Neural Restore and Hot Pixels

Hi guys,

I recently got to finally try DT 4.6 neural restore and it’s great! So to what better use we could put it than to breathe new life into old cameras? Off the shelf comes the Pentax K10D - famouse image quality, but max. ISO 1600, and that is a bit noisy. Can we remove the noise through AI? Turns out we can – but with a catch!

In this image, left-to-right, top-to-bottom, see the results of:

  • no denoising
  • denoise (profiled)
  • neural restore denoise after the pipeline
  • neural restore raw denoise

Clearly, each result a bit better than the previous, although there isn’t that much difference between conventional denoise (profiled) and NR denoise after processing. NR raw denoise on the other hand is a lot better, at least in the area of the pouch (it’s also a bit more aggressive, but I think it really is better).

But, alas! It amplifies hot pixels a lot, so much that they can no longer be removed by the automatic hot pixels filter, and at the same time, are clearly distracting. And the 20-year-old K10D has a lot of them! So I would have to use retouch on each one.

But as I understand it, there is no way to run any filter or processing step before the NR raw denoise, or is there? Even if I export minimal processing without hot pixels as a TIFF, that’s not RAW anymore, so no raw denoise.
Another idea would be to train a denoising model that removes hot pixels, but I am not sure how to go about that and if it’s feasible.

Any thoughts on that matter? FWIW, I will also upload the RAW if anybody wants to inspect a sample (license: public domain).

IMGP2101.DNG (16.1 MB)

1 Like

Forgot to say, that is with the default models, “raw denoise nind 1.0”.

There is a ‘remove hot pixels’ filter in g’mic, whose plugin you can use in krita or gimp. I haven’t tried it on this image.

Can’t you move the (dartable) “hot pixels” module below the NR raw denoise?

Neural Restore isn’t a pipeline module, but an operation that runs on the original raw and produces a noise-reduced raw.

I actually prefer to use denoise rather than raw denoise because then it can be applied after non-AI denoising and hot pixels. Willing to stand corrected if I am mistaken about this.

Tested with your file and the result without hot pixels is using the denoise (profiled) or better neural restore denoise after eliminating the hot pisels using the hotpixels option.

1 Like

Yes, I can remove hot pixels and then denoise. I was just wondering if there was any way to make NR raw denoise working without amplifying hot pixels, because it seems to give the best results.

You probably have more experience about it than I do, I need to test it out more. True, I havent tried the combination of either raw or non-raw NR denoise with conventional denoising that much yet.

I have very limited experience with the AI denoising in DT. Most of the time non-AI is sufficient for me but on rare occasions I will bring in the big guns of AI denoising. I am certainly glad of the option being in DT to complement non-AI denoising but it is not my first approach.