I just accepted the update today, and I must say that I'm impressed so far. I normally run DeNoise after my initial RAW processing, but this time I processed the RAW files directly for this writeup. Both were shot at ISO 1600 on my K3 II and I manually adjusted the settings in DeNoise, as I find I get better results than with the auto settings. Other than cropping and resizing, the images are untouched. I compared the prior version's DeNoise AI model to the new Severe Noise model. Denoise AI is renamed Standard model in the new version, but I did not compare that to either the prior DeNoise AI model or the Severe model. The first image is a screenshot from the program; the left side is the original RAW, and the right side is the preview using the new severe noise model. You can see there's a drastic difference to say the least, but the real question is how it performs compared to the old models.
For starters, processing speed is improved. That's not just me repeating what the email notification said, there is a noticeable speed increase. On the last version, it took about 15 seconds to export a photo. Now it's under 10 seconds. In terms of results, it's a bit of a mixed bag and I think which mode you use will depend on the specific photo you're working on. For example, look at the second and third photo here. The first photo is the Severe Noise model, the second is the older DeNoise AI model from the prior version (now renamed as Standard in the current version). Noise handling is almost the same, with the severe model probably slightly ahead as one would expect. The roll cage around the driver's head shows up more crisp than in the Standard model, but there's slightly more fringing along the top of the car's body.
In the second set of photos, the Severe model is a clear winner... almost too clear! The red posts in the center of the photo and the text on the banners are much more defined and with almost no artifacts. On the downside, the concrete barrier is smoothed out too much. It is supposed to have a speckled look from all the weathering (as shown in the last SOOC jpg photo), but the Severe model tried to completely smooth out the barrier. Even the DeNoise AI model tried to do this, but to a lesser extent. In a case like this, using the mask function to not apply the denoise to this area of the photo would likely be a necessity. Either that, reduce the amount of noise reduction, and/or increase the amount of original detail to retain. (Or simply leave it as is as an artistic change.)
I only played with it long enough for this little test. I think in general, I'll be sticking to the standard model, but mostly because I shoot low ISO and don't get above 800 very often. If I'm pushing 1600 and up, then I think this new severe model will come in quite handy.