NVIDIA DLSS 5 Launches Thursday With NBA 2K27, Here’s What To Expect
NVIDIA showed DLSS 5 again at SIGGRAPH 2026 in June, and while the firm shared more details about its new graphics technology, the demos were fundamentally similar to what it had showed before. This made me wonder if the state of DLSS 5 upon release would be much different. Well, the answer is mostly yes -- the tech that is launching this week is in-line with what was demonstrated back in March, but with the benefit of six months of intense optimization work.
So what exactly is DLSS 5? We know much more about it now, so let's explain...
NVIDIA's DLSS 5 Neural Rendering In A Nutshell
NVIDIA says DLSS 5 is a breakthrough in real-time rendering and photorealism. The company refers to it as "3D guided neural rendering" and makes many bold claims about how it is the future of graphics technology. We're not ready to evaluate those claims today, and we're also not going to be looking at our own third-party testing just yet; that will have to wait until Thursday at the earliest. Instead, we're going to look at NVIDIA's updates and talk about the technology.
Fundamentally, DLSS 5 is a diffusion transformer model that is intended to "make real-time rendering look more real." It's not a simple filter, but rather a greatly specialized version of a video generation model specifically optimized for improving the detail of 3D video game graphics. As NVIDIA's Edward Li explained at SIGGRAPH, DLSS 5 inherited broad world knowledge from the much larger model it was based on that allows it to understand more about a scene; in the process of specialization, it has lost generality (so it can't generate new media from text prompts, for example), but it gained the ability to run in real time.
Put even more simply, it's an image-to-image AI model, but it doesn't just take the finished output buffer and run that through Stable Diffusion or something similar. For one thing, the model is deeply specialized. But it's also much smarter than something like Stable Diffusion; while enterprising hackers have proven that it can work with relatively little input data, it can take lots of different inputs besides the finished color buffer, including what seems like the entire G-Buffer. That's the collection of 2D texture layers that stores albedo (base colors), surface normals, depth information, roughness and metalness, and motion vectors, among other values.
That's not all it can use, either. DLSS 5 supports engine-side inputs as well as a few different controls on both the engine side and the model side. The engine-side inputs include developer masking, where the developer can define objects in the scene as masks for DLSS 5 to treat differently from the surrounding environs, as well as two scene controls: structure intensity and tone intensity. Developers can fine tune the way DLSS 5 treats a scene by carefully defining masks and setting the intensity values; structure intensity affects high-frequency detail for things like better ambient occlusion and better reflections, while tone intensity affects low-frequency detail, impacting lighting and contrast.
Then, from the model side, developers can select from three different models for DLSS 5, each of which has a slightly different look and it optimized for different types of scenes. It's not clear how this relates to the seven presets that modders found inside the leaked DLL from NBA 2K27. Developers can also use a feature called "model automask" which apparently automatically detects characters in a scene and allows developers to adjust the intensity of the DLSS 5 effect on specific characters. It seems to work well, but it would be nice to see NVIDIA demonstrate this in a scene with more than one character.
The Three Biggest Challenges Of DLSS 5
All of this fine-tuning and configuration capability was added to DLSS 5 Neural Rendering to deal with one of the three biggest concerns facing the technology, which was preserving artistic intent. You see, even though DLSS 5 is deterministic, according to NVIDIA—meaning it always produces the same result—if the model wasn't trained and set up properly, it could change the scene too much, eliminating critical details or adding them where they shouldn't be. NVIDIA says that it was a goal that DLSS 5 must change the image in a way that "it does not change the story", so characters remain recognizable and true to the original creators' intentions.Whether NVIDIA succeeded in that goal is something we'll be looking at in the future, but that was just one of the three challenges. The second one had to do with temporal stability. NVIDIA says that conventional AI video models produce frames in bursts, working from the previous frame to ensure continuity. That, obviously, isn't going to work when running in real time; video games require response times measured in dozens of milliseconds to feel responsive. The solution was apparently to train the model specifically to work on exactly one frame at a time, just like a regular image diffuser.
Finally, the third challenge was performance. If you recall, NVIDIA's demonstration back at GTC was running on a pair of GeForce RTX 5090 cards, with one card to render the game and the other to run the DLSS 5 model. The tech was roundly criticised for that requirement despite that NVIDIA insisted that the steep hardware requirements at that time were an artifact of its experimental nature. As it turns out, that appears to have been true, as NVIDIA has not only gotten it running on a single GPU alongside the game, but also running on some relatively modest graphics cards like the GeForce RTX 5060.
Strictly speaking, I don't know what the minimum requirements to use DLSS 5 are just yet, or where NVIDIA will ultimately settle once the technology is further optimized, but NVIDIA gave us three sets of its own benchmarks in NBA 2K27. This is a fairly lightweight game in general; the recommended (not required) hardware for the game includes a Core i5-10600 and a seven-year-old GeForce RTX 2060. The enormous numbers in NVIDIA's benchmark chart were produced using 6× multi-frame generation, though. I've helpfully gone in and added the actual non-MFG render frame rate to each of the charts so that you can see what the performance is like.
NVIDIA is testing the game on what appears to be its highest settings with ray-tracing enabled; we don't yet know what ray-tracing effects the game uses, but in any case, the absolute framerates produced are pretty low without frame generation. Is this to do with DLSS 5 being heavy on performance? Partially, yes. The leaked version that modders have been shoving sideways into every game has a high performance cost, and some commenters have speculated that this could be an artifact of its nature as an unsupported, unoptimized hack. We'll know for sure when we can do our own testing on the matter.
The Real Question: How Does DLSS 5 Actually Look?
Truth be told, I haven't gotten the chance to experience DLSS 5 first hand just yet, beyond watching NVIDIA's materials; I'm planning to do some hands-on testing later this week. However, HotHardware head honcho Dave Altavilla got to see DLSS 5 in person at GTC and he was impressed. NVIDIA's demos certainly do look impressive. First, check out the side-by-side foliage demo at the top if you didn't click on that already--that's from NVIDIA's Zorah demo, which was created specifically to show off this tech. So naturally, it's impressive, but what about in a real game? Well, it looks pretty damn good in NBA 2K27, too, at least from screenshots.
In this comparison featuring the Detroit Pistons' Cade Cunningham, it's immediately obvious what DLSS 5 is doing for the scene. Nearly every single part of the image is improved, from the more accurate texturing and shading on the basketball, to the lines and skin on Cunningham's face, to the material of his sleeve and jersey. His hair could use some work, but curly like his is an enormous rendering challenge that only Rockstar seems to have tackled, judging by the footage released so far of Grand Theft Auto VI.
On the other hand, in this comparison showing a courtside videographer and some fans in the stands, the improvements are more subtle, yet no less noticeable when you start looking. Besides the young cameraman's face, the camera itself gains tremendously in material and shading detail, and subtle details like the shading on the camera operator's hands and the textures of the fans' faces are also present in the DLSS 5 version. There are other comparisons we could show you, but you probably get the point by now.
Earlier we noted that NVIDIA is testing NBA 2K27 with all the graphical features turned on, and you might wonder why that is considering that DLSS 5 is essentially going to modify all of the graphics from the game with its own interpretation. The reason is simple: having a better base render gives DLSS 5 better grounding to start with. There's no reference ground truth to match against like with DLSS upscaling or ray reconstruction, so it's important for the game engine to put its best foot forward to ensure the best and most lifelike results from DLSS 5.
Of course, not every video game is aiming for a photo-realistic look. DLSS 5, at least this version of it, is explicitly targeting photorealism, so it is simply not well suited for games that aren't attempting a high degree of realism. Still, the tech is planned for many more games than just NBA 2K27. Back at GTC, NVIDIA said that DLSS 5 is slated for Aion 2, Assassin’s Creed Shadows, Black State, Cinder City, Delta Force, Hogwarts Legacy, Justice, Naraka: Bladepoint, NTE: Neverness to Everness, Phantom Blade Zero, Resident Evil Requiem, Sea of Remnants, Starfield, The Elder Scrolls IV: Oblivion Remastered, Where Winds Meet, "and more." Most of these games do indeed have a visual style that trends toward photorealism, although I'm dying to see what it looks like in anime gacha NTE.
We'll have slightly more comprehensive reporting on DLSS 5 when NBA 2K27 launches with the feature on September 3rd, which is this upcoming Thursday. For now, what we can tell you is that the tech is real, and it's coming to a Blackwell GPU near you. No word on official support for Ada Lovelace chips at this time, but they do have the requisite FP8 support, so it's possible we might see neural rendering hit last-generation GPUs too. Unfortunately, folks with older chips are probably out of luck—the leaked version hammers RTX 30 series GPUs—but NVIDIA hasn't explicitly said such, so stay tuned for more details as we get them.
