Darktable 5.6.1 Ships as the First Bug-Fix for the Stable 5.6 Series
Darktable has published version 5.6.1, the first maintenance release for its stable 5.6 photography workflow. The team shipped this update purely to patch stability issues and security flaws discovered since 5.6.0 dropped. No new features. No new toys. Just cleaner code.
The release lands right on the biannual schedule the project follows. You can spot the release-5.6.1 tag at commit 03179f8 in the repository. Since the last stable build, the volunteer community merged roughly 629 commits to master, with 150 targeting the darktable codebase and the bundled rawspeed RAW decoder. That is a solid haul for a community project.
What Actually Changed
The release notes mention a reverted performance tweak. Someone tried to bump preview pipe dimensions for speed, but it caused rendering weirdness, so they rolled it back. The AI features got a practical win: model downloads now resume after network hiccups instead of restarting from scratch. Since those ONNX inference models are huge, that is a real quality-of-life win. Neural raw denoise also refuses to touch Fujifilm Super CCD files now, since the AI models were never trained on them.
Workflow metadata finally survives AI denoise and upscale operations. Color labels, ratings, and manual geotags are copied to newly generated images, which actually preserves user intent through the AI pipeline.
Stability and Platform Notes
The bug list is heavy on security. There is an out-of-bounds read in wavelet decomposition for previews, plus out-of-bounds writes in path masks when two nodes overlap. Memory management got cleaned up too, with fixes for CPU highlights out-of-memory crashes and over-allocations on small OpenCL devices.
AMD GPU owners on Windows finally see their OpenCL devices stay enabled, though you will still need to jump through a few hoops for full acceleration. The ONNX Runtime auto-detect crash is gone, and Wayland AppImages will now default to native sessions instead of falling back to X11.
Camera tethering finally connects when your UI language is not English. Sony tethered exports stop stripping embedded lens-correction tags, which was a persistent annoyance. JPEG 2000 files lose that weird every-fourth-pixel distortion, and 16-bit PNG exports actually render the blue channel correctly now.
If you are using Apple ProRAW or computational raws, the AI neural restore module stops crashing on LINEAR-class images. sRaw inputs are now routed through the LINEAR pipeline instead of being rejected. Keep in mind that several compression formats remain unsupported. Apple ProRAW DNGs, CinemaDNG lossless/lossy, DNG 1.7 with JPEG XL, Fuji lossy RAFs, and a handful of Phase One and Sony variants simply will not open. Camera support for older Leaf, Olympus, and Phase One models is suspended pending raw samples on raw.pixls.us. You can help restore those by contributing files under a CC0 license.
Platform requirements have not shifted since 5.4. The minimum hardware spec is 8 GB RAM, a quad-core processor, and an Nvidia GPU with 1,024 CUDA cores and 4 GB VRAM. The project still leans on Nvidia for reliability, though this build does fix several AMD rendering issues.
AI features remain opt-in. Build with -DUSE_AI=ON to enable them. CPU inference works out of the box. Apple Silicon gets CoreML and the Neural Engine. Windows supports DirectML. Linux users need CUDA 12/13 with cuDNN 9.x for Nvidia, ROCm 6+ with MIGraphX for AMD, or OpenVINO for Intel. You will need 4 GB of VRAM minimum for GPU inference. If it fails, darktable drops back to CPU automatically.
The documentation for 5.6 is not complete yet, and exporting to AVIF, EXR, JPEG XL, or XCF requires all metadata checkboxes to be ticked together. The project is CII Best Practices certified and runs on GNU GPL-3.0.
Head here to the darktable website to grab the latest release. Verify the hashes before installing, especially if you are building from the darktable-5.6.1.tar.xz source tarball. The next stable release will likely arrive in January, assuming the 5.5 development branch does not uncover more regressions.
