NVIDIA DLSS 5: GeForce Game Ready driver 616.64

5 septembre 2026 · Kris Wilhem

NVIDIA DLSS 5: Neural Rendering arrives, and ComfyUI is already starting to exploit it

NVIDIA officially launched DLSS 5 on 3 September 2026, together with the GeForce Game Ready 616.64 driver.

The major new feature has a name: 3D-Guided Neural Rendering.

With DLSS 5, NVIDIA no longer limits itself to rebuilding an image at a higher resolution or generating intermediate frames. The system now adds a genuine real-time generative neural rendering step, able to act on the final appearance of a scene.

Light, materials, skin, hair, shadows, light diffusion and micro-details can thus be enriched by a neural model.

The 3D engine nevertheless remains the structural reference: geometry, textures, lighting, motion vectors and other information from the render serve as a guide for the model.

The neural network then steps in as the last stage of the rendering pipeline.

The first game to officially integrate DLSS 5 is NBA 2K27.

An important change in the philosophy of DLSS

Previous generations of DLSS mainly sought to rebuild more efficiently an image the engine could have produced with more resources.

With DLSS 5, NVIDIA goes further.

The model holds visual knowledge learned from real images and can bring in information that the 3D engine does not necessarily simulate completely.

This concerns in particular:

  • the behaviour of light on skin;
  • subsurface scattering;
  • the diffusion of light in hair or vegetation;
  • the perception of materials;
  • certain micro-details;
  • the coherence and richness of lighting.

NVIDIA speaks of a single-step pixel-space diffusion model, designed to run in real time and preserve the structure defined by the artists and the 3D engine.

In other words: Neural Rendering does not replace classic rendering, it complements it.

What hardware is required?

At launch, 3D-Guided Neural Rendering is officially available on the GeForce RTX 50 Series.

GeForce RTX 50

This is today the platform officially supported by NVIDIA:

  • RTX 5090
  • RTX 5080
  • RTX 5070 Ti
  • RTX 5070
  • RTX 5060 Ti
  • RTX 5060
  • compatible RTX 50 Laptop variants

The launch goes in particular through the new driver:

GeForce Game Ready 616.64

GeForce RTX 40

The RTX 40 cards do not yet have official support for 3D-Guided Neural Rendering in NVIDIA's current matrix.

But the situation is more interesting than a plain "not compatible".

Community experiments have already managed to get DLSS 5 running on certain RTX 40 cards with unofficial runtimes and methods.

NVIDIA has also said it intends to extend official support to the GeForce RTX 40 later, once the model is further optimised on the RTX 50.

For the moment:

RTX 40 = experimental is possible, but not yet officially supported.

RTX 30 and earlier generations

Experiments also exist on certain RTX 30 cards.

The compute cost, however, becomes much higher and performance can drop sharply.

Here we are clearly in the realm of hacking and experimentation, not of a feature officially offered by NVIDIA.

And the RTX PRO Blackwell?

The situation is different for professional cards.

An RTX PRO 6000 Blackwell, for example, obviously has a recent enough architecture and considerable compute power.

But the RTX PRO cards use an Enterprise driver branch that differs from the GeForce Game Ready branch.

So the problem is not necessarily the hardware.

You also need:

  • the compatible driver;
  • the matching NVIDIA NGX runtime;
  • and in particular the Neural Rendering runtime `nvngx_dlssnr.dll`.

At the time of this publication, the professional branch does not yet provide the same DLSS 5 environment as the GeForce 616.64 driver.

For the RTX PRO cards, we will therefore have to wait for the compatible runtime to arrive in NVIDIA's professional branch.

And for ComfyUI?

This is probably the most interesting part for AI creators.

An experimental, unofficial open source project called ComfyUI-DLSS5-NR already gives access to NVIDIA's Neural Rendering directly from ComfyUI.

The project talks to NVIDIA NGX through a small native bridge using D3D12.

The pipeline then potentially becomes:

AI generation ↓ Krea / Klein / FLUX / another model ↓ Generated image ↓ DLSS 5 Neural Rendering ↓ neural work on light, materials, skin and micro-details ↓ Final image

It is not an upscaler

This is an important distinction.

The DLSS5-NR node does not change the resolution of the image.

An image of:

`2048 × 2048`

stays:

`2048 × 2048`

after processing.

DLSS 5 Neural Rendering should rather be seen as a form of:

Neural Finishing Pass

The model acts on the appearance and perception of the image, not on its definition.

This opens an approach quite different from our usual workflows:

Generation → Optional upscale → Neural Rendering → Finishing → Output

or, depending on the results obtained:

Generation → Neural Rendering → Upscale / Detail → Output

The two approaches will have to be compared experimentally.

A node that is still very experimental

It must nevertheless be made extremely clear:

ComfyUI-DLSS5-NR is not an official NVIDIA or ComfyUI integration.

The project currently exploits an NVIDIA NGX interface that is not yet publicly documented.

It requires in particular:

`nvngx_dlssnr.dll`

This proprietary DLL is deliberately not distributed by the project.

Compatibility therefore depends directly on the combination:

GPU + NVIDIA driver + Neural Rendering runtime

A GPU technically able to perform the computation can thus be refused by the runtime in use.

So this is still genuine research territory.

And already a temporal mode for video

Version 0.3.0 of ComfyUI-DLSS5-NR already goes beyond processing still images.

It introduces a mode:

`temporal`

intended for image sequences.

The node then uses the NVIDIA Optical Flow Accelerator (NVOFA) to estimate motion between frames.

The principle becomes:

Previous frame + Current frame ↓ NVIDIA Optical Flow ↓ Motion vectors ↓ DLSS 5 Neural Rendering ↓ Processed frame

The aim is to bring better temporal coherence to the neural processing.

For our AI video pipelines, this part obviously deserves particular attention.

Why DLSS 5 now goes beyond gaming

This is probably the most interesting point of this announcement.

Until now, DLSS was mainly associated with:

  • upscaling;
  • anti-aliasing;
  • Ray Reconstruction;
  • Frame Generation;
  • performance improvement in games.

With DLSS 5, another logic appears:

using a generative model as a genuine component of the rendering engine.

And only a few days after its release, this technology is already being experimented with in ComfyUI, independently of any game engine.

This potentially opens much wider uses:

  • finishing of generated images;
  • perceptual improvement of materials;
  • skin processing;
  • light enrichment;
  • micro-details;
  • AI post-production;
  • sequence processing;
  • generative video;
  • hybrid workflows, classic rendering + AI.

A young technology, but a frontier has just moved

It is far too early to consider DLSS 5 Neural Rendering a new standard production tool in ComfyUI.

The runtime remains restrictive, hardware support is still limited and the ComfyUI project is explicitly experimental.

But the technological shift is interesting.

We are starting to see real-time generative Neural Rendering leave the video game pipeline alone and meet AI image and video generation pipelines directly.

And this convergence between:

3D rendering + AI generation + neural processing + ComfyUI

could become far more important than a mere new DLSS version number.

Sources

  • NVIDIA: DLSS 5: Generative Neural Rendering
  • NVIDIA: DLSS 5 / 3D-Guided Neural Rendering
  • NVIDIA: GeForce Game Ready Driver 616.64
  • NVIDIA: DLSS Supported Hardware
  • GitHub: ComfyUI-DLSS5-NR
  • GitHub: ComfyUI-DLSS5-NR v0.3.0