Nvidia’s DLSS 5, explained

Nvidia knows that DLSS 5 left a bad first impression. In March, many gamers took one look at Resident Evil Requiem protagonist Grace Ashcroft’s DLSS 5-ified face and declared it AI slop. So this week, the company’s re-introducing its controversial “neural rendering” with firm messaging about how it “preserves artistic intent,” “respects the rendered frame,” and gives game developers full control.

Now, ahead of its September 3rd launch, with an official Nvidia briefing under our belt and some hands-on time with the version that leaked onto the web, we think we finally get the picture. Here’s what’s going on.

The critics weren’t wrong: DLSS 5 largely is a generative AI filter for video games. Kind of like a camera filter on your phone, DLSS 5 analyzes and modifies a real-time stream of 2D images after your 3D game engine is done spitting them out. It can indeed make game characters look terrible if applied poorly, like they’ve aged 20 years in a split-second. It uses a lot of GPU power, and can easily cut framerate in half — even applied well, you could argue that existing games look better without it.

But DLSS 5 isn’t what we typically think of when we think of generative AI. It doesn’t change shapes, or hallucinate words on a sign, or place brand-new fake objects in view. Instead, it imagines light, shadow, color, and detail that make the same shapes appear differently.

DLSS 5 can make dull, pasty characters look far more photorealistic. It can fill their hair and clothing and nearby foliage with signs of realistic light. It can ground characters and objects and clothing so they don’t look like they’re floating above video game surfaces.

Here’s an example I captured myself in Yakuza Kiwami: with DLSS on, see how contact shadows underneath each character make them seem like they’re actually standing on the ground?

(Later in this story, you should really check out how Cade Cunningham’s jersey casts shadows on his video game body, so it looks more like he’s wearing it.)

And unlike AI-generated photos and videos, that usually produce different results with the same prompt, Nvidia claims DLSS 5 is deterministic, so games should look the same every time.

DLSS 5 is also not one-size-fits-all. Despite what it sounds like, “DLSS 5 On” and “DLSS 5 Off” isn’t just an on-off switch. Developers (and apparently modders) can drag various sliders to change how DLSS 5 plays with color and light, and they have three different AI models to work with at launch, giving your graphics “default,” “natural” or “cinematic” looks.

Plus, developers can decide whether to apply DLSS 5 to everything in a scene, just a character’s face and skin, just foliage, just silverware or furniture or even just food.

You’re saying game studios could leave character faces alone?

Yep. The toolset can automatically cut out a mask, like in Photoshop, to isolate a character’s face and skin from everything else:

If they don’t like the automated mask, devs can cut one out themselves. That’s the little “Developer Masking” checkbox in the image above. “There’s no limit to how many masks, how many groups, they could implement,” says Nvidia’s director of real-time content technology Gabriele Leone.

Would game studios actually leave the faces alone?

Probably not! The first game brave enough to ship with the tech is NBA 2K27, and I suspect that’s because Cade Cunningham actually looks a bit better with detailed skin, a uniform that doesn’t seem glued to his body, and a bit more glint in his eye. Take a look at Nvidia’s images for yourself:

That basketball, too — doesn’t it feel a bit more lifelike now most of its detail is in shadow?

Truly deranged. But these games are supposed to look like animation, and surely their creators would never do anything like this to them, right? Right??

“The soul and goal of GeForce is to achieve photorealism in real-time rendering,” Nvidia’s director of applied deep learning Edward Liu told journalists. And games with more realistic art styles don’t look anywhere near this bad. But even realistic games can look wrong if the sliders are set too aggressively.

What do you mean by “too aggressively?”

I’m willing to bet that if you’ve seen the leaked DLSS 5 modifying an existing game, the unofficial DLSS 5 mod was set to its default level of around 80 percent intensity or even higher. Some creators will even pump it up to 200 percent to make the difference between “DLSS on” and “DLSS off” especially clear.

But I believe even 80 percent is far too aggressive for many existing games. It can feel like you’re artificially aging a character by weathering their skin and deepening the wrinkles and creases in their face.

To show you, I fired up Shadow of the Tomb Raider. Here’s what Lara Croft looks like with DLSS 5 at 0 percent, 80 percent, and 200 percent intensity:

She seems… weathered, no? But at 40 percent intensity, she still looks like the same video game character to me, only with a bit more life in her eyes:

Similarly, here’s Kazuma Kiryu from Yakuza Kiwami once more. In this part of the game, he’s supposed to be 27 years old, and nowhere near the grizzled veteran he is by Yakuza 6. But as I ramp up the intensity from 0 to 40 to 85 percent, you can see his furrowed brow gets more furrowed, his hair lightens, and the increased shadow across his cheeks makes his face appear slightly gaunt.

He looks older — even if the underlying geometry of his face hasn’t actually changed. In a cutscene at 85 percent, it even looks like his hair might be tinged with grey:

But again, it’s something developers and modders can control with some granularity. I’d probably enjoy a version of DLSS 5 for older games that ignores the faces but enhances a character’s gear. Kiryu’s sworn brother and antagonist Majima, below, doesn’t quite look right to me — but his eyepatch and necklace look excellent.

By the way, these are just me playing with the overarching DLSS 5 slider when set to its “default” model. There are actually three different AI models, each with a slightly different look, and two different sliders, one for “tone” and one for “structure.”

What’s the difference?

I’m telling myself that “tone” is about light and color, while “structure” is about shadows and definition. But it’s probably easier to show you by isolating each:

Officially, Nvidia says that “structure intensity” controls a mix of ambient occlusion, contact shadows, reflections and subsurface scattering (read: how light scatters through materials like hair, skin, and leaves), while “tone intensity” controls color and “broader lighting” details.

What about the three different AI models?

Nvidia only called them Model A, Model B and Model C in my briefing, but the DLSS 5 mod I used describes them as “default,” “cinematic” and “natural.” Here’s the image Nvidia provided to show the difference:

And here’s my own comparison in Shadow of the Tomb Raider — you may need to blow these up and cycle through to spot the differences.

Is this really an AI filter? It doesn’t use the game engine at all?

Nvidia says that DLSS 5 can take some cues from game engines. But it’s based on the colors and motion vectors that it sees in the flat 2D frames, after they’re rendered and after other techniques like DLSS Super Resolution are applied.

When we asked how DLSS 5 knows the difference between a character model, an object, and a piece of foliage, Liu replied that the model was trained to recognize what it sees. “Just like a person, if you look at anything, you know where things are.”

You can even technically run DLSS 5 on video capture from a Nintendo Switch 2, and modders have already found ways to inject it into earlier games that used OpenGL and DirectX 9.

Will game developers actually buy into DLSS 5? Will they use it well?

That’s one of the big remaining questions — and one that Nvidia left unanswered during and after our briefing.

While Nvidia said in March that DLSS 5 would come to at least 16 different games including Assassin’s Creed Shadows, Hogwarts Legacy, Starfield, and Resident Evil Requiem, the company did not answer questions about when DLSS 5 might come to any game beyond NBA 2K27, nor did it confirm that other titles are still coming following the original backlash. Perhaps Nvidia and its partners will wait and see how reviewers and gamers react first.

It’s worth noting that this may be a niche technology to start: while it no longer requires two RTX 5090 GPUs to function, it still uses a lot of horsepower. Officially, DLSS 5 is only coming to the latest RTX 50 series GPUs, which have become quite expensive: even a lower-mid-range RTX 5060 retailers for around $470 these days.

Nvidia claims that an RTX 5060 is enough to play NBA 2K27 at 1080p with ultra settings, ray tracing, and DLSS 5 enabled — assuming you’re pairing it with a fast gaming CPU and a whole heap of RAM.

Just how much of a performance hit are we talking?

With NBA 2K27, Nvidia claims the performance hit is 50 to 60 percent on average across all Nvidia RTX 50-series GPUs. When I ran the Shadow of the Tomb Raider benchmark at 1440p with on an RTX 5090 at Highest spec, DLSS Quality and an unofficial DLSS 5 mod, I saw framerate exactly cut in half — 312fps with DLSS 5 off, 156fps with it on.

Will gamers be able to tweak DLSS 5 the way they like?

The DLSS 5 mods are so easy to apply, it seems like the cat is out of the bag. But Nvidia’s message is that the developer has full control, and it didn’t answer questions about whether it’ll provide controls to gamers, too.

Which mod should I install?

I don’t yet feel comfortable recommending one — there’s been a feeding frenzy for DLSS 5 mods ever since the leak and we haven’t been able to properly vet. I installed one on a completely fresh system, just in case something went wrong.

How have gamers reacted to Nvidia’s re-introduction of DLSS 5 so far?

The reactions have been all over the place since DLSS 5 leaked and modders began posting their work. I’ve seen people say that DLSS 5 looks like a free next-gen remaster of any game you like, and I’ve seen people say it looks as terrible as it did before. I’ve seen people suggest that even if it looks good, it’ll make every game look too similar once developers start using it as a crutch.

But many of those impressions were based on images and videos of the tech at its extremes. Once people realize there’s granular control, perhaps reactions will become more measured.

Correction: An earlier version of this article misidentified characters in screenshots; they are NBA player Cade Cunningham and Majima from Yakuza.