How to make the Neural Network Image Processing tool actually work.

[STS21]

In a Canon USA thread (https://community.usa.canon.com/.../Canon.../m-p/395278...) I found this: 'It needs a keyboard shortcut to invoke the process'.

I've tried using the tool but it makes absolutely no difference to the images, no matter where I put the two sliders.  The manual is brief to the point of being almost useless so I'm wondering if this 'keyboard shortcut' is the missing piece in the puzzle.

Any help gratefully received,

STS21

[jrhoffman75]

Good info. I think you forgot to post the images you are describing.

[jrhoffman75]

I did a quick test with NNI tool.

The image was late dusk. It is underexposed - I limited the ISO to 25600 and shutter speed to 1/60.

Screenshot of original image.

100% crop of original image.

100% crop of NNI tool output. Increased Sharpness to 4 in DPP4.

Next test will be to output just a .CR3 file and then process to see what the difference is.

[jrhoffman75]

The NN suffix file can be opened in DPP4 but isn't recognized by Lightroom.

The .CR3 file generated by NNI starts out looking "better" (less noise) than the original .CR3 but processing in LrC produces pretty close results for both. Slightly more noise for the non-NN version. (Adobe Adaptive Color profile and AI Denoise were used on both images.) 

[Sports_N_stuff]

So, when you are asking the NNiP to generate a CR3 file you'll notice that it alerts you to say it won't be applying the neural network algorithm to "demosaicing." And you'll notice you get a smaller file than you would with a .CRN file.

What it's trying to tell you is this:  There are two advantages you can gain from this additional out-of-camera processing with Canon's advanced statistical image models:  1) additional elimination of noise, and 2) Deeper processing of the RAW file to find additional fine detail within the image.  To distinguish the level of processing a given file has endured, they will only eliminate noise when you save it as a CR3 file.  If you want to benefit from additional interpretation of the original pixels in RAW you indicate that by saving as a CRN.  When you have an image that benefits from this treatment, you'll notice that the additional detail can create a substantially larger file.  When you have an image that benefits from additional processing you will see the most benefit when you let it save to a CRN file. 

If you're most comfortable using external products like LRc, Topaz, etc.  You can open the CRN file and save it as a TIFF file.  Saving to this universal format will allow you to save most of the benefits of Canon processing while using other tools. 

[Sports_N_stuff]

.. so doing some follow-up testing in Lightroom, with an actual  photo I had been upgrading for submission in DPP4, I do find that a CRN file saved as a TIFF does save virtually all of the Canon enhancements when imported into LRc. To my eye, the coloration and dynamic range in the DPP4 - generated image is ever so slightly better, so that the rounding of the facial features might be a little more realistic, but that's  just a bias of mine and they are essentially the same.  I tried both a straight NN-based CRN, saved as a TIFF in DPP4 as well as a CRN saved as a TIFF, which underwent a shade of additional focus-polishing in Topaz saved as another TIFF and LRc was able to produce a JPEG image from both of these that was essentially comparable to DPP4.  An additional advantage in this scenario is that DPP has already filtered out the noise.  For me, if I'm trying to export a regular CR3 from LRc as a JPEG, I have to tune it a bit to get rid of the pixel grain in LRc: Turn the detail down to 0 and run denoise for about 30 seconds to match the image quality I get out of DPP4.  The DPP4-generated TIFF already has that denoising/degraining baked into it even for the NN-based files.

[Sports_N_stuff]

Oh yeah!  When you click on the image posted above, letting it go full screen, and cycle between the two you do see the difference in noise. Sweet.

[JamesHarvey]

From which camera body, with which FW release, did the images come from on which you noticed no change? 

Thanks, I find the same, but only with one of my 3 (different model) bodies.  With the other two the improvements from NN are outstanding.  I'm going to start a thread on this issue that you may want to follow.

/JH