AI noise reduction still needs much work

Yeah. I fully agree. You are 100% spot on
Pixel peeping your pictures might be really disruptive and needleslly time-consuming IMHO. I am guilty of this myself…

1 Like

Yes. I blithely made capture sharpening automatic. Then it gradually dawned on me that it was unnecessary and (sometimes) ugly.

Less is more, as they say.

For darktable at least, the new capture sharpening tool in the Demosaic module is not designed to be a sharpening tool per se, but rather a correction for lens blur caused by diffraction, AA filters, etc.
Personally, I think it’s fine to have this applied automatically as I’ve found it just helps with fine detail and doesn’t really make the image look sharpened.

1 Like

I’d heard that AI noise reduction was really good, but the examples I’m seeing are pretty nasty. When people praise it, are they saying it gets rid of noise effectively and provides a coherent if inaccurate image, or are they saying it provides an image which is closer to what the image should look like than alternative methods (i.e. as if the shot had had a few EV more light, but was otherwise identical)

There is no “it”, ML-based noise reduction is an umbrella term, there are wildly different implementations in terms of models, complexity, training data, etc.

It’s like saying something about “cake” in general. Some cake is good, some are bad, some are chocolate, some are not, it is impossible to describe them all.

At the moment, reasonable middle ground seems to be training on mid-size patches of structure so that the algorithm gets come context, and punishing large, contiguous deviations in the output (to discourage generative inpainting).

3 Likes

I can report Apple’s Photomator (and I assume Pixelmator Pro in the Creator Studio subscription uses the same code) AI de-noise is also terrible - any details in the dark must be noise, so good luck doing a dramatic low-key edit. And it is oh so slow… like exporting with high quality resampling in Darktable without GPU acceleration slow. It leaves a huge sidecar file four times the size of your raw file, despite which if you want to adjust the denoise applied you somehow have to clear it and start over again… And it’s the only option, no fancy level-of-detail selective wavelet decomposition module with your annual subscription there.

1 Like