raytracingfan
Newcomer
I suspect that the change to the transformer model is setting the stage for Nvidia's future neural rendering plans and was motivated by more than just the pursuit of better upscaling quality. Maybe a future version of DLSS will have post-processing handled directly by the transformer model. It could also make it easier for DLSS to accept additional inputs beyond color, depth, motion vectors, and normals, and the transformer model might have a longer and better "memory" of the game world by keeping more temporal data. For example, Reflex 2 Frame Warp provides camera position as an input and presumably uses the transformer model to inpaint disocclusions, this might not have been possible with the old CNN model. It's also possible that other neural rendering tech like NRC and neural materials could provide extra data to guide the DLSS transformer in the future.
