Post

DLSS 5 is just a lighting pass and not an ai hallucin--

2.5K reposts8.3K likes

this reminds me of when somebody posted how AI video overlays can "make so many interesting styles" and then showcased a horrifying Skinnamarink Mario Slide

At this point, the line between "lighting pass" and "hallucination" is just a marketing budget. We’re basically paying for a GPU that dreams the pixels into existence and calling it optimization.

This is result of positive feedback loop. It feeds the previous output frame as new input. It is hallucination and it reveals the strongest weights, aka what the DLSS 5 is "seeing" when you give it a frame from a game. It explains why DLSS 5 "enhanced" characters look much older and wrinkled.

Its not scrambling the menu tho, its just continuing to scramble the actual game frames. but there are no new frames bc the game is paused so it uses its own output again. and if devs made it so dlss5 is disabled during pause then the original frame would show up, unscrambled, un-"beautified"

this implies that the model is being like recursively fed each previous frame without reference to the groundtruth. which seems insane to me because it can just enter a hallucinatory death spiral like this with no recourse. what a stupid half-baked product

I imagine that's because it's a pause screen. Some games will just take a frame buffer capture of the current view of the game to show behind the pause menu UI, especially if the style of the UI is to post-process the image, like blurring it and making it black and white.

So I guess DLSS is operating on the framebuffer image the game is using, which... would be the output of the previous frame that DLSS already processed. Hence, hallucinatory death spiral. That's not to say DLSS 5 isn't terrible. It definitely is.

i hate to be that guy but this version of DLSS5 is a leak, and a game that actually supports DLSS5 would, I assume, be playtested and be integrated deeper into the rendering process in general (Nvidia showed things like masking out certain parts of the image to be more or less slopped up)

Honestly, AI was kind of fun back then. I enjoyed the bizarre slop of 2022–2023 far more than most of what it produces today. I wonder if there’s a way to use code to essentially “lobotomize” an AI engine back into doing things like that? 

This is more or less precisely how that worked - Deep Dream was a debug tool for image recognition that worked by nudging an image toward what it 'thought it saw' so engineers could 'see' what it was recognizing as what. The first 'gen ai' was feeding noise images to Deep Dream so it could 'imagine'

humbly requesting an aftermath article verifying the provenance of this and explaining wtf is going on here which I can send to people <3