Article · AI translation
Thanks to IT Home reader 卡了别叫 for the tip!
IT Home reported on November 17 that developer Eren Doğan posted on X yesterday (November 16) that he had extracted Nvidia DLSS models from the Nintendo Switch 2. He found that they include the first-generation Transformer model introduced with DLSS 4, with the same model weights as the PC desktop version, but found no DLSS 4.5 model.
According to the code he examined, the Switch 2 has two dedicated presets named H and I. Doğan's analysis indicates that they use a convolutional neural network, or CNN, model. The model is the same size as the standard desktop version, but its input feature grid is only one-quarter the size, which could reduce the computational load and speed up processing. The actual speed improvement has yet to be tested and verified. He believes this may be the lightweight approach previously called “DLSS Lite” or “Tiny DLSS.”
IT Home consulted publicly available information: the size of the input feature grid refers to how large a grid of information the AI needs to analyze when processing an image. A larger grid generally means more information to process; shrinking the grid may reduce the computational load.