Original DLSS 1.0 worked by "hallucinating" a full-res image from a lower-res source image. This didn't work all that great and seems to be abandoned at this p

Original DLSS 1.0 worked by "hallucinating" a full-res image from a lower-res source image. This didn't work all that great and seems to be abandoned at this point. DLSS 2.x is a fundamentally different approach, it is standard TAA where the sample weights are set by a ML model based on previous frames and motion vectors. DLSS3 takes this further and actually does interpolation (or, extrapolation?) of additional frame(s) based on the previous "concrete" frame and the motion vectors for objects inside the game.