lot.
rmuk@feddit.uk 3 weeks ago
Right. Even back in the eighties UK broadcasters were upscaling American NTSC 480i60 shows to 576i50. The results were varied. High-ticket shows like Friends and Fraiser looked great, albeit a bit soft and oversaturated, while live news feeds looked terrible. If you’ve never seen it, The Day Today has a perfect example of what a lot of US programmes lookd like converted to PAL.
CeeBee_Eh@lemmy.world 3 weeks ago
Ya, I knew there were analogue “upscalers”, but I’m not familiar enough with them to confidently call them an upscaler vs a signal converter.