What is “AI upscaling” in video game development?
Last Updated: 02.07.2025 00:44
There are landing aid lights which are super-bright so you can see them from a very long way away. One type of these lights are on the ends of the runway and they are designed to show whetherer you’re approaching the runway on the right “glide slope” - so they change color depending on how high/close you are to the touchdown point.
That’s a hell of a lot of math - and as display resolutions and frame rates get better - the problem rapidly escalates.
The AI might MAYBE notice that the light was there on the previous frame and “fill it in” - but what happens if you went off the glide slope at that moment. The light SHOULD change color - but because the AI doesn’t know anything about the glide slope - or the particular angles at a specific airport - it’s highly likely to screw up.
Brain ‘Reset Wave’ May Explain How ECT Rapidly Relieves Depression - Neuroscience News
So if we render the scene at reduced resolution - we might not draw that critical light point at all because it fell into one of the gaps.
So usually - if you’ve got an 8 MegaPixel screen - the graphics card has to calculate the color of every single one of those 8 million pixels (maybe more with anti-aliasing) - probably 70+ times per second.
These light points are tiny - they only just cover a single pixel.
But there are issues.
I’ve been working in Flight Simulation -where this problem is massively exacerbated by clients who have maybe 16 display screens - all at some crazy high resolution!
However, consider a common thing people to with flight simulators…practicing landing at night.
Sam Darnold on final games in Minnesota: "We laid an egg on offense" - NBC Sports
So using this technique could be very handy.
But that’s a kinda special situations - and we could maybe handle it by drawing those specific lights AFTER the AI had filled in all the gaps.
I’ve seen demo’s from Nvidia that do this AMAZINGLY well.
No One in the Movies Stays Dead Anymore - Yahoo
But the point remains - as it does with ALL AI - that you can’t really trust it not to do something very strange.
That’s A VERY BAD THING!!! When you’re trying to train a pilot to react correctly to those lights and they’re being displayed incorrectly - you can end up doing what we call “Negative training” - where you make the pilot WORSE - not better!
So the idea is that you train an AI to look at just (say) 1 million pixels - and try to guess what the others should be.
Dozens of Israeli children develop malignant tumors from sperm donor - The Jerusalem Post