Meet the new AI tools in darktable 5.6 | darktable
Andrii Ryzhkov
Meet the new AI tools in darktable 5.6
posted on<br>Tue 23 June 2026
by<br>Andrii Ryzhkov
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.
To follow along, you can grab the five sample raws used in the walkthroughs:
Download “a Cat named Ebby_ND7_5751.NEF” (24 MB) – used for the AI object mask walkthrough.<br>“A cat named Ebby” © 2020 by Suki2019, licensed under CC BY-SA 4.0.
Download “DSC01318.ARW” (20 MB) – used for the multi-object mask example.<br>“A whole lot of penguins!” © 2025 by Klogg, licensed under CC BY-SA 4.0.
Download “_DSC0488.NEF” (15 MB) – used for the title image.<br>“Chrysanthemum Flowers” © 2019 by Xavier Bartol (XavAL), licensed under CC BY-SA 4.0.
Download “07 Monkey ISO 6400.CR3” (43 MB) – used for the neural denoise walkthrough.<br>“A Raw Denoise Cross-comparison” © 2025 by Dave22152, licensed under CC BY-SA 4.0.
Download “20250629_0012.NEF” (13 MB) – used for the upscale walkthrough.<br>“Hoverfly on flower” © 2025 by Fenny, licensed under CC BY-SA 4.0.
Thanks to Suki, Klogg, Xavier, Dave, and Fenny for sharing these.
Ground rules
A few principles we set early on, that ended up shaping every later decision:
Optional. Built without USE_AI, the whole subsystem disappears. The official binaries include it, but distros can opt out, and so can you – none of the existing modules depend on it.
Local. Nothing leaves your machine. No cloud inference, no telemetry, no “we’ll improve the model with your data”. Every model runs on your CPU, or your GPU if you have one. As a photographer myself, I’d never paste a raw file into a black-box service, and I wasn’t going to ask anyone else to.
Curated open models. Every model we offer is documented in a separate repository – darktable-ai – with its source code, training data and license. Binary weights are not bundled inside darktable itself, and models whose provenance isn’t clear don’t get included. That bar means our catalog is smaller than commercial competitors. That’s the trade we chose.
Extensible. The catalog is a starting point, not a fence. Each AI task in darktable (object masking, denoise, upscale, …) has a defined ONNX interface – input shapes, output shapes, expected pre- and post-processing. Any model matching that interface can be installed manually and used like a first-party one. If you know how to convert a published model to ONNX, or train your own, you can bring it to darktable without waiting for us to ship it. The catalog is for everyone; the door is open if you want to go further.
Off by default. Nothing is downloaded until a feature you enable needs it.
Turning it on
AI features stay off until you turn them on – nothing is installed or loaded unless you ask for it. Open preferences → AI and tick enable AI features. The easiest start is the download / update default action: one click pulls down the recommended set so every feature works immediately. Once installed, models live on disk and load on demand.
If you have a GPU, wire up a matching execution provider – CoreML on macOS, CUDA on NVIDIA, ROCm on AMD, DirectML on Windows, OpenVINO for Intel. Setup varies per platform; the GPU acceleration page in the manual walks through each one. CPU works too, just slower.
The AI object mask
This is the feature that’s changed my workflow the most. Drawn masks in darktable have always been powerful – brush, path, gradient, ellipse, parametric intersection – but masking, say, an animal against a busy background has always meant either patience with the brush or settling for an approximation.
The AI object mask short-cuts that. A click on an object gives you an accurate, hard-edged selection of that object, converted on the fly to a regular vector path – the same kind drawn masks produce, with edge nodes you can move, feather you can adjust, and intersections you can combine with other masks.
Walkthrough: masking Ebby
Open a Cat named Ebby_ND7_5751.NEF in the darkroom. Pick any module that supports masking – I’ll use color balance rgb for the demo. Open its mask manager and pick the AI object mask tool.
The first click on a new image triggers the encoder pass – 1 to 10 seconds depending on hardware. Once it’s done, the model is ready to answer click prompts interactively for as long as you stay on this image.
Click somewhere near the centre of...