Meet the new AI tools in darktable 5.6

Software darktable
1

darktable 5.6 ships its first AI features: an AI object mask that turns a single click into a vector mask around any subject, and a neural restore module for ML-based denoise and upscale. Our focus was on building the foundation that loads and runs open computer-vision models locally on your CPU or GPU – so darktable can host AI tasks suitable for photography as the open model landscape evolves. This post is the hands-on tour: what each tool does, how to use it, and where it falls short.

This is a companion discussion topic for the original entry at https://www.darktable.org/2026/06/meet-darktable-5.6-ai-tools/
8 Likes
2

Well done !

Thanks a lot, I tried out raw denoise with some nightly shot at Stromboli using standard conditions. Worked right away. Happy to share a sample pair of raw file (arw) and the raw-denoised file (dng).
2026-04-29_042347.ARW (19.6 MB)
2026-04-29_042347_raw-denoise.dng (38.8 MB)

Looks very promising for pictures taken under challenging conditions.

4

It’s certainly taken out the noise; but the vibrancy of the image has also gone (vibrancy in the general meaning of the word)

1 Like
5

Some people say that the masking capabilities are the greatest feature of darktable. I am not worth a damn at masking, so I’m sure the new AI features will be a boon for me.

1 Like
6

What computer resources would be required to use these on Win 11 without waiting and waiting for the module to complete the task? Has someone made an assessment?

7

This post slipped under my radar and I don’t know why. It looks a real excellent way to introduce new features that are added to DT. I will have to explore the post more fully.

1 Like
8

Any reasonable dedicated GPU with at least 4 GB VRAM.

9

I suppose it will be fair to assume that these AI modules will not work on my laptop, which has Iris Graphics of Intel and shares the main memory as GPU RAM . Would that be a correct assessment?

10

these AI modules will not work on my laptop, which has Iris Graphics of Intel and shares the main memory as GPU RAM .

At least at this time all darktable AI stuff works on CPU, too. Don’t expect high speed, though.

1 Like
11

Integrated GPU can still be faster than CPU. For Intel cards, I’d recommend installing ONNX Runtime OpenVINO for better performance and hardware support.

2 Likes
12

Kudos on version 5.6! The new RAW AI denoise tool does an amazing job.
However, on my astrophotography shots, the generated DNG file contains many hot pixels, and it is currently impossible to use the “Hot pixels” module on this file to fix them.
Would it be possible to allow this module to be applied to the generated DNGs, or to include hot pixel removal before the AI processing?

13

Which camera do you use?

Neural raw denoise writes two DNG variants:

    • Bayer CFA path (standard Bayer sensors): DNG carries CFA data. Hot Pixels module should work - it fires before demosaic in the raw pipeline.
    • Linear DNG path (X-Trans, and pre-demosaicked raws that Bayer round-trip can’t handle): DNG carries demosaicked 3-channel linear RGB. Hot Pixels module can’t apply here by design - no CFA to look at.
1 Like
14

I am still unable to run any of the AI modules. The models are failing to download. I am getting this message on my system which is a win11 with Intel IRIS Graphics GPU. This is the screenshot of the error message Please help

image
image1121×477 28.9 KB

15

You can download and install models manually in case you have connectivity issues

16

Besides, I don’t think the denoising times of other commercial apps are that much faster.

17

Can the ai mask be used for the retouch module? I’d like to take advantage of the ai mask to remove an object from a picture but I can’t quite figure out how.

18

The retouch module cannot be used to do inpainting, which IIUC is what you are after. You can use it to heal, clone, fill or blur a patch. For this things you hardly ever need (or want, for that matter) an exact contour mask, so I don’t see how the AI making would help there.

1 Like
19

Since I’ve been using nind-denoise in my own script, I tested out the raw-denoise as it has the advantage of having much more data to denoise from RAW than a demosaiced JPG/TIFF.

I guess it’s still early and improvements are still being made. It’s still not handling X-Trans correctly (sadly I have X-H1, X-T2, and X-T20), LinearRAW DNG pathway still has blotchy blobs with wrong colors.

The strength slider might not be the best approach as it simply leaves uniform chroma noise to be addressed later. I haven’t found an easy way in darktable to achieve the effect of Richardson-Lucy deconvolution, it’s super helpful to bring the most details out of the lens. But it tends to accentuate noise, so a completely clean image is needed, the strength slider need to be at 100% for me.

Compared to my nind-denoise workflow, instead of blending noise uniformly, my script does:

  • RGB-version of nind-denoise then RL-deblur to sharpen up details
  • use darktable’s denoise profiled to create a grainy/luma-noise image (retaining details)
  • diff the grainy against the denoised to determine noise level
  • selectively blend the grainy image in (e.g. more noise in midtone, less noise in shadow and highlight, simulating film grain). Amount of blending can also be adjusted as an S-CURVE approximation, e.g. alpha = 0.65 * log10(1 + (noise / 150)**2)

Anyway, the neural restore module and ONNX approach is a huge achievement in the right direction. I’m excited for upcoming development and improvements.

A few test cases for comparison (identical history stack except demosaic):

Ballroom dancing under blue light - DSCF6882.RAF (16.1 MB)

Creative Commons, By-Attribution, Share-Alike.

For X-Trans, Markesteijn has been the best demosaic, even better than Fuji’s algorithm. The generic LinearRAW DNG pathway might not be making use of it yet.

DSCF6882_raw-denoise
DSCF6882_raw-denoise5958×3901 1.66 MB

raw-denoise: blotchy/wrong colors, grid-pattern, artifacts

DSCF6882
DSCF68825958×3901 3 MB

nind-denoise + RL-deblur, with grain blended in dynamically

Schoena at the cottage - Bayer

TZ5_3917_raw-denoise
TZ5_3917_raw-denoise3230×3891 1.26 MB

raw-denoise: 100% strength, no sharpening, bokeh is real smooth, image has slight green cast (some existing nind-denoise models also have red/green color cast)

TZ5_3917
TZ5_39173230×3891 2.72 MB

nind-denoise + RL-deblur, with grain blended in dynamically

Bratislava castle at night - Bayer

DSC_2998_raw-denoise
DSC_2998_raw-denoise5985×3815 1.4 MB

raw-denoise: 100% strength, no sharpening, blotch/artifact near high-contrast (wall edge, light bulb, …), visible circular steps of gradient near the top-center

DSC_2998
DSC_29985985×3815 3.19 MB

nind-denoise + RL-deblur, with grain blended in dynamically

20

Makes sense I guess even though retouch can remove some objects with a bit of patience. Is there a way to do inpainting in darktable that I might not know about then?