DLSS 5 Tested: NBA 2K27 Image Quality + Benchmarks, Mods And More

Rich (0:02)

Well, it's finally here. NVIDIA's DLSS5 has actually shipped within an actual game, implemented by actual artists within Nba 2K27. It's the culmination of a run up to launch that began with the contentious March reveal at GTC this year, where we got to spend about four hours of the technology. Things didn't go too well after that. It was followed by a pause where essentially NVIDIA rethought its marketing process in the wake of this massive backlash.

The reintroduction of DLSS 5 began with a SIGGRAPH presentation exhibiting how the technology works, the three new rendering models on offer, and how developers can tune the effects to best suit their artistic vision. And this was followed by closed doors presentations at Gamescom 2026 with a wide array of press and a follow up virtual presentation for those like me who couldn't make it. So basically the key aim throughout since this reintroduction has been to emphasize developer input and even validation of the technology.

However, embargoes in place were blown out of the water when the technology leaked. Back in March, I predicted an opti-scaler injector, but I can't say I expected it to arrive before the first game did. But here we are. So in this video, we're going to talk about DLSS5 in Nba 2K27. Going to experiment with a modded experience in an attempt to learn more about how it works and what it does. And we'll be talking about performance. Joining me first of all here, our man at Gamescom 2026 Fair to say a staunch critic of the original presentation. John Lineman, hello.

John Lineman (1:38)

Yes, I think that's putting it mildly and in general not a fan of generative AI at all.

So everything I say today here is kind of should be viewed through that lens, that I'm not somebody that champions this stuff. But I was surprised by some of the results. And then the modding started to happen. I'm just like, oh boy. But yeah, we'll get into all of that today.

Rich (2:02)

And next up, well, the first time I actually met him in person was at that ill-fated GTC event back in March. But here to talk more about DLSS 5 is Oliver McKenzie.

Oliver McKenzie (2:12)

Yes, I had fun at that event regardless, although some of the response to that was not quite as I anticipated. But you know, I think that we're seeing a very interesting and compelling technology here today and I'm happy to talk about it.

Rich (2:27)

Okay. Where do we want to begin then? Let's talk about what is DLSS 5? What is its intention?

Oliver, do you want to start on that?

Oliver McKenzie (2:35)

Yeah. So DLSS 5 is basically a diffusion model, AI diffusion model that aims to improve lighting fidelity in games.

Technically, it's a family of three models, three models that are available to the users in the mods, for instance. But only one is used at any given time. It tries to basically add complex lighting phenomena to a rendered frame, tries to impart those characteristics onto a rendered frame, while staying anchored to the geometry and the materials in that frame. So it essentially takes the input frame and it's conditioned through a training process to make it look as much like a photograph as realistically possible. They seem to have trained in photographic data for this model, which is something that a lot of people, including Alex, speculated when the model was initially debuted, while only hopefully significantly changing the lighting qualities and some of the material properties of the image. That's the intention. It's intended for games that are aiming for realism. It's not intended for games that are very stylized because photographic like image transformations may not be appropriate for those games, although with the mods, you're free to do as you like, maybe with some good results, maybe with some not so good results. They do importantly flag that they're continuing to work in this space and they intend to introduce additional artistic styles and controls for future releases that they might have, of products like DLSS 5 in the future potentially. From the game, it only requires color and motion vectors, and then any masks the developer has provided as well, may be used in DLSS 5 process. Integration relative to other things in the sphere should be fairly easy. That's not to say that it's a painless process, it can be a very involved process, but relative to a lot of other lighting changes you might see in games, this could be a relatively easy thing to integrate.

That's DLSS 5 Now, most of this was known back in March to some degree, either kind of inferred or directly stated. The key thing is though we know more about how it was trained, and that it does intend to produce photographic-like results, because presumably it was trained on photographic input, probably a lot of studio photography is my guess. That might suit some games well, realistic games quite well, stylized games less well, maybe you can get a good result out of them by tweaking things. Could it depend on the title potentially there. But that's DLSS V in broad strokes. It's basically a diffusion model that intends to produce realistic lighting phenomena in games. That's the intention, and I guess we'll see how well it does.