Debuting alongside the new Blackwell GPU architecture, Nvidia gifted a remarkable new technology to owners of all existing RTX GPUs – the DLSS 4 transformer model. We’ve already talked about how the new ray reconstruction produces some outstanding results, but what about upscaling, or super resolution? Outlets like Hardware Unboxed have already put out excellent analysis of the new DLSS, so we went for a slightly different approach. Based on our years of testing games and isolating specific DLSS issues within these titles, we decided to go back and re-test DLSS pain points, swapping out the old convolutional neural network (CNN) model for the brand-new transformer alternative with a view to seeing how well the new technology copes in known trouble spots.
The concept of being able to improve existing games with brand-new DLSS technology is wonderful – but how do you do it? Well, new games offer the choice between CNN and transformer models within their menu systems, while others have been patched. Beyond that, Nvidia has added a function into its new app that does the same job. This is handy when it works, but sometimes it does not – like in Assassin’s Creed Shadows for example.
Thankfully, third party apps have been injecting new DLSS into older games for some time now, and DLSS Tweaker is a good alternative. With this you merely grab the latest DLSS super res DLL from a place like Tech Power-Up, drop it into the targeted game folder, click on DLSS tweaker’s config .exe and then change the DLSS version to preset model “K” – which is the latest transformer model. It’s slightly more convoluted but easy enough to get to grips with – but ultimately, Nvidia really needs to address the issue properly within its app.
In terms of our test suite – well, this is definitely a situation where watching is more illuminating than reading, so do check out the video above for actual head-to-head comparisons – but before we go deep into the weeds, let’s not forget that the transformer model has been proven to be more effective in terms of quality, albeit with a larger hit to resources than the CNN alternative. However, those fewer frames are traded for higher image quality, to the point where in many cases, you can comfortably lower the DLSS setting while still achieving the same – or better – image quality.
However, the focus of our testing is to see just how well the transformer model improves over known DLSS drawbacks and we start with Death Stranding, or more specifically the Director’s Cut, one of the earliest DLSS 2 games. We tested at 1440p resolution with the balanced mode and it’s obvious to see improved detail with a lower level of softness, without introducing aliasing. It’s a big improvement. However, there was a weakness: rain around Sam Porter, with DLSS virtually eliminating rain around his backpack – which is there in super-sampled ‘ground truth’ comparisons.
DLSS was taking those singular reflective points where rain drops hit that and most likely interpreting them as image noise or flicker, cleaning them away, although they are a part of the game’s art. However, transformer actually seems to make the situation worse: the rain drops are still not perhaps as bright as they should be, in fact, they are almost harder to see than they were before. It requires much more input resolution for the effect to resolve – as seen with DLAA. There’s a net boost to quality here but clearly, there are still limits.
In God of War Ragnarok, I originally noted how you can see trails coming from thin objects flush against the sky – and in motion it almost looks like those objects are ‘smoking’ a little. Here, I noted that the artefacting is still there but is significantly reduced with the transformer model. I saw a similar issue with Red Dead Redemption, and there’s still similar – if not markedly better – improvement here with the transformer model too. On the same game, I also noted dithering problems with hair – an issue that’s common to many titles, actually. The good news here is that once again, the transformer model offers a big improvement.
Coming over to Forza Horizon 5 – this game was updated with DLSS post release, and while it was a welcome addition it was not without issues. The first issue was with the telegraph wires above the tracks in the world. With DLSS on and using the CNN model, you would often see a break-up issue similar to God of War or Red Dead Redemption 2, but more intense. The wire in these games appears to be made of real geometry, and it is often sub-pixel nearly the entire time it is on-screen, especially at 1440p or lower. So DLSS would typically make a mess of them – flickering and with break-up, a stark reminder of the base resolution.
Flipping on the transformer model, this is still clearly a problem. Either CNN or transformer can look better depending on the content and both do not pass for native resolution rendering. This type of detail is extremely difficult to reconstruct and only by using MSAA at native resolution is the break-up taken care of – a stark reminder that Forza Horizon 5 was built with this kind of anti-aliasing in mind. By extension, super-sampling also works – a brute force form of AA.
However, Forza Horizon 5 does illustrate a strength of the transformer model – a lessening of soft detail in motion inherent to pretty much all forms of temporal anti-aliasing. This elimination of detail softening is one of the key advantages of switching to the transformer model, though it will be more or less visible depending on your display type – something like an OLED or strobing display will make the transformer model’s enhanced clarity during camera movement more obvious than an LCD or other display that has larger image persistence issues.
Next up, I was curious to check out how ray traced reflections resolve in the Nixxes ports for insomniac games, like Ratchet and Clank. There, when the ray traced reflections were set to high quality, they would be checkerboarded to save performance just like on the console version. The problem is that these reflections would resolve with big chunky pixels on any more mirror-like surface. Curiously, this only happens with DLSS and not with any of the other upscalers. With that in mind, this sounds like a game-specific problem as opposed to something off with DLSS – borne out by the fact that swapping over to the transformer model does not help.
The last game specific issue I want to look at concerns Dragons Dogma 2. At launch, I noticed how grass ghosted heavily in movement when the wind picked up. In this scenario, the grass tends to ghost into itself, looking smeary and unfocused, with all semblance of individual blades of grass completely disappearing. Flipping on the transformer model, there is a difference and I would say it is largely positive. Grass blades in stronger wind movement now keep their form much better and ghosting is eliminated. That’s great but as a negative side effect I think the grass now looks more strongly aliased when it does bend rapidly in the wind – which in aggregate looks fizzly – ghosting is not good of course, but it did mask aliasing due to the blur. Overall though, I would say this is a win for the transformer model.
Altogether going back and applying the transformer model over the legacy DLSS CNN model yields a lot of positive results, in some cases removing prior issues completely. However, it’s not a silver bullet or a complete cure-all for prior DLSS issues. The rain drops in Death Stranding, for example, suggest that sometimes the base resolution is still insufficient, or – in the case of Ratchet and Clank – DLSS is not being fed correct inputs. Another thing I noticed when applying the transformer model to older games is that there can be image quality regressions.
For example, in Control, I found that the transformer model has issues with the game’s ray tracing and working in combination with Jesse’s hair, adding extra noise over area lights. Presumably, the transformer model does not like the diffuse ray tracing here. Another issue I saw in many games was an increase in disocclusion issues over the previous model. For example, in Dragon’s Dogma 2, the game may look better overall with the new DLSS, but you can see that the area trailing the character’s head while running fizzles in the new model. In spite of its many faults, the older CNN model does not have the same problem.
The last issue I found is more easily seen in a title like Assassins Creed Shadows. The transformer model seems to have some issues with volumetric fog in general at times, making it so that objects transitioning into the fog ghost heavily, while the fog itself shows off a stippled ordered grid look. So, the transformer model is excellent and offers profound improvements in many categories but it also has issues at the moment which prevent it from being appropriate for all games.
That said, this is the very first iteration of the new DLSS transformer model, so I would expect improvements. Nvidia says that the older CNN technology has effectively run its course, with only iterative improvements since its 2020 debut – while the sky’s the limit for the transformer technology. Alongside further improvements to ray reconstruction and frame generation, we’ll be following super resolution’s continued progress with much interest.
