What does Capture-NX "clean" noisy images (even with NR turned OFF)?

Software
1

Continuing the discussion from Help understanding Darktable's non-local means algorithm:

Since this topic is diverging from the initial subject (which is basically solved), I decided to start a new one to better follow the discussion.

I have a rather old Nikon D300, an excellent camera that however suffers from a rather large noise at high-ISO (compared to more modern models).
The Nikon software, Capture-NXD, is really a pain to use, but nonetheless does an really good job in getting out the details from noisy NEF files.
The fact is that, event with noise reduction disabled, the processed images look really clean, particularly in terms of chroma noise.

Below I have put a sample of the individual RGB channels, for the same NEF file processed in either NX-D and NR off, and PhotoFlow with LMMSE and 4 false color correction steps (no additional NR applied).

Green channel (left: NX-D, right: PhF):

Red channel (left: NX-D, right: PhF):

Blue channel (left: NX-D, right: PhF):

I’m putting here the NEF file and the NX-D output in TIFF format, in case anyone wants to experiment.

Any idea of what could be the math behind those NX-D results?

2

I’ve no idea about the math, but the PhF image looks sharper than from NX-D,
Maybe the higher noice is coming from there.

3

The PhF images are the direct output of the LMMSE demosaicing, without any further sharpening, while the NXD images are the output with NR switched completely off.

That’s why I think NXD is doing some “magic” smoothing before or during demosaicing… I do not know yet how to make the PhF image smoother.

4

There;s definitely some smoothing going on. The are some red hot pixels which turn into red blobs.

5

Over the years we found that many raw development programs apply processing under the hood, hidden from the user with no indication that something is being done. That’s where problems arise, as users of these programs try other software and complain that these other programs are “adding noise”, produce “wrong colors”, etc., when in fact they’re seeing the real data for the first time.