Ghostwire: Tokyo is now out on PC and PS5 and contrary to what the official system requirements would have you believe, it’s a fairly decent port. Sure, there are drops here and there, but the overall performance is quite satisfactory. In this post, we’ll explore the various upscaling technologies included in Ghostwire. These include DLSS 2.3, FSR 1.0, and TSR from Unreal Engine 5. Along with Intel’s XeSS, these represent the primary image upscalers of this (and future) generation. It’s worth noting that AMD’s newly announced FSR 2.0 upscaler is more or less identical to TSR so you can think of the two as different variants of the same algorithm.
Without any further ado, let’s begin. Above you have a comparison of the balanced presets of DLSS 2.3 and FSR 1.0. As expected, a purely spatial upsampler like FSR 1.0 doesn’t stand a chance against DLSS 2’s temporally fed AI-upscaler. There’s plenty of loss in texture and mesh detail as well as partial lighting scaling. DLSS does an exceptional job, retaining nearly every bit of detail, both small and big.
Moving to TSR, the image quality is virtually identical to DLSS. On very close inspection, it can be seen that TSR trades a bit of sharpness for a smoother image while DLSS does the opposite. Overall, it’s a tie.
Once again, FSR 1.0 proves to be an old and ineffective upscaler compared to DLSS 2 and TSR.
TSR is incredibly accurate at upscaling highly detailed images as can be seen in these examples. At both the balanced and performance presets, it not only keeps up with DLSS but outright beats it in some cases. It’s remarkable how far video game graphics technologies have come. Even the performance preset which is as expensive as 1080p produces native 4K-grade image quality.
DLSS and TSR continue to impress the closer you look. Even at the performance preset, we’re getting near-flawless upscaling. At the end of the day, the only difference between the two is that the former prioritizes texture sharpness and detail which results in a bit of aliasing along thinner object edges. TSR, on the other hand, produces a smoother image that is better anti-aliased at the cost of an ever-so-slight loss in detail.
These comparisons make you realize one very important fact. TSR manages to achieve the same level of quality that NVIDIA took more than three years to achieve with DLSS 2.x with its very first implementation. Really makes you question the role of that neural network that has been supposedly trained using hundreds of thousands of images. At the very least, it isn’t conferring on DLSS any meaningful advantage over TSR (and very soon FSR 2.0) which is a “simple” temporal upscaler.
When FSR 2.0 releases next quarter, it’ll be hard to recommend DLSS 2.x if both produce the same quality output image. The former will work across on pretty much any modern GPU while the latter will be limited to a dozen products from one vendor. Unless NVIDIA releases the source code for DLSS, it’s likely going to die a slow death like most Gameworks technologies.
You can view the full-sized 1080p images below (click on each to view in another window).