Just one implementation in one game - but the results are compelling.
Face-off by Alex Battaglia Video Producer, Digital Foundry
Published on
73 comments
With PlayStation Spectral Super Resolution (PSSR), Sony has joined the machine-learning based image reconstruction race started by Nvidia almost six years ago. At Digital Foundry, we're excited to see this kind of technology reach into the console space and so as part of our recent exclusive PS5 Pro capture opportunity, we spent some time using PS5 Pro and Ratchet and Clank: Rift Apart to grab a series of reconstruction-based shots that we know from prior experience truly puts reconstruction technologies through their paces. To what extent can PSSR best the FSR-level technology typically used by consoles? And how close is PSSR to the acknowledged state-of-the-art: Nvidia DLSS?
First of all, let's discuss caveats and limitations. We compared PS5 Pro to PC on approximate quality settings, but it's impossible to get a complete match. Secondly, Rift Apart uses dynamic resolution scaling and the PC implementation is very, very different - so we pixel-counted each individual test shot on the Pro, then matched resolution precisely on DLSS using a tool called DLSSTweaks, whereas we were limited to FSR 3.1's quality mode, which typically uses a slightly lower base resolution than the Pro. Beyond that, while we tried to disable motion blur on PS5 Pro for increased clarity, turning it off only seems to lower the intensity of the effect, so we had to match that on our PC captures.
Finally - and perhaps the biggest caveat of all - this is just one game and the quality of upscaling implementations vary from title to title. We've not seen enough of PSSR to draw more solid conclusions in the way we can with FSR and DLSS, but there's a good chance that what we're seeing here with PSSR may be indicative. For now though, it's very much just a Ratchet and Clank head-to-head.
I'd strongly recommend watching the video above as the very nature of the exercise means that video is a better medium to communicate the differences, and as you'll see when you watch, comparisons up against FSR 3.1 were the first order of business. AMD's compute-based upscaler is heavily used in the console space and in many scenarios, this is the technology PSSR looks to supplant. I began by looking for 'disocclusion fizzle'. A third-person game like Rift Apart is great for testing this: it's essentially the quality of newly revealed detail which has little to no 'history' to draw on from previous frames.
This is my major complaint with FSR 3.1's quality - it fails to effectively anti-alias newly revealed detail, 'fizzling' in motion. Thankfully, like DLSS and XeSS, PSSR does not suffer from this issue - and it's as plain as day to see. Ratchet and Clank runs at relatively high resolutions on PS5 Pro, so this gap in quality may widen when dealing with much lower quality input pixel ranges - something we'll test in due course.
FSR also has problems with particles, whether it's the particles when Ratchet collects his first gun, or the fluttering confetti in the background during the opening stage. Particles have a wispy, ghost-like look when processed with FSR, to the point where the confetti can fade in and out of existence. PSSR does a comparatively superior job and actually seems to present with more confetti. That's not actually the case: it's just that PSSR particles are more temporally consistent, remaining on-screen in stark contrast to FSR 3.1. Hologrammatic signs also look better with PSSR. FSR 3.1 lacks anti-aliasing coverage here and while PSSR is not perfect, the presentation is less jagged on edges.
Finally, processing grass in the first stage is challenging for FSR, to the point where it can seemingly present flowing artefacts that look almost liquid-like. Machine learning-based upscalers tend to discriminate better and PSSR tries to clamp down that issue. It still is not perfect, but I would say it does manage to keep the grass from looking as odd as it does with FSR. If we consider that this is the first publicly seen iteration of PSSR, we can safely assume that it manages to do the job required of it - it offers a substantial upgrade to upscaling quality, but do bear in mind we're still dealing with relatively high base resolutions. It's at 1080p - and even lower - where PSSR truly needs to prove its worth.
Comparisons against Nvidia DLSS perhaps have less practical use. Unlike FSR 3.1, you will never see a Sony console using this upscaling technology. However, more holistically, DLSS holds the crown as the most accomplished machine learning-based upscaler on the market - a position Nvidia has earned with almost six years of iteration - and so we can get some idea of how close Sony has got with its debut iteration. Across all my test clips, I noted a kind of moving fizzle. I am not exactly sure what it is - some sort of meta instability, perhaps - but it is found across every image of the game I have looked at so far. It's not there in FSR 3.1 either, to be fair, and I also found that in some areas, DLSS resolves moving geometry with less aliasing.
This also bears out when you compare a rapidly moving object. PSSR does a better job than FSR in movement, but against DLSS it does not anti-alias as well. Another thing you can notice about PSSR is how it produces a softer image than DLSS. Nvidia's upscaler no longer has an inherent post-process sharpener that is built into the model anymore, instead only having games add it if they want it. And even with post-process sharpening off, I think we can see how the resolve of PSSR is noticeably softer. This can be viewed as a detriment, but I found it highly preferable to instances where Ratchet's in-game post-process sharpener is ramped to the max. Of course, the resolve of a reconstruction technique is essentially subjective and a matter of personal taste.
Less subjective is something curious I noted - ray tracing ghosting on the PS5 Pro version of Ratchet and Clank, seen in its RT reflections. Ray tracing meshing with image reconstruction is complex, and the noise pattern input into the reconstruction can have a huge impact on the output image. Insomniac says that the sampling pattern is customised for PSSR to better mesh with it. Even if the reflections in movement have negative aspects, there can be better stability in certain aspects. For example, on lower quality settings, checkerboard RT reflections clearly shows checkerboard artefacts with DLSS. PSSR merges the checkerboard correctly, delivering a higher resolution output.
Beyond these differences, we will need more time with PSSR in this - and other - titles to deliver a more definitive verdict. Based on today's testing, there are areas where PSSR is vastly superior to FSR 3.1, the technology it is most frequently likely to replace and to the eye, it's easily preferable. Assuming these differences translate to other games, PS5 Pro owners can appreciate significantly improved upscaling quality. The issues with stability are something to watch in the future, to see if it presents in other games, and if it's not, whether it's improved in future iterations of PSSR.
But perhaps it's games like Alan Wake 2 that will best prove its worth: an internal 864p in its 60fps performance mode, upscaled to full 4K is a daunting task for the tech - and we're looking forward to testing this, along with other titles that have perhaps pushed FSR technology too far on the standard PS5 and Xbox consoles. The PS5 Pro's November 7th release date draws closer, and we'll have more coverage as soon as we can.