The albedo buffer is in linear space.
The normals are simply the world space normals. Miss events have a null-vector as normal.
(EDIT: No, correct is camera space normals. See below.)
I found this in the docs for what the normal buffer should contain:
This buffer is expected to contain the surface normals of the primary hit in camera space. The camera space is assumed to be right handed such that the camera is looking down the negative z axis, and the up direction is along the y axis. The x axis points to the right.
This is a little more specific than simply supplying world space normals. My question is, does it matter? Should we expect better results by supplying normals in the coordinate system specified above.
This means that the world space normal needs to be transformed by the inverse camera orientation.
For the usual pinhole camera that would be a projection of the world space normal onto a basis built from the unnormalized U, V and -W vectors.
(Negative W because the OptiX pinhole camera is defined as left-handed system.)
The result should look like a normal map without scale and bias then.
That said, the OptiX DL Denoiser currently gains little from providing the normal buffer!
It’s much more effective to provide the noisy beauty buffer and the albedo buffer.
That’s also the reason why I didn’t notice my error. I wasn’t expecting dramatic changes.