NVIDIA DLSS 5 Is Coming This Week, and It Could Change What PC Game Graphics Even Mean

NVIDIA’s DLSS 5 is finally moving from tech demos and internet discourse into an actual game. The new technology will be available from September 3, and, at least for now, NBA 2K27 is the only game officially confirmed to launch with it. It will work across NVIDIA’s GeForce RTX 50-series desktop and laptop GPUs, along with GeForce NOW Ultimate.

That might sound like a surprisingly small debut for something NVIDIA has spent months positioning as its next major leap in PC graphics. But DLSS 5 also isn’t a normal DLSS update. It isn’t primarily about upscaling a lower-resolution image, generating extra frames, or cleaning up ray tracing. This time, NVIDIA wants AI to help determine how the finished game actually looks. And that is precisely why DLSS 5 has been so interesting — and so controversial.

DLSS 5 goes well beyond upscaling

DLSS has become a fairly broad umbrella over the years. Super Resolution can take a lower-resolution image and reconstruct it at a higher resolution. Frame Generation creates additional frames between traditionally rendered ones. Ray Reconstruction uses AI to improve ray-traced effects. With DLSS 5, NVIDIA is adding something it calls 3D-Guided Neural Rendering, and this operates much closer to the end of the rendering pipeline.

In simple and layman terms, the game still renders its scene normally. The characters, geometry, textures, lighting, and everything the artists created remain the foundation. DLSS 5 then analyses that rendered frame and uses a specialised AI model to add details that would otherwise be extremely expensive to calculate in real time.

That could mean more believable shadows where a shirt meets somebody’s skin, light passing through hair or leaves, or subsurface scattering that makes skin react to lighting more naturally rather than looking like a flat plastic surface. NVIDIA says DLSS 5 analyses things such as objects, materials, lighting, color, surface information, and spatial relationships to produce those enhancements. It works on a one-frame-in, one-frame-out basis and uses motion vectors from the game engine so that whatever it adds should remain consistent while the scene moves.

The devs’ side

The other big piece of DLSS 5 is how much developers can supposedly decide what the AI is allowed to touch.

There are multiple neural models available through the DLSS 5 SDK, and developers can even use different models depending on the scene. One could be used for an outdoor environment full of foliage, for example, while another could be selected for an indoor cutscene.

Then there are two particularly important controls: Structure Intensity and Tone Intensity.

Structure affects details including ambient occlusion, reflections, contact shadows, and subsurface scattering. Tone handles broader lighting and color changes. Developers can lower these effects substantially or push them much harder depending on what they actually want the game to look like. And this is where things get particularly interesting.

DLSS 5 also supports semantic masking, allowing its neural rendering system to distinguish between different elements in a scene so developers can decide where the effect should and shouldn’t apply. A studio could enhance the environment while leaving a character relatively untouched, for example, or separately adjust elements such as foliage, glass, or water. Developers can also provide their own masks when they need even more precise control.

We recently got an unofficial look at those controls

Interestingly, we didn’t have to wait until September 3 to see some of this in action.

A DLSS Neural Rendering DLL was recently discovered inside the early-access PC build of NBA 2K27, and modders from the RenoDX community quickly started experimenting with it in other games. The leaked implementation was first demonstrated in games including Control, before spreading much further through the modding community.

More importantly, those experiments exposed just how adjustable the technology can be.

The modding tools include controls for overall neural-rendering intensity, local tone, local structure, skin structure, automatic skin masking, and different presets or styles. The game looks completely different depending on how the settings are being used.

That leaked build shouldn’t be treated as a proper benchmark for NVIDIA’s finished implementation, especially when DLSS 5 is being unofficially injected into games that were never designed around it. But it does make one thing much clearer: “DLSS 5 On” isn’t really one single look.

And that brings us to the elephant in the room.

DLSS 5 has already had a rough introduction

When NVIDIA originally showed DLSS 5 in March, the reaction wasn’t exactly subtle.

Early demonstrations showed the technology applied to games including Resident Evil Requiem and Hogwarts Legacy, and plenty of players weren’t impressed with what it did to character faces. Some thought the results looked unnaturally glossy or excessively beautified, while others felt the AI was interfering too much with the original art direction. The whole thing quickly became a meme.

“DLSS 5 Off / DLSS 5 On” posts started appearing everywhere, usually taking an ordinary image and replacing it with something hilariously overproduced or completely ridiculous. Even some game developers joined in on the joke.

Underneath the memes, though, there was a genuine concern: should an AI model be deciding that a developer’s character, lighting, or texture needs to look more “realistic” in the first place?

That criticism is worth taking seriously because realism isn’t automatically better graphics. Fortnite doesn’t need to look like Alan Wake 2, and neither should a stylised JRPG suddenly have photorealistic faces simply because the GPU is capable of producing them.

NVIDIA’s answer is essentially that developers remain in control, and looking at the sheer number of controls available, that argument makes more sense now than it did when DLSS 5 was first demonstrated. But ultimately, we'll need to see how studios actually use it rather than how good the SDK looks on paper.

NBA 2K27 might actually be the perfect first test

That also makes NBA 2K27 a fairly logical place to start.

A basketball game is already trying to recreate real players, real arenas, realistic skin, clothing, hair, and broadcast-style lighting. There isn’t much of a stylised visual identity for DLSS 5 to accidentally steamroll.

Visual Concepts is using the technology to improve effects such as skin scattering, contact shadows around jerseys, hair lighting, and the way light passes through them. The performance side is slightly less straightforward.

NVIDIA says the RTX 5090 can reach up to 370fps at 4K with Ultra settings and ray tracing, while its 1440p figures reach as high as 590fps. But those results use the broader DLSS package, combining DLSS 5 with trickeries such as Super Resolution and Multi Frame Generation.

Neural Rendering itself is actually computationally expensive. NVIDIA says it has improved DLSS 5 performance by around five times since the technology was shown in March, when it required two RTX 5090 GPUs. It can now run on a single RTX 50-series GPU, with NVIDIA promising further optimisation later this year. That may ultimately be the most fascinating part of DLSS 5.

For years, NVIDIA has used AI to help GPUs render less work while producing something close to the same result. DLSS 5 flips that idea around. The neural rendering is now doing additional work because NVIDIA believes it can create lighting and material detail that traditional real-time rendering still struggles to deliver.

Whether that ends up being the next major step for PC graphics or simply another expensive graphics option people turn off will depend heavily on developers.

For now, NBA 2K27 is our first proper chance to find out.