Model catalog

The model lineup changes fast. Rather than duplicate it here (and let it go stale), this page explains the structure – what tasks exist, how they map to darktable, and what each model card contains. For the current list of models, see the README in the main repo:

→ Model list in darktable-ai/README.md

Each row in that table links to the model's own card under models/<model-id>/README.md.

Tasks

A task is a darktable feature slot. One model can be active per task; multiple models for the same task can be installed and switched between.

Task	Where it shows up in darktable
mask	darkroom → mask manager → AI object tool
denoise	neural restore module → denoise
rawdenoise	neural restore module → raw denoise
upscale	neural restore module → upscale
embed	tagging / similarity search
depth, erase	reserved for upcoming features

If you want to add a new task, that's a coordinated change across this repo and darktable itself — open an issue first.

Model types

The type field in model.yaml controls how the model is packaged and how demo() is called:

Type	Description
single	One ONNX file. demo(model, image, output, **kwargs).
split	Encoder + decoder pair (e.g. SAM 2.1). demo(encoder, decoder, image, output, **kwargs).
multi	Several ONNX files in one model directory (e.g. BSRGAN 2× + 4×). demo(model_dir, image, output, **kwargs).

What a model card contains

Every model has a README.md under models/<model-id>/. The card includes:

• Source – upstream repo URL, paper, license
• Architecture – brief description
• ONNX I/O – input/output tensor names, shapes, dtypes, normalization, tile support
• Selection criteria – table with every row from AI model policy

When reviewing or adding a model, the selection-criteria table is the part that matters most — it's how the project verifies the model meets policy.

Picking a model as a user

If you just want to install models and use them, head to the Users portal. The catalog link at the top of this page is the place to browse what's available.