Your RTX received an upgrade: DLSS 4 with transformer!

Nvidia launched, along with the RTX 50 series, an update to DLSS, DLSS 4 with Transformer, which affects all Geforce RTX. The company introduced a new neural model for the operation of increasing DLSS (UPSCALING) and also for the new frame generator (Frame Gen), the transformer, which promises to deliver better results than the previous version used.

What changes in this exchange? First we need to understand what these neural networks are, the “engine” that creates new DLSS pixels and frames.

To apply Deep Learning Super Sampling (DLSS), Nvidia’s video cards work with a set of information. They include information from the game engine, which gives information from the movement vectors (where things are going) and also the pixels themselves of each picture generated.

This information is thrown into a neural network, which is a series of algorithms that seek to find patterns in gross data. A neural network is created at the base of brute force: Nvidia trains models in successive attempts, and then compares with the result that should have delivered, until it reaches a model that hits more.

What is changing is precisely this model. Nvidia has so far used a convolutionary neural network (Convolutional Neural Network or CNNS). These networks have a broad use on photo recognition algorithms in photo, for example.

But Nvidia is now migrating to the transformer, and this neural network has more potential to further stagge the performance of DLSS. Transformer has wide use in the latest generative technologies, with T, chatgtp, being precisely from Transformer (Generative Pre-Trained Transformer, GTP).

And what does it bring potential for DLSS? This new neural network model will be able to retain more information about each pixel, as well as being able to analyze the image more broadly. While the CNN model evaluated the image by region, the transformer will see the scene as a whole and will identify more relevant pixels. This model is also more efficient in using information from previous pictures of the image.

What does this impact? According to Nvidia, the transformer will be better to avoid problems such as sparkling, which is when the edges of an object are “flashing” because DLSS does not stop “changing idea” about the color that a pixel must have. It will also reduce problems such as ghosting, caused by a pixel that has been retained too long with the same color.

But not everything is positive. According to Nvidia, the transformer works with twice as much information, so it needs the quadruple of performance to run. Fortunately, in our tests, even a popular sign like GeForce RTX 4060 does not seem to have been much impacted by using the transformer model. The RTX 4090, in turn, showed variations below our margin of error.

In our tests, we do not notice such an impact on the use of the new model in 4K. Even using performance mode, ie reducing 4x resolution, the CNN model does well in expanding an image from Full HD resolution.

But bringing the test to Full HD, things got more interesting. Here, connecting DLSS in performance mode is having to start from an image with only 560p resolution, a base with little information for the AI ​​work model.

Here we have more noticeable differences, especially in retention of information and the smoothing of graphic artifacts, such as border sparkling. This makes DLSS in performance mode deliver a graphic result more comparable to the native 1080p.

Anyone who has enough performance should still keep DLSS in quality or even Dlaa mode in Full HD. But if the DLSS 4 performance mode is lacking in your system with transformer is delivering much better results in 1080p, with fewer image defects, even starting from such a low resolution of rendering.

The DLSS Transformer model will be available on all DLSS compatible plates, which includes GeForce RTX 20 series forward. Games may bring native support, as is the case of Cyberpunk 2077, but players will also be able to use the NVIDIA APP to “force” the replacement of the CNN model with transformer on games that already have DLSS support. Much better than copying DLL files from one game to another or use third party software.