release 5.6.0

We're pleased to announce the first release of the darktable AI models, 5.6.0!

The github release is here: https://github.com/darktable-org/darktable-ai/releases/tag/release-5.6.0.

These models are intended for darktable 5.6.0 and later. They power the AI features introduced in this darktable release – interactive object masking and neural restore (raw denoise, image denoise, upscale). Models are installed and managed from the AI tab in darktable's preferences.

Models

Object masking – used in the darkroom mask manager's AI object tool

• mask sam2.1 hiera small (mask-object-sam21-small) – Segment Anything 2.1 (Hiera Small). Click on an object to generate a precise mask; click again with foreground/background prompt points to refine. Encoder + decoder pair: the encoder runs once per image (with optional GPU acceleration), the lightweight decoder produces masks interactively. Default choice for most users.

• mask sam2.1 hiera tiny / base plus (mask-object-sam21-tiny, mask-object-sam21-base-plus) – lighter and heavier variants of the same SAM 2.1 pipeline. Use tiny on low-memory systems or for faster encoder runs; use base plus when small doesn't capture an object cleanly.

• mask segnext vitb-sax2 hq (mask-object-segnext-b2hq) – SegNext ViT-B SAx2 HQ fine-tuned for interactive masking. Alternative to SAM 2.1 with openly documented training data. Also useful when SAM 2.1's segmentation behaviour doesn't fit a particular scene.

Neural restore – used in the neural restore module in the lighttable/darkroom sidebar

• denoise nind (denoise-nind) – UNet denoiser from the NIND (Natural Image Noise Dataset) project, trained on Wikimedia Commons noisy/clean pairs. Drives the module's denoise task on demosaiced RGB.

• denoise nafnet small (denoise-nafnet) – lightweight NAFNet denoiser trained on the SIDD smartphone dataset. Alternative denoise task model – tuned for noise patterns typical of small-sensor cameras.

• raw denoise nind (rawdenoise-nind) – UtNet2 raw-domain denoiser trained on RawNIND. Bundles two variants in one package: a Bayer model that denoises and demosaics in one step (pre-demosaic input), and a linear Rec.2020 model used for non-Bayer sensors (X-Trans, Foveon, 4-colour CFAs). Drives the module's raw denoise task.

• upscale realplksr (upscale-realplksr) – RealPLKSR (Partial Large Kernel CNN) at 2× and 4×. Faithful upscaling that preserves detail without smoothing or inventing texture. Best for clean sources like edited RAW exports or high-quality scans. Drives the module's upscale task.

• upscale bsrgan (upscale-bsrgan) – BSRGAN blind super-resolution at 2× and 4×. Combines upscaling with implicit denoising/deblurring, useful when the source already has visible noise or compression artefacts. Drives the module's upscale task.

Compatibility

• Darktable version: 5.6.0.
• All models are statically-shaped ONNX with tile sizes declared in the manifest. Tiles are sized for both CPU inference and the GPU execution providers supported by darktable (CUDA, ROCm/MIGraphX, DirectML, OpenVINO, CoreML).