DLSS 5: What It Actually Does, What It Costs, and Why Artists Are Angry

DLSS 5 arrived on 3 September 2026 in NBA 2K27, six months after a GTC reveal that got it branded “AI slop” before anyone outside NVIDIA had run it. It is the most interesting thing to happen to real-time graphics in years and also the most misunderstood, partly because the name it ships under is misleading about the one detail that matters most.

It Is Not an Upscaler

Every DLSS before this one took a cheap frame and made it bigger or more numerous. DLSS 1 through 3 upscaled; frame generation interpolated. All of it was in service of frame rate.

DLSS 5 does neither. It takes a frame your GPU has already finished rendering at full resolution and relights it — changing lighting response, material appearance and fine surface detail using a generative model NVIDIA calls a pixel-space diffusion transformer. Geometry, camera, animation and material assignment all still come from the conventional engine pipeline. The model works on top of that.

The effects it adds are the ones that are expensive to compute honestly: occlusion in an eye socket, a contact shadow where a shoe meets the floor, the shadow a collar casts onto a neck. Without it, NBA 2K27 at ultra with ray tracing still reads slightly plastic. That is a real problem and this is a real attempt at it.

“3D-Guided” Describes the Training, Not the Runtime

This is the part most coverage skips, and it is the key to the whole argument.

The name suggests the model is fed the engine’s 3D understanding of the scene — the G-buffer, with albedo, surface normals, depth, roughness and metalness. Those channels are involved, but NVIDIA has confirmed they are used during training only. At runtime the model receives the rendered output frame and motion vectors, and nothing else.

So a shipped DLSS 5 frame is produced by a model looking at a 2D picture, inferring from appearance alone what each surface is made of, and then relighting it according to what surfaces like that looked like in its training data. It was taught using ground truth about geometry and materials. It does not get that ground truth when you play.

That single fact explains the artistic objection more precisely than most of the argument about it has. A model inferring material from appearance will regress toward the average of what it was trained on, because that is the only thing it can do when the image is ambiguous. If it was trained substantially on professional photography, the average it regresses toward is professionally photographed skin.

Conventional Neural

What the Developer Actually Controls

NVIDIA’s answer to the art-direction complaint is that studios hold the dial, and the controls are more granular than the discourse suggested:

• Masking. A developer can mark objects in the scene to be treated differently from their surroundings — push a crowd and leave the hero alone, or the reverse.
• Structure intensity. How strongly the model reworks fine surface detail and micro-geometry cues.
• Tone intensity. How strongly it reworks lighting and colour response.

Usefully, none of this costs anything at runtime. The model receives a screen-space buffer of control values and processes it identically whether there are five masks or a hundred, at any slider position. Tuning is free; the inference is what costs.

The remaining worry is not really technical. It is that a control which is free to use is also free to overuse, and that a studio under deadline may reach for it as a substitute for lighting work rather than a finish on top of it.

The Price: Roughly Half Your Frame Rate

This is where the marketing and the measurements part company. In NBA 2K27 at 4K ultra with ray tracing:

A 5090 losing well over half its throughput on a basketball game is a striking number. It is also the honest one.

NVIDIA’s own figure for the same GPU and game is up to 370 fps at 4K. That is not a rendered frame rate. It is display output after Multi Frame Generation inserts several generated frames for every frame the GPU actually draws. Both numbers are true and they describe different things; quoting the second without the qualifier is how a 68 becomes a 370.

DLSS 5 is also RTX 50-series only. Blackwell tensor cores, no RTX 40 support at launch.

The Artistic Case Against It

The objection is not that it looks bad. It is that it looks like something in particular, and that something is not necessarily what the art team drew.

Digital Foundry’s Alex Battaglia described the model as “over-averaging and over-perfecting things based upon presumably the training data”, and called that “ethically very problematic”. Steve Karolewics, a rendering engineer at Respawn, put it more bluntly: it looks “like an overbearing contrast, sharpness, and airbrush filter”. Danny O’Dwyer made the art-direction argument. The flashpoints were character faces — Resident Evil Requiem’s Leon Kennedy and Starfield in particular.

Take the runtime constraint seriously and the criticism becomes structural rather than aesthetic. A stylised game is a deliberate departure from photographic reality. A model that only sees the final image cannot distinguish a deliberate departure from a deficiency — it has no channel that says “this skin is matte on purpose”. Pointed at stylisation, its instinct is to correct it.

That is why masking matters more than the slider. Masking is the only mechanism that lets an artist say not this object, which is the thing the model cannot infer.

What the Backlash Got Wrong

The March demo that started it was NVIDIA running the effect at maximum, and it is worth being fair about how much of the reaction was to a setting rather than a technology. Hands-on builds at Gamescom, and the shipping version, are markedly more restrained. Notebookcheck’s testing concluded the “AI slop” framing had been overblown.

Some of the reaction was worse than wrong. Digital Foundry’s Richard Leadbetter and Oliver Mackenzie received death threats over their coverage. Leadbetter called it “completely unacceptable”, which it plainly is, and it is worth naming here rather than folding into a summary of the discourse.

NVIDIA’s own posture did not help. Asked about the criticism, Jensen Huang said the critics were “completely wrong” — a response that treats an argument about authorship as if it were an argument about image quality.

What to Take From It

• If you are buying a GPU for this, budget for the cost rather than the headline. Half your frame rate is the realistic planning figure on the first shipping title, and the marketing number includes generated frames.
• If you are shipping a game, the masking system is the feature, not the intensity slider. It is the only control that expresses artistic intent rather than degree.
• If you are watching the technique, the thing to watch is whether the runtime inputs widen. A model given real G-buffer data at inference would not need to guess what a surface is, and most of the artistic objection follows from the guessing.