Neural Network Processing and/in the R5ii

[JamesHarvey]

I'm trying to determine the degree to which, and how, the Canon Neural network processor is incorporated into the firmware of the R5ii, so I can best configure my workflows, which commonly use the Canon NN processor app as the first treatment for selected raw images. 

I recently ran a set of tests of various R body & lens combinations using the moon as a target on a night with relatively low atmospheric distortion issues.  Bodies used were the R5ii, 6ii, & 7.  Lenses were the RF100-500, 200-800, & 800f/11, plus the 1.4x TC. I did not try all combinations, but only those I was most likely to use for general super-telephoto applications, considering resolution, diffraction, DR etc. Select results from these tests may be found here:  

https://photos.app.goo.gl/DMMEAusMWZ6s1ESXA 

One very interesting finding was that whereas the images from the R6ii & R7 showed considerable improvements, as expected, when processed through the NN app, those from the 5ii did not show as much improvement, as though some NN processing had been applied to the raws in camera.  I am aware that the R5ii does have some NN processing capabilities, but I had assumed that these were applied to camera-generated jpeg versions of the images, and the controls on the menu seem to point to this.  However, I also note that the output from the NN app is itself a raw (uncompressed CRN or compressed CR3) image to allow further processing, so would I be wrong to assume that, in the R5ii, some NN processing is automatically applied to the raw images?  If so, is there a technical brief that describes how this may relate to how NN is applied in the R5ii (& R1); I can't find reference to it per se in the 5ii (or R1) manual? 

[Cyrilbrd]

Hi, not a professional here but the R5 II spec sheet explicitly lists Neural Network Upscaling and Neural Network Noise Reduction among its image processing features.

It seem that the “in-camera NN noise reduction” is likely applied only in the path that prepares JPEG/HEIF (or in capture preview) rather than to the raw capture file that you would feed to your external NN app...

Again not a specialist here but it is also possible that certain internal firmware settings (e.g. default noise profiles, default lens correction / shading / distortion / chromatic aberration) are more aggressively baked in for R5 II, so the raw baseline is “cleaner” or “closer” to what NN processing would produce, hence the smaller incremental benefit from your post-NN pipeline.

[JamesHarvey]

Thanks for your response but effectively your two last paragraphs seem potentially contradictory.  It's unlikely that the baked in processing would not use NN processing, as otherwise the results would be inferior to what NN can achieve, certainly for NR; why would you apply better quality corrections to the jpeg output than to the raw? .