Games hand data to all DLSS algorithms through named parameters, so the parameter list is effectively the input specification, and we can fish that out of the file. For DLSS 5 it is short: the rendered image ("Backbuffer," "Color"), depth and motion vectors ("Depth," "MVec"), UI handling ("UI," "UIAlpha," "UICorrection"), masking ("ControlMask," "UseAutoMask"), some kind of bidirectional distortion field shared with Frame Generation ("BidirectionalDistortionField"), plus output and sizing parameters including a "ScalingRatio."
Interestingly, this is a much smaller set of inputs than with DLSS 4.5 Ray Reconstruction, which accepts roughly 25 buffers—diffuse and specular albedo, normals, roughness, ray hit distances, transparency layers, subsurface scattering guides. DLSS 5 accepts none of them.
At SIGGRAPH, NVIDIA said they used "the render's internal buffers such as albedos, surface normals, and some of the lighting information" so the model would know what to stay consistent with. This confused us at first, but then it became clear that the spokesperson framed this statement about training: "we trained the model in a way to..." We also checked whether DLSS 5 might receive G-buffer data by grabbing that data from DLSS Super Resolution or Ray Reconstruction—it does not reference either file.
Two further details. "UI," "UIAlpha" and "UICorrection" alongside "Backbuffer" place this at the very end of the frame pipeline, after tone mapping, with the interface potentially already composited—the model has to be told where the HUD is so it does not break it. And "ScalingRatio," together with log messages tracking network dimensions separately from output dimensions, shows DLSS 5 runs at its own separate resolution, i.e. it does some sort of scaling, too.
