Nvidia DLSS 4 explained: Multi Frame Generation, the Transformer model, and the GPUs you’ll need to run them
Testing out the biggest DLSS overhaul yet
Nvidia DLSS 4 has launched under not-terribly-happy circumstances. It’s a mostly AI-powered technology at a time when mistrust in artificial intelligence, fuelled by underbaked applications and anti-creative policies, is at all-time high – not to mention how the new GeForce GPUs it’s released alongside, the RTX 5080 and RTX 5090, have immediately entered 2020-style stock shortage purgatory.
However, having tried it out in a few different games, I suspect time will prove kinder to DLSS 4 than the RTX 50 series launch has been. It is, in fact, a rather nifty collection of upscaling improvements that can help out older graphics cards as readily as the very newest, while building on DLSS 3’s frame generation tool with Multi Frame Generation (MFG) to send visual smoothness skyrocketing. If you care even the slightest iota for how your PC performs in modern games, DLSS 4 demands your attention.
Here, then, I’ll break down DLSS 4’s various tricks and components, looking at how they really affect performance – and whether it’s worth upgrading your rig to maximise compatibility with it.
- DLSS 4 part I: What "DLSS 4" actually means, and what can run it
- DLSS 4 part II: Upscaling image quality and the Transformer model
- DLSS 4 part III: Performance and Multi Frame Generation
- DLSS 4 part IV: The latency issue
- DLSS 4 part V: The verdict
DLSS 4 part I: What "DLSS 4" actually means, and what can run it
DLSS version names got a bit tangled up in themselves during the previous generation – DLSS 3 was all about the frame gen, then DLSS 3.5 showed up with Ray Reconstruction but nothing else to do with DLSS 3’s defining feature. This turned the simple act of talking them into a minefield of mismatched acronyms, so before anything else it’s worth defining what exactly DLSS 4 is.
The easiest (and, fortunately, most accurate) way of looking at it is that DLSS 4 is the umbrella term for a collection of GPU-based upscaling and performance features. In DLSS 4’s case, that includes the latest edition of its classic upscaling, as well as new renditions of tech that was previously referred to simply by the DLSS version that introduced it. MFG, for instance, is the next-gen take on DLSS 3 frame generation, and DLSS 4 also includes an improved version of DLSS 3.5’s Ray Reconstruction for luxury ray tracing/path tracing effects.
This does mean that statements like "Game X supports DLSS 4" might not be entirely accurate, as DLSS 4 covers so many different components that individual games likely won’t always work with the full set: they might support its upscaling and MFG, for instance, but not Ray Reconstruction. Thus, some confusion – and twisting, multiple-clause explanations by tired hardware editors – will remain. But there is an upside, namely that a lot of DLSS 4’s features will function on older RTX GPUs dating back to the original RTX 20 series.
Indeed, of all DLSS 4’s biggest doodads, only MFG outright requires an RTX 50 series graphics card. The improved upscaling, Ray Reconstruction, and the Nvidia app’s DLSS overrides (more on those later) are all backwards-compatible with the RTX 20 series onwards. You’ll still need at least an RTX 40 series GPU to use frame generation at the same speeds of DLSS 3, though DLSS 4 also includes improvements for this, reducing memory usage and potentially making a slight increase to frame output. Just not as much as MFG, obviously.
DLSS 4 part II: Upscaling image quality and the Transformer model
Image quality is hardly one of DLSS’s traditional weaknesses. Arguably the biggest single reason to use it over alternatives like AMD FSR or Intel XeSS is that it simply looks better: its AI-powered anti-aliasing produces the cleanest, sharpest edges and the fewest blurry textures among the big three upscalers. Even so, this component has been touched up further in DLSS 4, especially in games that let you switch from the previously employed Convolutional Neural Networks (CNN) AI model to its new Transformer model.
Nvidia’s explanation for how these models differ is several mouthfuls of impenetrable, "AI is the future bro" turbojargon, but the very basic version is that the Transformer model is smarter about where it places individual pixels when piecing together an upscaled frame. It’s slower than CNN on framerate enhancements – by about 9-10%, according to my testing on the RTX 5080 and 5090 – but within seconds of switching to Transformer mode in Cyberpunk 2077, the visual improvements became apparent. Including, yes, the ones that Nvidia promised. Reduced ghosting? Subtle, but yes. Higher detail in motion? Definitely. Smoother edges? Yep, especially on very fine details like chainlink fences or individual blades of grass. DLSS upscaling has nearly always looked good; DLSS 4 makes it look better.
I will say that these improvements are much more noticeable in motion than in static screenshots, so I don’t want to dwell on too many side-by-side, its-the-same-picture comparisons. But even here, you can see some of the Transformer model’s upsides, with less visual noise and finer detailing on the tape around V’s shotgun. Zoom in closer and you might notice less jagged edges than on the CNN model as well.
The Transformer model also has the knock-on effect of improving how frame generation looks, as the more intelligent pixel placement technique applies to AI-generated frames as well as the ordinarily rendered ones. Again, fast motion looks cleaner and sharper, with less ghosting, and that goes for ye olde DLSS 3 as well as the latest, RTX 50-exclusive MFG. Speaking of which…
DLSS 4 part III: Performance and Multi Frame Generation
I suspect that MFG is already the most contentious of DLSS 4’s constituent parts. There’s definitely a perception among some PC players that the advantages of frame generation are "fake", and there is some truth to that in the tech itself: AI-generated frames don’t take your control inputs into account, as rendered frames do, so a game that uses frame gen to turn 30fps into 60fps will look nicer but still have the 'writing your initials in unset concrete' feel of the former.
Still, if you approach DLSS 3 – and now DLSS 4’s MFG – with the full knowledge and understanding that frame gen is a purely visual enhancer, then both become a lot easier to get on with. And in these simple terms of raising frames per second, MFG is unmatched. It has the option to generate extra frames at a 2x rate, as DLSS 3 does, but also 3x and 4x, and while that fastest option might not have literally quadrupled the framerates of the games I’ve tested, it did punt them so high upwards that it’ll probably be another two graphics card generations before they’ll reach those heights through traditional rendering means.
Here’s how MFG polished up a few different games on the RTX 5080:
Nearly 144fps for fully path-traced 4K? Hey, if you’re going to have fake frames, you might as well have absolutely loads of 'em. More than the (much more expensive) RTX 4090 can pump out with DLSS 3, for the record, and more than you’d ever get from the equivalent frame generation behind FSR 3.
Still, what both prospective GPU buyers should remember – as should game developers thinking frame gen can make up for performance shortfalls – is that these kinds of AI-assisted performances are still reliant on having a sturdy base of plain, old, PC-rendered frames. Running S.T.A.L.K.E.R. 2 at 200fps+ is fun because it looks silky smooth, yes, but it only works because I’m getting a nicely playable 64fps underneath the frame gen, which is really all that matters for keeping my aiming feeling sharp. You probably could take something that only runs at 15fps, slap 4x MFG on it and get the counter running at a respectable 50fps or so, but it’s still going to feel like spluttering dogshit because the screen is still only reflecting user inputs 15 times a second. Which, in PC game terms, is spluttering dogshit.
Perversely, then, you do need to get games running well without frame gen help before they... become suitable for frame gen. But I do think the technology still has value as an add-on bonus, and with DLSS 4 and MFG, that bonus has never been fatter.
Game support looks decent as well, with 75 games confirmed to either add MFG compatibility immediately or in the future. Some of these won’t have baked-in, native support, but that’s where another nice DLSS 4 feature comes in: DLSS overrides. Applied via the Nvidia app, these overrides can essentially add forwards compatibility with newer DLSS features if they supported an older equivalent. That means adding DLAA or the new Transformer model to games that launched with more limited DLSS upscaling support, or, if they worked with DLSS 3, you can toggle them to play nice with 3x and 4x MFG as well. That’s now I got Dragon Age: The Veilguard and S.T.A.L.K.E.R. 2 into these tests: they weren’t updated to include MFG as such, but I could force it via the Nvidia app.
This is a great feature that commendably changes tack from Nvidia’s usual inclination towards solely the latest and greatest. It doesn’t need an RTX 50 series card (unless you want an MFG override), and some of these 75 games are years old, breathing new life into them by unlocking hitherto blocked-off tools for boosting performance.
DLSS 4 part IV: The latency issue
If MFG has a problem, it’s not so much that its frames aren’t real, but that it adds input lag. That’s delay between making a mouse movement or keyboard press, and your PC registering it as an input, leading to the appropriate action on your screen. Both DLSS 3 and now MFG can add more than a few precious milliseconds to this latency, which can result in a less responsive, more sluggish feel as you play.
The exact extent will vary by game, largely because – and I hope you appreciate the irony here as much as I do – latency is partially determined by framerate. As in, the real, regularly rendered framerates, with a higher FPS bringing down lag. Here’s how those four games tested above fared with both 2x and 4x frame gen:
| Cyberpunk 2077 | Alan Wake II | Dragon Age: The Veilguard | S.T.A.L.K.E.R. 2 | |
|---|---|---|---|---|
| DLSS Quality | 52-70ms | 64-72ms | 26-30ms | 51-67ms |
| DLSS Quality with 2x FG | 66-87ms | 78-95ms | 38-46ms | 55-76ms |
| DLSS Quality with 4x MFG | 69-89ms | 92-107ms | 36-50ms | 53-78ms |
Veilguard and S.T.A.L.K.E.R. 2 saw the lowest latency spikes, in part because they had a higher ‘base’ framerate without frame gen) to help keep it down. I actually didn’t feel much of a tangible difference with these two, so I personally wouldn’t have much beef with flicking MFG on.
Alan Wake II, on the other hand, clearly got a bit soupy as it rose through the frame gen rank, being the only one to break 100ms. Cyberpunk 2077 was okayish – I could sense a certain loss of snappiness in the aiming and car camera controls, though I didn’t actively struggle with it.
I’m saying "I" a lot because like framerates themselves, input lag is one of those tricky game things where people perceive them differently. There are surely some shooter fans who’d spit in disgust at anything over 20ms, whereas even the 80-100ms range won’t feel alien if you’ve played a lot of 30fps console games. Neither approach is wrong, but it is a fact that frame generation raises latency.
The good news, in a sense, is that 4x MFG only adds a relatively tiny dash more lag to what 2x MFG already does. And while you might notice a a 20ms-odd jump from turning on any kind of frame generation, the extra 3ms or so that 4x has over 2x is probably going to be harder to spot. As such, if you do end up with an RTX 50 series GPU and a taste for AI frames, my advice is that you might as well go all the way.
DLSS 4 part V: The verdict
Otherwise, there isn’t really much wrong with DLSS 4. Its AI applications are a world apart from the dodgy, joyless, art-stealing stuff, and even if some individual components aren’t to everyone’s tastes, it’s still a broad collection of tech tools that make your games look nicer. Or run faster. Or both. Consider especially that several of its best upgrades are available to the older and cheaper members of Nvidia’s graphics card catalogue, and it’s already far more of an impactful update than DLSS 3 was.
Should you, though, go as far as to buy an RTX 50 model that can take full advantage of it? The answer is slightly awkward, as while MFG isn’t enough of a faultless panacea that everyone should immediately join the months-long waiting list for RTX 5080 stock, I do think tech like this is becoming more and more of a difference-maker when it comes to choosing future upgrades. The RTX 4080 Super, for instance, costs about the same as the 5080, and performs very nearly as well in games without frame gen gubbins. If all else is equal, why would you not choose the one with more features?
That might not be a ringing endorsement for the state of the graphics industry as a whole, and I guess there is the risk that continued reliance on DLSS magic could further exacerbate issues with questionable games performance and sky-high system requirements. But that would be a misappropriation of DLSS 4, not a misstep by the technology itself. Which, regardless of which RTX model you have, remains worth using.