Denoising high ISO images // Test: darktable vs Neat Image (Linux user)

Dear all

I remember very well Deed's denoising challenge,Topaz AI noise reduction versus other software.

Recently I had to take photos at ISO 6400, it was portraits. I am not eligible to publish them. But I was not satisfied with my skills in denoising with darktable. So I was seeking a more satisfactory solution.

Of course many use Topaz DeNoise AI. But I am on Linux, and switching the machine for denoising would mean some hassle for me.

Sluggy warrior recently mentioned Nind Denoise , a Linux tool using convolutional neural networks that was written as a Master thesis. To me, it seemed a bit complicated to get it running. Maybe I am wrong.

A friend reminded of Neat Image. I tested it one week and found it good. Big advantage for me, it runs on Linux.

Below please find two examples that I can publish. It is Pablo, one of our dwarf rabbits

Sample 1. 6400 ISO.

X-T4, XF 50/2, ISO 6400, OOC, NR -4 (my standard).

What I could get from darktable. Denoise profiled, raw denoise (I apply raw denoise only for real high ISO images), sharpen.

What I exported from darktable as a tiff (no sharpening, no denoising).

What Neat Image made from the darktable tiff export. I got this result very quickly. No hassle.

Sample 2. 12800 ISO.

X-T4, XF 50/2, ISO 12800, OOC, NR -4 (always my standard).

What I could get from darktable. Denoise profiled, denoise raw (I do this only for high ISO images), sharpen.

What I exported from darktable as a tiff (no sharpening, no denoising).

What Neat Image made from the darktable export. Very quick procedure.

Screenshot from Neat Image, 12800 ISO. The selection shows a preview of the result after denoising & sharpening. The rest is from darktable tiff export, without any denoising and without sharpening.

My workflow with darktable and Neat Image

• My standard workflow: My darktable workflow to provide a decent baseline // Updated (darktable 4.0)
• When I apply Neat Image, I do everything but I deselect sharpen and any kind of denoising in darktable
• I export the image as a tiff
• Import into Neat Image
• Process to taste
• Export final jpeg from Neat Image
• DONE

My assessment

• Meanwhile I am convinced that those professional denoising tools have a real justification.
• Neat Image delivers a way cleaner output than Fujifilm OOC as well as what I can do with my personal darktable skills. It is pleasing, smooth, sharp, and not waxy (if not overdone).
• I cannot tell how far really skilled darktable users can come. But with Neat Image, I get a great result within one minute.
• I also tested Topaz DeNoise AI and found it good. I decided for Neat Image because my OS is Linux. Otherwise Topaz would certainly have been a good choice, too.
• Now I am not afraid any more of high ISO color photos.

What is important with Neat Image

• Acc. to my experience, a lot depends on the proper selection of the sample region. Oftentimes, Neat Image does a good job but not always. If you're not happy with the outcome, try a different region. It must not contain a structure.
• Don't overdo denoising. I sometimes reduce the Luminance by 10 % points. I never increase. If I am not happy, I start anew starting from a new sample region in the image.
• I usually do apply sharpening with Neat Image. Standard settings are mostly good but I typically enhance the medium level a bit.

Btw, animal hair are a real challenge.

Regards,

Martin

I suggest you use selective denoise and sharpening.  You will get much nicer results and will be able to use more aggressive denoise.

Morris

Morris0 wrote:

I suggest you use selective denoise and sharpening. You will get much nicer results and will be able to use more aggressive denoise.

Morris

Thank you Morris.

In contrast to darktable, Neat Image does not offer the option to work with masks, at least not in the current version 9 and when used as a standalone. But of course one can produce multiple outputs and then selectively overlay with a software that allows to work with layers, such as GIMP.

In case you meant to work with darktable only and selectively denoise certain areas more and others less with darktable - well, yes, darktable offers excellent masking (no automatic recognition of shapes, though), but I personally with my skills am unable to achieve the same or similar level of denoising with darktable as Neat Image does, also locally. So that's not an option for me.

Regards,

Martin

Rightsaidfred wrote:

Morris0 wrote:

I suggest you use selective denoise and sharpening. You will get much nicer results and will be able to use more aggressive denoise.

Morris

Thank you Morris.

In contrast to darktable, Neat Image does not offer the option to work with masks, at least not in the current version 9 and when used as a standalone. But of course one can produce multiple outputs and then selectively overlay with a software that allows to work with layers, such as GIMP.

In case you meant to work with darktable only and selectively denoise certain areas more and others less with darktable - well, yes, darktable offers excellent masking (no automatic recognition of shapes, though), but I personally with my skills am unable to achieve the same or similar level of denoising with darktable as Neat Image does, also locally. So that's not an option for me.

Regards,

Martin

I paint my masks with a soft brush in Photoshop. With some practice it's quick and easy.  Possibly you can do something similar

Morris

I've had good luck with running Topaz Denoise in wine.

rlx wrote:

Morris0 wrote:

...

I paint my masks with a soft brush in Photoshop. With some practice it's quick and easy. Possibly you can do something similar

Morris

I think NeatImage does that automatically but with its own algorithm. It won't de-noise the contour areas and it won't apply sharpening on low contrast parts of the image. I find it quite good and quite useful for batches of pictures when I need it.

Interesting, I use V8.  That might be a V9 feature worth the upgrade

Thank you,

Morris

Morris0 wrote:

rlx wrote:

Morris0 wrote:

...

I paint my masks with a soft brush in Photoshop. With some practice it's quick and easy. Possibly you can do something similar

Morris

I think NeatImage does that automatically but with its own algorithm. It won't de-noise the contour areas and it won't apply sharpening on low contrast parts of the image. I find it quite good and quite useful for batches of pictures when I need it.

Interesting, I use V8. That might be a V9 feature worth the upgrade

Thank you,

Morris

Morris

I have v9.0.2, 64 bit, standalone.

The Pro version for 16 bit images because I like to work lossless with tiffs.

Bought it about one week ago.

Before I tested.

But I assume this has been part of the algorithm already in earlier versions.

Cheers,

Martin

Thank you OP. This is a fantastic thread. I'll check out Neat Image.

Have you had any luck getting X-H2S raw images to work with Darktable? I am unable to open any raws using Darktable. I'm wondering if there is some kind of patch I can run to get it to work.

Samuraidog wrote:

Thank you OP. This is a fantastic thread. I'll check out Neat Image.

Have you had any luck getting X-H2S raw images to work with Darktable? I am unable to open any raws using Darktable. I'm wondering if there is some kind of patch I can run to get it to work.

You're welcome, thank you for your kind words.

Haha I don't own the X-H2S But seems to be a wow camera when it comes to AF speed and tracking.

I was a late adopter with the X-T4. Nevertheless I had circumstances to get it show up in the corrections. But this was a Lensfun topic, not a darktable issue.

darktable // How to make the X-T4 show up in the lens correction profile

Regards,

Martin

Rightsaidfred wrote:

What Neat Image made from the darktable export. Very quick procedure.

I always thought Neat Image did a great job (still does) but, for me, Topaz DeNoise (with a little fiddling) does a significantly better job of retaining and enhancing fine detail. I'm a Windows guy, so issues for me on that count...

Your ISO 12800 "before" image with Topaz DeNoise

*Edit: I couldn't help warming Pedro up a bit..

Kameratrollet wrote:

Samuraidog wrote:

Thank you OP. This is a fantastic thread. I'll check out Neat Image.

Have you had any luck getting X-H2S raw images to work with Darktable? I am unable to open any raws using Darktable. I'm wondering if there is some kind of patch I can run to get it to work.

X-H1 support

You will need to add X-H2S support. It is like copy/paste. You can get support already today but start by reading this https://github.com/darktable-org/darktable/wiki/Camera-support

That worked! Thank you.

Kaoticphoto wrote:

Give a try to dxo pure raw 2

I use deep prime as well and have been fairly happy with the high iso NR.

Note the OP did mention running on linux os. DxO (PR/PL) is windows/mac.

Erik Baumgartner wrote:

Rightsaidfred wrote:

I always thought Neat Image did a great job (still does) but, for me, Topaz DeNoise (with a little fiddling) does a significantly better job of retaining and enhancing fine detail. I'm a Windows guy, so issues for me on that count...

*Edit: I couldn't help warming Pedro up a bit..

Not bad Erik.

Thanks for warming up Pablo. I like the 'retracted' colors. I love Classic Chrome

So we need a direct comparison. Sitting in front of a Windows machine.

• Same tiff source file (darktable export).
• Standard settings, more or less.
• Effort about 1 minute each.

12800 ISO. Screenshot from a Windows machine. Left: Topaz DeNoise AI test version v3.7.0 standard settings. Right: Neat Image v9.0.2 standard settings. Both processed from exactly the same tiff exported from darktable. Topaz provides a smoother image. Neat Image maybe a little more vivid an image.

12800 ISO. At 109 % we see a clearly better result of Topaz (left) vs Neat Image (right). I do not know whether this shows the full performance each. But if it needs to be quick, in this case, yes, Topaz is the winner.

The respective section in the user interface of Topaz AI. Left: original (tiff), right: preview. "Severe noise" was auto detected from Topaz. I did not change the settings.

12800 ISO. Topaz DeNoise AI v3.7.0 test version. Standard/auto settings.

12800 ISO. Neat Image v9.0.2. Standard settings, sharpen, mid details slightly enhanced. Same image as in parent thread. Second copy for convenience of the readers.

So indeed, Topaz produces an excellent result.

Regards,

Martin

I'm sure it could be attributable to user error, but I downloaded the demo version of Neat Image and didn't like it so much. I think it is great for smoothing out the grain, but I feel it degraded the sharpness too much. I even had the sharpness setting to high. Maybe my expectations were too high. I've seen what Topaz can do.