| shikamaru317 said: This looks so bad, like a bad AI phone filter. It also ruins the creative vision that the character artists on these games had for these characters. And worst of all, AI slop like this is part of why it costs $370 bucks to buy 32 GB of DDR5 RAM (up from about $90 in 2024) and like $150 to buy a 1 TB SSD (up from about $70 in 2024), part of why PS6 will be $800+ and part of why Project Helix will be $1000+, because the AI datacenters that power this kind of slop are taking all the RAM and NAND memory, causing supply and demand issues because we can't build new fabrication facilities fast enough to accommodate the increased demand from AI. |
So for it to be able to inference in real-time these models need to be a lot smaller than what you find in video generation or image generation models, about two-orders of magnitude smaller (think 20-30 million parameters vs. 12 billion +.)
The datasets are also a lot smaller, and synthetically produced (provided by game developers, with motion vectors available as inputs.)
In terms of actual compute costs, training DLSS features is a drop in a very large bucket, and Nvidia has even experimented with online training (continuously training tiny special-purposed models as you play the game on your hardware.)
The synthetic data production probably is more computationally expensive than the actual pre-training.
Last edited by sc94597 - 5 hours ago
