DeNoise-AI Auto settings rarely in the ballpark

Maybe it’s my system, but the Auto feature almost always guesses wrong…badly. Does it typically get in the ballpark for others?

1 Like


I find it rather hit and miss.
On average images which don’t have significantly visible noise, the auto remove noise settings mostly seems reasonable, but when an image has significant noise that is often not the case.

As what I think is a good example.
The link will allow you to download an original ex camera RAW photo of some books in a bookshelf, taken with a Sony SLT-A57 camera at 1/640s F/8 at ISO 16000. The filename is Sony_a57_20.arw.

When I open that image in DeNoise AI v 1.2.1, the auto setting for Remove noise is 0.56 with Enhance Sharpness at 0.49. If at 100% zoom I position the image so that the large green “Photoshop Elements for Mac” is positioned at the left of the view, and the dark covered “Gulbins Fine Art Printing …” at the right of the view, then flicking between the updated preview and original shows that it has removed a fair amount amount of the noise from the dark cover book at the right, but only had little effect on the noise on the light green book on the left hand side.

If I increase the Remove noise slider to its maximum 1.0 there is some limited improvement, but there is still significant noise on the green book. By contrast, if I set the model to AI Clear the Auto Setting for Remove Noise is High, and the noise on the green book is gone. I can even reduce the Remove Noise to Medium and still get a good result.

The link to the full Photography Blog web page is

Scrolling down that page to Sample Raw Images will enable the download of the same view taken at different shutter speeds and ISO settings for further comparison.

All my previous testing showed that Denoise AI works best for low to medium noise while for high noise pictures AI clear is best. Both are available in the same app. The Denoise auto setting seems to work fairly well for the low to medium noise samples but on some pictures it under estimates the noise level.

Oddly, that seems to be the reverse of what I would expect. I get an underestimate virtually every time from AIDN…typically, in the opposite direction as needed.

I think we both said that it underestimates the noise level in auto mode. I just found that if noise is low it does an adequate job. One work around is to use auto and then raise the noise reduction slider before updating the preview. I think Topaz should evaluate the auto mode and have it choose a higher reduction level.

Yes, it is self-evident that manually raising the Noise slider is the obvious solution because one doesn’t really need to use Auto at all; for me Auto is useless; hence, my point. When I say “opposite” I mean Auto driving the initial setting to the left of the default staring point…0.15. If Auto drove the setting to the right of the default (0.15), but still not enough (ie needs more), that might be more understandable. But, the value that Auto came up in this example, and in almost all of my cases, was less; 0.12 (ie less than the default)…which typically leaves glaringly obvious noise still visible, in my experience.

A long way of saying that Auto just doesn’t seem to work at all (for me, anyway).