The meme is not even wrong. If the "original game's artistic intent" is not included in the teaching dataset AI model is likely to hallucinate one. Lately I've greatly enjoyed "Juice Galaxy" and I do not think a AI filter would improve it. (It's early access on Itch and Steam with free demo available. If €$£ 5 is a lot to you don't buy yet, it needs a lot more content.)
The meme is not even wrong. If the "original game's artistic intent" is not included in the teachin
This actually would work well if you could provide "per-game neural network weights" -- you could make expensive photorealistic renders in the studio - or even record actors (where applicable for a photorealistic title) and finetune the network just on that data, then ship those weights with the game.
The issue is, that's expensive and time consuming. If you instead want a "magic wand" that just makes every game high res without effort -- then you end up with shitslop.
Also for finetuning the network architecture and training code would have to be open.
This actually would work well if you could provide "per-game neural network weights" -- you could ma
Teaching 'all human knowledge' to neural network model is expensive but limited knowledge like 'intended high quality rendered look of a game' might actually be in reach of game studios, at least for those wasting $100000000s for boring duds. This Corridor Crew video about neural net applied in a different cinematic graphic problem might be of interest: It Took Me 30 Years to Solve this VFX Problem
Teaching 'all human knowledge' to neural network model is expensive but limited knowledge like 'inte
