Nvidia’s DLSS AI transformer model is now out of beta so we might see more games get patched to make use of it

While it’s fair to say that Nvidia’s RTX 50-series of graphics cards might not go down in history as the greatest lineup of GPUs ever created, the release of DLSS 4 has been far more warmly received. This latest collection of AI-powered upscaling, frame generation, and ray tracing denoiser technologies is the best yet, though one element of it has been in beta form only. Well, not anymore, as the new DLSS SDK now lets anyone use Nvidia’s transformer model.

If you have no idea what that actually is, let me give you a brief rundown as to how DLSS upscaling works. To improve the performance of games, programmers can implement a section of code that activates DLSS during the rendering of a frame. It drops the resolution right down, making the shading stage much quicker to process.

Once that’s all done, the frame is then scaled back to the resolution it needs to be for the monitor. That’s super quick and easy to do, but it often produces small artefacts in the frame, so before the upscaled frame can be rendered further, your GeForce RTX GPU runs a neural network on the image to tidy it up.

In DLSS 2, 3, and 3.5, that network uses what is known as a convolutional neural network or CNN. After six years of developing said network, Nvidia had reached the limits of quality and performance that it could offer, so for DLSS 4 upscaling, DLAA, and Ray Reconstruction, it created a so-called vision transformer (ViT) model.

It’s more complex than the previous CNN, but due to having twice as many parameters (think of these as ‘control knobs’ for the model), the new ViT can produce a better quality output. It is more work for the GPU to process, though, but one can’t have everything.

Anyway, although the option to use the Transformer model was added to some games, such as Cyberpunk 2077, very quickly, it’s officially been in beta form since launch.

However, as spotted by Videocardz, the latest DLSS SDK 310.3.0 (software development kit) on GitHub highlights the fact that it’s now out of beta. That doesn’t mean Nvidia has stopped improving the model; it’s just saying that it believes it’s ready for widespread use by any developer.

I’m not expecting every future game that supports DLSS to automatically offer the Transformer model as an option, as Cyberpunk 2077 does, nor that they will use it by default, but it does mean that developers now have the option to implement it.

More importantly, we might also see current games get patched to give us the choice of faster upscaling or slightly slower upscaling, with a better output. Or better, ray tracing denoising, for that matter.

Whatever one thinks about Nvidia and the RTX 50-series, DLSS 4 is seriously impressive stuff, and with regards to the transformer model, at least it can be used on any RTX graphics card.