-
DarkRoom 0.22.0
StableBenchmarks / CPU and I/O (per commit) (push) Successful in 8m36sBenchmarks / Frame budget (on demand) (push) SkippedTraceability / Requirement traces (push) Successful in 1m31sBuild and test / Android (aarch64) (push) Successful in 48m34sBuild and test / android-image (push) Successful in 4s🐳 Android image / Build and push (push) Successful in 3sBuild and test / Desktop (Linux) (push) Successful in 1h32m33sBuild and test / windows-image (push) Successful in 2s🐳 Windows image / Build and push (push) Successful in 2sBuild and test / Layer separation (push) Successful in 29sBuild and test / Windows (x86_64, cross) (push) Successful in 31m14sBuild and test / Publish the release (push) Successful in 1m9sreleased this
2026-10-04 12:13:39 +00:00 | 25 commits to master since this releaseCorrection: in 0.22.0 the tablet's NPU is not used. The app could not load the NPU's driver library on Android, so every model, AI denoise included, ran on the tablet's processor (Best takes 30 seconds to two minutes there). 0.22.1 fixes this; the NPU timings below apply from 0.22.1 on.
AI denoise becomes how every raw is developed, with three networks to choose from and its result kept on disk; the tablet runs it on its NPU; colours are richer by default; and presets follow you between devices.
AI denoise is how raws are developed. It now heads the Adjust panel, and every Bayer raw develops through it unless you choose otherwise. A Method control replaces the on/off switch: Best (the default), Medium, Fast, or Bilinear for the ordinary conversion. On a laptop GPU a 20 MP photograph takes about 2.5 s with Best, 0.8 s with Medium and 0.6 s with Fast, plus about half a second to read the file. The ordinary conversion shows while it works. Strength eases it off, putting some of the removed noise back as fine grain. The networks are new: two specialists blended per pixel for Best, trained on several times more photographs, with sharper edges and less false colour, and stuck photosites are repaired before the network sees them. Results are kept on disk, so a photograph reopened or exported does not wait again. The first launch prepares each network for the graphics card in the background, which takes a few minutes.
On the tablet, AI denoise runs on the NPU. A 20 MP photograph takes about 9 seconds with Best, the default, about 2 with Medium and about 1 with Fast — against more than half a minute on the processor — with no measurable loss of quality, confirmed on the device. The other models that run on the tablet — face landmarks, segmentation, scene recognition, the panorama border fill and keypoints — move to the NPU too, and face detection finds faces it missed before at small sizes. The APK is about 85 MB larger, for the new networks and the NPU's forms of the models.
Richer colour by default. The camera profile's look table is now off by default; its strength is still under Camera Profile. The colour mixer's saturation now acts on muted colours such as skies and shadowed snow, where it used to do almost nothing. The Vivid presets and Punch are retuned to measured amounts of colour on the new rendering.
Earlier edits come across more faithfully. A photograph opened with the edit its file already carries now brings its tone sliders and per-colour saturation at measured strengths, with each saturation band spread over the colours it really affects.
Presets sync. Your develop presets travel with each library to your other devices. Saving a preset now works on Android and Windows, where it failed before.
Sensor health groundwork. The app can now find stuck and hot photosites without repairing them, the first step to tracking a camera's sensor over time. Nothing in the app shows it yet.
Every raw looks different from 0.21.0 — the denoise network, the look table and the colour mixer all changed. Previews made before refresh when a photograph is rendered again. Update tablet and desktop together: the catalog's schema is unchanged, but two versions render the same raw differently. An older build reading an edit saved by this one ignores the new Method setting and uses its own network.
The APK, the desktop binary, the Windows installer and SHA256SUMS are built by CI from this tag's commit.
Downloads