A few months back, Nvidia started making a bold claim. It said that DLSS 4 and its new transformer AI model provided better image quality than native resolution. The claim came up during the launch of the ill-fated RTX 5060, but it's a sentiment that's been floating in the air of the PC gaming ecosystem for years now. Sometimes, it's worth turning on DLSS because it looks better than running a game at native resolution, even if you don't need a performance uplift, or so the story goes.

It sounds like nonsense dreamt up by a marketing team that's playing fast and loose with the definition of "true," but there's some legitimacy to the claim. Although the idea that upscaling can produce a better image than native resolution seems counter-intuitive, it actually makes a lot of sense for the games we have today. And with the new transformer model in DLSS 4, you'll find games where running at a lower resolution produces a better image than running at a higher one.

Let's define "native resolution"

It's not just about rendering all the pixels

I've intentionally left the image above unlabeled. This is Marvel Rivals, and in one of the images, I'm running at native resolution, and in the other, I'm using DLSS set to Performance mode, which means only a quarter of the displayed pixels are actually rendered. You make up your mind about which image looks better, and we'll revisit the labels at the end of this section.

A lot of the confusion around the idea that DLSS looks better than native resolution comes down to a misunderstanding of what "native resolution" even is. At a basic level, native resolution means your GPU is rendering every pixel on your display. You'd assume that it would be impossible for DLSS (or any other upscaling tool) to beat native resolution because these tools don't render every pixel. Your GPU only renders a subset of the pixels you'll be shown with DLSS on, and the rest are inferred by an AI model.

Just because your GPU renders every pixel doesn't mean you see the raw pixel output on your display. You almost never do with modern games, in fact. Any time a diagonal line crosses over square pixels, you'll be left with jagged edges, and as those edges move, they'll sputter and flicker. Native resolution shows the rough edges of rendering, and a straight-to-the-heart output doesn't look good. Most games don't handle "native resolution" this way. They're using some form of anti-aliasing to clean up those rough edges, even if you don't have anti-aliasing options in the graphics menu.

Nowadays, especially, games mainly use cheap forms of anti-aliasing, that is, anti-aliasing that has a negligible impact on performance. Largely, games use TAA, but you'll sometimes find FXAA, as well. More demanding forms of anti-aliasing like MSAA and SSAA don't show up often in modern games. When looking at games that use DLSS 4, we're not comparing "native resolution" to DLSS. DLSS and just about any other form of anti-aliasing will win that battle. We're comparing DLSS to rendering every pixel and using cheap anti-aliasing to clean up the image.

And to prove that point, let's return to Marvel Rivals. The image on the left is "native resolution" without any anti-aliasing, while the image on the right is DLSS set to Performance mode with only a quarter of the pixels rendered. You wouldn't play Marvel Rivals at true native resolution; you'd play it with TAA, at the very least. And, indeed, DLSS 4 with its transformer model looks better than most TAA.

DLSS is exposing, not extracting, detail

It didn't invent new details

The example Nvidia usually points to when talking about DLSS 4 and how it looks better than native resolution is Avowed, and in particular, an image like the one you see above. Look at the intricate details in the spellbook. With DLSS 4, it looks far sharper than it does at native resolution. DLSS 4 didn't just invent details to go into the texture sprawled out over the page, or at the very least, it's not applying some intense sharpening that applies across the entire image. DLSS 4, in this case, exposes the detailed texture that native rendering was smoothing over.

This makes a lot of sense if you know about Avowed, too. The game is built using Unreal Engine 5, which is an engine that is built around TAA. You can technically use MSAA in UE5, but you need to give up Lumen and Nantie with the forward renderer. Regardless, UE5 uses TAA, and turning it off produces a pretty terrible-looking image — just look at Marvel Rivals in the above section, which is also built on UE5. DLSS with its transformer model is able to expose the detail that TAA glosses over.

The transformer model is able to do this because of the self-attention mechanism within this type of model. Without getting too deep in the deep learning weeds, a transformer model is much more computationally expensive than CNN, but it also considers a broader range of context. With CNN, you could get an image that looks like native resolution — that is, every pixel rendered and washed in TAA — but with a transformer model, you're able to push beyond native resolution. The transformer is able to take in all the pixels at once, understand them, and rebuild the image with the context of a full scene.

Look at the final image

Listen to your eyes

When I recently wrote about Nvidia's DLDSR, which helps your games look even better at native resolution, I received some feedback that the "DLSS looks better than native resolution" claim is a load of hogwash. And, to a degree, that's correct. Rendering every pixel will leave you with more detail than inferring a portion of the pixels, which is a big reason why tools like DLSS work so much better at higher resolutions where there's more input data. As we've established, though, native resolution doesn't mean just rendering all the pixels. When we're talking about native resolution, we're talking about the final image you see, which involves clean-up after the pixels have been rendered.

And, as you can see from the images I've shared here, DLSS 4 with the transformer model looks better than native resolution, at least in more modern games that lean on TAA heavily. That's not the case in every game. A game like Forza Horizon 5 includes MSAA, for example, and the battle between it and DLSS 4 is much closer. You really don't need to dig into the engine or understand what graphics features are going on behind the scenes that you don't have access to. You just need to look at the final image.

I know it sounds obvious, but so much of the conversation surrounding DLSS (and other upscaling/frame generation tools) hinges on one image being objectively better than another. And usually, that falls in the camp of native resolution, weighted by the fact that the GPU had to do all the hard work of actually rendering all the pixels. The conversation doesn't hinge on a full image of a game you're playing. It focuses in on the minutia to extract differences, and frames native resolution as some sort of objective truth from which all other methods of rendering must be measured.

In reality, there have always been various ways that developers smooth over the rough edges of rendering, and DLSS Super Resolution is the latest tool that's caught on. Years ago, you'd have to go about it the hard way with SSAA, but then MSAA offered much better performance despite a drop in quality. FXAA caught on as an extremely cheap, but terrible-looking, alternative, and then the industry turned its eyes to TAA, which offered exceptional performance and great image quality, albeit with some visual artifacts.

DLSS has cemented itself as one of the key graphics features available on PC, and that's not going to change. If you don't like the way it looks, or you think that "native" rendering is better, that's great. For everyone else, look at the final image you get. Especially in recent games that hang their hat on TAA, DLSS' transformer model will usually come out ahead.