DLSS 5 in detail: Three AI models on one Geforce graphics card

Special Jul 21, 2026 10:50 AM Sven Bauduin Add as a preferred source on Google

Source: Nvidia / Montage: Sven Bauduin

Nvidia has explained DLSS 5 in detail for the first time at SIGGRAPH 2026: three AI models, masks for individual objects and operation on a single graphics card are all confirmed. Geforce RTX 5000 owners are set to gain access in the autumn.

Under the title "Next Era of Graphics: Neural Rendering, World Models & Simulation", Nvidia used its research keynote at SIGGRAPH 2026 in Los Angeles to provide the first detailed technical breakdown of the technology announced in March this year. The US company offered an extensive look at what players can expect from this autumn.

A complete reinvention of computer graphics is taking place.
— Jensen Huang, Nvidia-CEO

Lead developer Edward Liu had already outlined the basic concept behind DLSS 5 ahead of the conference, while Neural Rendering first appeared in the graphics driver in May. Nvidia has now provided specific details about the model, its tools and the required hardware. The keynote also answered how much control developers will retain and whether a single Geforce graphics card will be sufficient.

Neural Rendering in games: How DLSS 5 works in detail

The presentation focused on three areas: the architecture of the AI model, the control tools available to developers and the hardware requirements. The details were presented by Edward Liu, Director of Applied Deep Learning Research, and Gabriele Leone, who leads Nvidia's team of technical artists and art directors.

Recommended editorial content [EMBED_URL] At this point, you will find external content from [PLATTFORM]. To protect your personal data, external content will only be displayed once you confirm this by clicking "Load all external content": I agree that external content may be displayed to me. This may result in personal data being transmitted to third-party platforms. More information can be found in our Privacy Policy.

External content Find out more in our Privacy Policy.

According to Nvidia, DLSS 5 must overcome three challenges: preserving artistic intent, maintaining stability from one frame to the next and delivering the highest possible performance with real-time graphics at 4K/UHD.

Challenge 1: Preserving artistic intent

The first challenge is preserving artistic intent.

Even if we give a generative model a rendered frame, it will change the image. That is ultimately what it is designed to do.

Every change, however, creates a risk of moving away from the original intent.

Suppose a character has a scar. The model may not like the scar and might remove it. The image could look cleaner afterwards, but it would no longer show the correct character.

That is why we have drawn a clear boundary. We train the model to respect these artistic constraints.

To do so, we use internal information from the renderer, including albedo data, surface normals and lighting information. This allows the model to understand which elements must remain unchanged.

Within those boundaries, it may improve the properties that increase realism:

more realistic subsurface light scattering through skin,
more detailed material responses,
improved light transmission through hair and leaves,
more realistic contact shadows,
better ambient lighting.

The scar may therefore receive more realistic skin shading, but it must never disappear.

DLSS 5 must change the image without changing the story.

That is the first challenge.

One-step diffusion: How the DLSS 5 AI model processes images

DLSS 5 is based on a compact, one-step pixel-space diffusion transformer distilled from substantially larger foundation models. It processes each image causally according to the principle of "one frame in, one frame out". The new model does not look ahead to future frames, but instead enhances each rendered image individually in real time.

DLSS 5

Preserving Artist Intent Frame by Frame Streaming Real Time 4K

Full-screen comparison

Its inputs include the colour buffer and the engine's motion vectors, keeping the output anchored to the original 3D content. According to Edward Liu, DLSS 5 does not alter the geometry of a scene. Instead, the technology enhances contrast, ambient occlusion, contact shadows, reflections and subsurface scattering at 4K/UHD ("2160p") and in real time.

Challenge 2: Frame-to-frame stability

The second challenge is temporal streaming.

Most video generators work with several frames at once. They analyse a group of frames and then generate several new images as a batch.

That is not possible with real-time rendering. A game must respond to input immediately and render every frame individually.

Within a game engine, however, we have access to motion vectors. DLSS has used them since its first version.

Motion vectors show the model how objects move between consecutive frames. The model therefore does not have to infer that movement from the pixels itself.

With their help, along with other techniques, we were able to make the model operate causally and frame by frame:

One frame in, one frame out.

The model does not need to look at future frames. Even so, the image remains temporally stable and coherent.

No flickering, no drifting and no detail swimming.

Three AI models and object masks for developers

DLSS 5 allows developers to choose between three AI models and switch between them for individual scenes or cutscenes. It also provides "Structure Intensity" and "Tone Intensity" controls, along with masks that can limit the effect to individual characters, objects or environmental details.

Generative models can produce remarkably photorealistic images. But a game needs more than a beautiful image. Every single frame must remain consistent with the game world created by the developers. — Edward Liu, Nvidia

Gabriele Leone demonstrated the tools using several game characters, which appeared visibly more realistic than the underlying rendered versions thanks to higher-contrast lighting and more detailed contact shadows.

Source: Nvidia

Source: Nvidia

The presentation itself was also notable. Rather than replaying the controversial comparison video from GTC, Nvidia showed a clip lasting just under 30 seconds and demonstrated live how precisely the AI effect can be controlled.

The response to the backlash in March is unmistakable. At the time, the "AI faces" produced through Neural Rendering sparked a heated debate about artistic control. Integration into games will continue to use the established Streamline SDK.

Challenge 3: Speed

The third challenge is speed.

A 4K image consists of around eight million pixels. At more than 60 frames per second, there are fewer than 16 milliseconds available for each frame. The game itself already consumes most of that time.

The underlying generative models capable of producing photorealistic images are large and slow. They can generate almost any kind of content, but require several seconds to do so.

Real-time rendering needs something more specific: a model that is smaller, more efficient and specialised for a single task.

We therefore transferred the knowledge of large generative models about how the world looks into a substantially smaller model.

It is a one-step diffusion transformer that works directly in pixel space and has just one task:

making real-time rendering look more realistic.

Even so, the model is not a conventional image filter. It has inherited the broad understanding of the world held by large foundation models.

It understands characters, lighting, materials and the context of an image. What it has lost is general versatility. It can no longer generate funny cat videos, for example.

Because it only has to perform this one task, however, we were able to make it compact and efficient enough to run alongside a game without the experience ceasing to feel like a game.

This solves all three problems:

preserving narrative and artistic intent,
maintaining temporal coherence from frame to frame,
staying within the available time budget.

Generative AI can therefore become a direct part of a renderer.

DLSS 5 launches for Geforce RTX 5000 in the autumn

DLSS 5 is due to launch in autumn 2026 and is expected to require only one graphics card at release. The first GTC demonstration still ran on two Geforce RTX 5090 cards, one of which was dedicated entirely to processing the AI model. By launch, however, Nvidia intends to have optimised the technology for operation on a single graphics card.

The manufacturer has yet to provide specific system requirements or performance figures. The Geforce RTX 5000 series ("Blackwell") is currently regarded as the baseline. The first supported titles will include Starfield, Assassin's Creed Shadows, Hogwarts Legacy and The Elder Scrolls IV: Oblivion Remastered. Publishers including Bethesda, Capcom and Ubisoft have already pledged their support. The full research session is listed in the SIGGRAPH 2026 keynote programme.

Quelle: Nvidia

DLSS 5: Three AI models, single-GPU operation and full control Nvidia has explained DLSS 5 in detail for the first time at SIGGRAPH 2026: three AI models, masks for individual objects and operation on a single graphics card are all confirmed. Geforce RTX 5000 owners are set to gain access in the autumn.

Send via email