With the start of the new Blackwell GPU generation, NVIDIA not only released new hardware, but also a powerful software upgrade for all RTX graphics cards: The Transformer model for DLSS 4. This new AI model not only improves the image quality in the high scaling of games, but also addresses many weaknesses of the previous CNN models (Convolutionary Neural Networks). In an extensive series of tests, Digital Foundry has known problem areas of older DLSS gameslooked at again, but this time with the transformer model, which visibly improves many of the previous weaknesses.
Visible improvements in DLSS 4
The new DLSS model can therefore be selected directly in the menu in current games, but older titles need third-party tools such as DLSS Tweaker to manually integrate the current "K" version. Although Nvidia also offers a DLSS selection in its new app, it does not always work reliably-a problem that the manufacturer still has to fix. Overall, however, it shows that the transformer model offers higher efficiency with a comparable or better image quality than earlier versions, albeit with a slightly higher resource consumption.
In practice, according to Digital Foundry, the new DLSS model is clearly superior in many games: in Death: Director's cut, for example, is improved by the sharpness without creating new aliasing problems. Nevertheless, there are also restrictions here, for example in the raindrops, which the AI are still incorrectly interpreted and suppressed by the AI. In God of War: Ragnarök orReduces the new model visible artifacts on fine objects such as telephone wires or hair, which is clear progress compared to previous DLSSION.
Again, how difficult to reconstruct details such as telegraph lines for DLSS remain, regardless of whether you use CNN or transformers as a model. According to the technology experts, both variants reach their limits here, which rather reflects the engine and its original use of MSAA than weaknesses in DLSS itself. However, the transformer model visible motion blur and ensures more clarity for camera swings, especially on high-quality displays such as OLEDs. However, there is no improvement in the reflections in Ratchet and Clank: here the game itself seems to be the cause of artifacts.
Inthe transformer model was also tested, with mixed results. The ghosting effects in moving grass were significantly reduced, but at the same time the new clarity led to more visible aliasing. Nevertheless, the overall result was positive. But even if many weaknesses have been eliminated, there are still challenges: According to Digital Foundry, DLSS 4 leads to additional image noise in combination with hair rendering, and inOccurrence artifacts in volumetric fog.
The conclusion of the technology experts: The Transformer model from DLSS 4 significantly increases the image quality of many titles and addresses long-term criticisms from DLSS. But it is not a panacea: some visual problems remain or shift, and not every game benefits to the same extent. Nevertheless, there is a lot of potential in the new technology, the first iteration of which already shows promising progress. Since NVIDIA has set the development of the CNN model and relies on the transformer future, DLSS 4 should be further refined in upcoming updates.
