DLSS 5 is brilliant.

[HoloDust]

@sc94597

I think you're confusing nVidia's quote of how the model was trained with what it actually takes as inputs at runtime.

Nvidia's own technical paper have clarified that while they used G-buffers like albedo, normals, and depth during training as 'consistency supervision' to teach the model, the network does not receive them as explicit runtime inputs. At inference, DLSS 5 only takes the rendered color frame, motion vectors, and temporal history. It does not have live geometric awareness or material data. It's literally looking at a flat 2D image and using its training to guess (or hallucinate) what the lighting and materials should look like.

That is exactly why it can't understand off-screen or occluded geometry, and why AMD's paper is IMO better path forward. Their approach doesn't just look at a flat frame, it's natively integrated into the rendering pipeline. It explicitly uses actual runtime scene data (normals, albedo, roughness, metallic) and live radiance cues to generate the indirect illumination. From what I understand, AMD approach actually knows the live geometry and material properties of the scene it's generating illumination for (although it's still screen-space approximation of GI), whereas DLSS 5 is just a highly trained 2D post-process approach.

That said, I understand why some folks are mesmerized with DLSS 5 - it works today (albeit at heavy cost), and it gives...well, (arguable) results.