Since I’m currently involved in astrophotography, I noticed that the PHAI denoisemodels may contain astrophotos for training.
I find this as such critical since these naturally flow into the final result and thus falsify it, because the photographic genre as such are very different among themselves and have different image processing goals.
Also, it could be that in the training data are images from old
cameras that had a problematic noise.
Problematic in the sense that these were often simply cut off in the shadows when editing.
For example, it was known in the past that Canon cameras before 2016 were significantly worse in terms of dynamic range than Sony and Nikon cameras, because both of the latter used Sony sensors, which were superior at the time.
So the training data from these “problematic” sensors would have to be treated separately.
It would make sense to have four models that are each trained with the specific data.
Normal (RAW Cameras 2017+)
Old (RAW Cameras 2017-)
Processed images (tiff, jpeg etc.)