Leaked Insomniac Slide Suggests AI Upscaling For PlayStation 5 Pro
It's for that reason that we've largely declined to report on the leak after our original coverage, but this story isn't really about the leak. Instead, it's about the idea that Insomniac may have been working on implementing machine learning-based image upscaling for its upcoming Wolverine game. A slide that was supposedly included with the leaks describes a number of technologies that were apparently intended to appear in the game.
There's only one real surprise here; Ziva is a machine-learning-based model deformation system that the company first implemented in Spider-Man: Miles Morales. Most of the rest of the items on the list are unremarkable, too: real-time global illumination is an industry standard at this point, and hierarchical destruction has been around a while now. We must say it's nice to see the company targeting 60 FPS with ray-tracing enabled, though. The entry two items down is likely part of how the company would do that: "AI upscaling (Machine Learning)."
Insomniac is a Sony studio, and so it primarily develops games for PlayStation platforms. There is no PlayStation with hardware AI acceleration, so what does this remark mean? Well, it's been rumored for some time now that Sony has been seeking to implement AI accelerators into its mid-generation PS5 Pro refresh. Given that, it's entirely possible that this slide (which originates from late 2022) is talking about AI upscaling performed on new hardware in the PS5 Pro.
However, we have to note that it's completely possible to do AI-based image upscaling on the extant PlayStation 5 (or Xbox Series X); a solution like Intel's XeSS that relies on generalized GPU compute could certainly work. It would consume a considerable portion of the available GPU compute resources, though, and it's not completely clear that it would give superior results to AMD's FSR2, which is known to work well on the PS5.
Comparing the results on PC hardware, Intel's DP4a path for XeSS typically gives a softer, less-detailed output than FSR2, but the trade-off is that you typically end up with a much more temporally-stable image. Neither are as good as NVIDIA's DLSS, of course, but DLSS relies on the presence of NVIDIA's Tensor Cores and thus can't be used on consoles. It's possible that Insomniac, Sony, or perhaps even AMD itself are cooking up something similar as we write this, though.
