NVIDIA DLSS 5 Technical Preview: Inside 3D-Guided Neural Rendering

DLSS 5 is now official, and it is not another upscaler stepping. 3D-Guided Neural Rendering is a generative AI stage attached to the end of the render pipeline, and its job is to take the frame the game engine already drew and repaint how that frame responds to light. Skin gets subsurface scattering, hair and foliage get light transmission, contact shadows tighten, and reflections resolve, all without the renderer spending a single additional ray on any of it. Basically, the works of which generative AI has evolved to be very good at. The technology goes live on September 3, 2026 in NBA 2K27, it is exclusive to GeForce RTX 50 Series hardware, and it will punish your frame rate. In this technical breakdown, I’ll talk about how it works, what developers can control e.g. guide the model to lock probability, and what to expect from the performance hit DLSS5 gives.

The full technical rundown is on NVIDIA’s DLSS 5 announcement page, and the NBA 2K27 is available on Steam.

Every DLSS Before This One Was Solving One Problem

Super Resolution, Ray Reconstruction and Frame Generation are all reconstruction solutions. NVIDIA has always put DLSS a way to renderer or draw something specific; then model’s job is to work out what that something was, and there is a technically correct answer sitting at the other end that you can measure against. You then train the network then compare its output to a reference frame, and score it.

But neural rendering is a different beast altogether. DLSS 5 is literally doing a Stable Diffusion render with every frame in real time, adding material that the renderer never attempted to produce: the way light bleeds through an ear, the sheen on a fabric, the darkening where a shoe meets the floor. There is no ground truth image to check the output against, which makes it far harder to validate and opens a door to ambiguity that none of the earlier DLSS technologies had to worry about.

NVIDIA’s framing of the problem is a compute challenge. Offline renderers spend minutes or hours refining a single frame; games get a handful of milliseconds. That is why studios routinely scale back their intended look to hit a performance target, and that’s the gap NVIDIA is aiming for generative models to close.

The Three Problems NVIDIA Had To Solve

Generative models learned photorealism by watching the real world rather than by simulating the physics of light. That makes them good at looking real and terrible at behaving predictably, and NVIDIA had three specific obstacles to clear before one could be dropped into a real-time renderer.

• Artistic intent: Oh they go roasted hard for this during the DLSS 5 teaser reveal. Generative models alter images by design. DLSS 5 is trained to hold geometry, textures and the light-to-shadow relationships exactly as the artist built them, so character identity, composition and scene semantics stay locked. The model is also deterministic, which means the same input always produces the same output, unlike a prompt-driven model that will hand you ten different results from ten identical requests.
• Temporal consistency: video models generate frames in chunks to keep them coherent, which a game cannot do because the next frame depends on player input. DLSS 5 uses engine motion vectors and runs strictly one frame in, one frame out, which is what keeps the output free of shimmer, swimming and drift.
• Speed: a general-purpose generative foundation model takes seconds per frame. DLSS 5 is a narrow, specialized model that does exactly one job, running on the fifth-generation Tensor Cores in RTX 50 Series silicon.

One Frame In, One Frame Out

This is an important bit of the process, because it is what separates DLSS 5 from an AI filter. The model is not conditioned on a text prompt. It is conditioned on the rendered frame itself (READ: OUTPUT IS BASED ON RENDERED FRAME), the one that the game visual artists designed and the engine built, and image generation becomes a final stage instead of being in image-creation stage. The frame dictates what remains and on the image; the artist directs what may change.

NVIDIA is also explicit that this is not a post-process filter. The model retains what the large generative models knows about the real world, so it understands object types, human features, and light sources, and it knows which cues make each of them read as photoreal. In NVIDIA’s own example, it knows to add scatter and glow to hair and fur and not to a metal crate.

What DLSS5 Actually Changes On Screen

NVIDIA demonstrated the effect with two internal tech demos that we saw in the original announcement and the SIGGRAPH keynote. On the ramen chef scene, geometry and facial structure are untouched with DLSS 5 enabled. What changes is how light responds: ambient occlusion around the head and fabric is more pronounced, contact shadows are more realistic, the ears pick up natural subsurface scattering (that glow that passes thru your skin) as light passes through them, and the eyes gain clearer reflections.

The Zorah clay render is more instructional because this scene ships with no materials and no textures at all. Neural rendering alters none of the shapes on-scene. The proportions around the eyes and the shape of the mouth stay intact, while the lips and skin gain self-shadowing and reflections, and the hair gains strand definition, volume, and light response.

Garbage In, Garbage Out

DLSS 5 is tied to whatever the devs hands it from the rendering engine, and that goes in two directions: NVIDIA ran the same Zorah scene three times to make the point in their release trailer: rasterized without ray tracing, the neural rendering pass still produces a visible visual improvement but the result is limited because the input frame lacks accurate lighting information. With ray tracing enabled, the output is marginally more detailed. Move to path tracing and the model has more accurate information again, barring performance penalties.

The practical application is that DLSS 5 is not a shortcut around a weak renderer. More detailed materials, better assets and more accurate lighting all feed the foundation that the DLSS 5 model uses, and the better that foundation is, the more useful data the model has to work with.

The Control Knobs

Making an image more realistic is an artistic decision with no single correct answer, so NVIDIA built the controls into the SDK and handed them to the studio. Some controls come from the model, others from the engine, and together they let a developer art-direct where neural rendering lands and how hard it hits.

NVIDIA showed structure intensity at 25, 50, and 95 percent on the same model. Twenty-five percent stays close to the rendered frame; ninety-five percent moves toward the model’s maximum uplift. Run the same intensity value through different models and you get distinctly different results, which is the point of shipping more than one.

Masking: Where the Effect Is Allowed to Work On

A mask is just a way of telling DLSS 5 where to work. Switch the effect on for one thing, leave everything else exactly as the engine drew it. There are two ways to make one.

The first is automatic. DLSS 5 can tell what it is looking at, so it already knows which pixels are the character and which are the room behind them. It draws that mask itself with no work from the developer. Move the intensity slider and only the character changes while the background sits still. A studio can push the character hard and leave the environment alone, or do the exact opposite.

The second is manual. The developer tags objects in the engine, and DLSS 5 then treats each tagged group separately. NVIDIA showed a kitchen scene where the pitcher, the glass and the bowl of grapes each carried their own mask, with the cutting board and the props on the counter handled as a different group. There is no limit on how many groups a studio can create, and every group gets its own structure and tone values.

Together they let a studio work as broadly or as narrowly as it wants. Boost subsurface scattering on plants, foliage and food across an entire game and touch nothing else. Or tune characters, environments, gameplay and cutscenes as four separate jobs.

DLSS 5 Debut Title: NBA 2K27

NBA 2K27 is the launch title and, as of writing, the only announced DLSS 5 game. NVIDIA says more titles are coming “in the coming weeks and months” but has not named any of them alongside this launch. Integration runs through the existing NVIDIA Streamline framework or an Unreal Engine 5 plugin, and DLSS 5 is an independent, optional feature that sits alongside Super Resolution, Multi Frame Generation and Ray Reconstruction rather than replacing any of them.

Performance: Read the Footnotes First

NVIDIA published three NBA 2K27 charts, one per resolution, all on a Ryzen 7 9800X3D with 64GB of RAM on Windows 11 x64, at the Ultra preset with ray tracing on. The key number is 370 FPS at 4K with an RTX 5090, and that number is important for context.

Every bar on those charts is DLSS 5 stacked on top of Multi Frame Generation 6X and Super Resolution. The 4K chart runs Super Resolution in Performance mode, which is a 1920×1080 internal render scaled to 3840×2160. The 1440p and 1080p charts run Quality mode. Multi Frame Generation 6X means five of every six frames on screen are generated rather than rendered, so the underlying rendered frame rate is roughly 1/6 of the numbers shown. So the numbers line up around the estimates below:

NVIDIA has not published a DLSS 5 on versus off comparison at matched settings, so the cost of the neural rendering pass by itself is not something we can pull from these charts. What the charts do tell us is that the shipping configuration NVIDIA is proud enough to publish leans on aggressive upscaling and 6X frame generation to get there, and NBA 2K27 is a basketball game rather than an open world. We will not put a percentage on the overhead until we have run it ourselves.

Five Times Faster in Six Months

The one number in this launch that is straightforwardly impressive is the optimization curve. NVIDIA first previewed DLSS 5 running on two RTX 5090s in March 2026. By May it was running on a single RTX 5090, and by August NVIDIA’s chart puts model performance at 5X the March baseline across the RTX 50 Series, with the trend line still climbing into October. NVIDIA says more optimizations are coming.

Hardware and Game Support

What We Still Want To Know

NVIDIA closed its own presentation by saying it is not claiming photorealism is solved and that this is the very beginning of real-time AI generation, which is a more measured position than the launch materials around it. A few things are still open from where we sit.

• The isolated cost: without an on versus off comparison at matched upscaling and frame generation settings, the frame time price of the neural rendering pass alone is unpublished.
• Behaviour outside a basketball court: NBA 2K27 is a controlled, well-lit, indoor scene with a small cast on screen. A dense open world with volumetrics, weather and dozens of characters is a very different workload for both the renderer and the model.
• VRAM: NVIDIA has not published a memory footprint for the model, and the RTX 5060 sits at the bottom of the supported stack.
• The RTX 50 Series wall: NVIDIA’s position is that DLSS 5 targets fifth-generation Tensor Cores. Whether that stays a hard hardware line foisted on the current generation or will it eventually softens into a performance tier is the question a very large installed base of RTX 40 Series owners is going to keep asking.
• Studio discipline: the controls exist, but they only matter if studios actually spend time in them. A shipped default is a decision too.
• How good the feet renders are: for science

Our Take

DLSS 5 is a genuinely different animal from the rest of the DLSS stack, easily the first part of the tech that actually punishes performance rather than enhancing it like the rest of the family, and NVIDIA is effectively sidelining the RT in RTX for its main moneymaker right now. It is purely coincidental that NVIDIA has come to this stage but at the end of the day, visual photorealism has always been the goal of computer graphics and generative AI has been its newest friend in the good part of this decade. With NVIDIA leading the way, they’re taking charge of potentially the easiest crutch to enhance games visually while still relying on the RT stack to fuel the ground truth that fuels the DLSS 5 model.

Easily the most divisive addition to the RTX tech stack, its going to swim up the same river that ray tracing had to struggle with in the last decade and hopefully the current memory supply crisis doesn’t hinder the hardware evolution of this tech unless NVIDIA is willing to optimize and get it to an acceptable state with current hardware.

Related Videos