Artificial intelligence¶
All of Darkroom's AI features run on your computer: no connection, no photo uploads, no subscription. They are marked in the interface by a colored border, which moves while the AI is working.
The features¶
| Feature | Where | Model |
|---|---|---|
| Magic Brush | Masks, Remove | MobileSAM: snaps the stroke to the edge of the painted object |
| Subject, Background | Masks | BiRefNet lite: what stands out in the photo, people or the main object |
| Depth | Masks, Lens Blur, Haze and Relight | Depth Anything V2: relative distance of each point, from the foreground to the far distance, refined along the edges of the photo |
| Sky | Masks | Depth Anything V2: the farthest part of the photo, uniform, bright or blue, at the top of the image |
| People | Masks | MediaPipe (Google): faces, then the hair, skin and clothes of each person, and 478 face points for the eyes, eyebrows and mouth |
| AI Erase | Remove | LaMa: fills the erased area with a plausible background |
| AI Denoise | Detail | DRUNet, computed on the graphics card |
| My Style | My Style | DINOv2: compares the content of photos |
Processing time¶
The models are loaded the first time they are used (a few seconds). After that, on a recent computer:
- magic brush: one to two seconds of analysis on the first stroke on a photo, then almost instant;
- depth: one to two seconds, on the first click on Depth or on the lens blur Apply; the blur itself is then recalculated in real time;
- sky: two to three seconds; subject and background: a few seconds (about 4 seconds on a recent computer, more on a modest laptop);
- people: a few seconds, about one second per person;
- AI Erase: from one to a few seconds depending on the size of the area;
- noise reduction: from a few seconds to a few tens of seconds depending on the sensor and the graphics card (about 15 seconds for a 33 MP photo on a recent card).
Results are kept in memory: going back to a photo does not rerun the computation.
Photos exported with AI-generated pixels¶
When an exported photo contains AI-generated pixels (AI Erase in the Remove tool,
including removals picked up from Lightroom, or an AI fill in the Retouch screen), Darkroom
says so in its metadata: the IPTC digital source type "composite with elements produced by an
AI" (DigitalSourceType), in JPEG, PNG and TIFF files, even when the other metadata is not
exported. The European AI Act requires this marking, readable by software and by the sites
that publish the photos. Noise reduction, masks, red eye and censoring create no content:
they are not concerned.
Working without AI¶
View › Turn Off Artificial Intelligence removes every AI feature: no model is loaded, the AI buttons (magic brush, subject, sky, depth, people, AI erase, AI noise reduction, my style) are no longer offered, and the rest of Darkroom works as usual. The choice takes effect the next time Darkroom starts and stays after updates; untick the box to get AI back. Edits made earlier with AI (removals, masks) stay applied to their photos.
The models folder¶
The models are in the models folder next to Darkroom.exe. Without it, Darkroom
works normally; only the AI features are unavailable. To stop using AI, prefer the option
above: an update puts the models folder back.
Their licenses (MIT or Apache 2.0) allow their use in Darkroom: see Licenses.