Two presets that reduce the reconstruction workload
DLSS lets a game render a frame at a lower internal resolution and then reconstruct it at the output resolution. On Switch 2 this step uses the NVIDIA chip’s AI capabilities. Reconstruction itself takes up some of the time available to display each frame. Reducing that cost is particularly useful for a console that also has to run in handheld mode. NVIDIA’s technical overview.
In his 16 November post, Eren Doğan describes two convolutional networks, presets H and I. According to his analysis, their weights take up as much space as those of standard convolutional models on PC. The difference lies in the number of positions processed in their internal grid. The factor of four refers to this grid. It tells us neither the final display resolution nor that the frame rate is four times higher. Eren Doğan’s original post.
Preset K offers another option
Doğan also says he found the weights of DLSS 4’s first Transformer model, preset K. The newer DLSS 4.5 models do not appear in the files he examined. Here is how he maps the available options on the console.
Setting or preset | Mapping identified by the developer |
|---|---|
Quality, Balanced, Performance | Preset E |
DLAA and Ultra Performance | Preset F |
Ultra Quality | Preset K, from the DLSS 4 Transformer family |
Presets H and I | Convolutional networks with a smaller grid |
These mappings show what the models make possible. A studio may choose differently to suit its engine or rendering budget. On PC, NVIDIA’s documentation lists H and I as reserved presets. NVIDIA’s DLSS programming guide. On PC, DLSS Swapper also lets you manage installed DLSS libraries for individual games.
What can already be seen in games
Back in October 2025, Digital Foundry compared two ways DLSS behaved on Switch 2 and found visible differences in how details were reproduced in motion. Presets H and I offer a possible explanation for the mode sometimes called “Tiny DLSS”. Establishing that link would still require identifying which preset a game actually loads and comparing the same sequences. Digital Foundry’s video analysis.
Key takeaways
The finding is particularly interesting for the flexibility it suggests developers may have. A lighter model can free up time for rendering the rest of the frame. The balance between detail and smoothness will have to be assessed game by game.