| Garrus said: DLSS2 is based on TAA, temporal accumulation is used to render each frame (several previous frames are combined to render the current one), so yeah, frame rate makes a huge differences to TAA and all it's derivatives FSR2 and DLSS2 |
There are two ways DLSS 2 circumvents the artifact issue that TAA doesn't, which in the end make it less dependent on prior frame deviations.
1. With TAA, there are heuristics hard coded into the algorithm. For example, there is a heuristic that the data used from previous frames shouldn't deviate too much from the current frame. If you have a low framerate, then the deviation is more likely to be too much, and the TAA solution won't be as effective. With DLSS 2, these heuristics are not hard-coded but determined in the training stage, many of them fixing the image beyond what a high-resolution target would even show. That's why DLSS 2 can give better image quality than a native image at the target resolution, whereas TAA could never do that.
2. DLSS 2.0, unlike DLSS 1.0 (and of course unlike TAA because it isn't based on an autoencoder), has a generalized model which isn't trained on a per game basis. The heuristics learned by the auto-encoder can be applied generally to all games. Again, this means that there is less dependency on prior frame's data, and much more dependency on generalized "learned patterns."
In the end, this makes frame-rate matter a lot less for DLSS 2.0 than TAA. The model's feature weights and training input are far more important than anything a run-time.
