Comment by viktorcode
3 days ago
This video is a direct continuation of the one where Cerny explains logic behind PlayStation 5 pro design and telling that the path forward for them goes into rendering near perfect low res image then upscaling it with neural networks to 4K.
How good it will be? Just look at the current upscalers working on perfectly rendered images - photos. And they aren't doing it in realtime. So the errors, noise, and artefacts are all but inevitable. Those will be masked by post processing techniques that will inevitably degrade image clarity.
It only takes a marketing psyop to alter the perception of the end user with the slogans along the lines of "Tired of pixel exactness, hurt by sharpness? Free YOUR imagination and embrace the future of ever-shifting vague forms and softness. Artifact stands for Art!"
I’m replaying CP2077 for the third time, and all the sarcastic marketing material and ads you find in the game, don’t seem so sarcastic after all when you really think about the present.
If you think those are uncanny, wait until you hear the ads in GTAV.
1 reply →
I don't know, I think it's conceivable that you could get much much better results from a custom upscale per game.
You can give much more input than a single low res frame. You could throw in motion vectors, scene depth, scene normals, unlit color, you could separately upscale opaque, transparent and post process effect... I feel like you could really do a lot more.
Plus, aren't cellphone camera upscalers pretty much realtime these days? I think you're comparing generating an image to what would actually be happening.
> I think it's conceivable that you could get much much better results from a custom upscale per game.
> You can give much more input than a single low res frame. You could throw in motion vectors, scene depth, scene normals, unlit color, you could separately upscale opaque, transparent and post process effect... I feel like you could really do a lot more.
NVIDIA has already been down that road. What you're describing is pretty much DLSS, at various points in its history. To the extent that those techniques were low-hanging fruit for improving upscaler quality, it's already been tried and adopted to the extent that it's practical. At this point, it's more reasonable to assume that there isn't much low-hanging fruit for further quality improvements in upscalers without significant hardware improvements, and that the remaining artifacts and other downsides are hard problems.