Just started playing with it today. These are my early observations
It is a little klunky to get started. You're basically trying to surface a dialog box, which will show the file name of what ever file you've got highlighted and you can add more. You surface the box by going to "extended functions" and then "Start neural network image processing tool." It will open up a second DDP box on your task bar which you may have to click on it show on top of DPP.
Re. Your earlier questions
You can save in either format and see an improvement in the JPEG it will produce not sure if the CRN is the better of the two. Have been using the CRN so far. As the software notes, the "demosiacing" process is not included if you output to CR3, which has something to do with converting from RAW to image in your software. Not sure I'm doing it right, but my install seems to write over the original CR3. You'll note that the CRN is a gigantic file whereas the CR3 stays small and makes JPEGs that are even smaller than the original form of the CR3.
Does it do anything? I think so. Quite a bit actually. Depends on the shot though. When I'm pulling more detail into faces that I'm cropping out of a tiny area the effect is modest. At least it is for an R5ii image on a 70-200 2.8. E.G. if I've got a human head that is 8% of the image area width and 5% of the height, it will definitely smooth out the noise from that camera, and give a somewhat better definition to the eye bulges and the seam under the lower lid. It's not going to surface eyelash streaks out of that dark patch. Is it enough to turn a discard into a keeper? Maybe, if I'm willing to drop my standards a bit for clarity. I'm considering it. On a subject with firmer edges it seems to make a noticeable difference. 2nd E.G. a close-up on the coach's face. Dark lighting & noisy as usual on the bottom of the face. head width is about 16% of narrow image dimension - this time shot with an R1. Big difference decisively moves something out of the "if I have to" or "junk" folder into the keeper folder. chin wisker clouds surface about half the individual whiskers. Even more so with fuzz area that was on top of bald area. The arc, shape and expression of the ridge in his nose takes on dimension. The story is now the intensity with which he is addressing his team. A line on the shoulder becomes individual stitching. Dark fabric aqquires a texture. Brighten in DPP, lift the darks and it's a keeper. 3rd EG R5ii again. Head is about 8% of width. Definitely improves the scattered noise an R5ii will impart in these situations. Some additional detail? Yeah. maybe. Some. I don't use the R5ii as much so that may be part of it. Lack of experience. And it's not going to focus in at the speed you need it like an R1. And I think there may just be more residual background data lurking somewhere in an R1 image file.
As you might expect from a Canon product, I'm not seeing any full "pedal-to-the-metal" settings or output. they try to keep you from hurting yourselve. I've not yet been able to produce images that look like bent plastic as I might wander into trying to rescue something in Topaz. Neural network models tend to jump to conclusions and can insert things that aren't there if you don't know what you're doing but when this tool isn't having much effect, it also doesn't seem to break things. Have not been able to make it put a Topaz type distortion on letters but haven't really tried yet. Have seen it pull embroidery stitching lines out of a letter where you could barely tell they might be there before. Again depends on the picture and how much it can latch on. Haven't yet compared Topaz with Canon's NNIP on an image. In the cases where it can latch on, I'm thinking it might be better.
As best I can tell you're trying to surface a dialog box which will show the name of whatever file you're touching and you can add more. Neural networks are not tiny technology and they take a while to generate the "CRN" file. Similar to the time it takes to compute, smooth and draw a JPEG from a raw file. Don't know much about cuda cores, but it does stack up a few cores with minor loads on my AMD chip. I'm guessing that applying a NN to an image file doesn't distribute as well as basic image processing.
Overall, the didn't invest much in the interface, and the way it behaves with multiple windows enlarged is annoying but overall it seems like it might have soldi fundamentals at a good price.
