At CES 2026, one of the highlights from Nvidia’s showcase was the announcement of DLSS 4.5, the next (half) step in the company’s upscaling tech in its RTX cards. And one of the main upgrades is an updated, second generation transformer model.
If you recall, Nvidia previously made the shift to a transformer neural network model in DLSS 4 last year for Super Resolution, Ray Reconstruction and DLAA features. Compared to the convolutional models in previous DLSS versions, the transformer architecture allowed for more stable images, with less ghosting and smearing.
DLSS 4.5’s transformer model is also based on an expanded training data set.
Image: Nvidia.






