Home

Welcome to the darktable-ai wiki – the home of darktable's AI model conversion and packaging pipeline.

If you're here as a darktable user wanting to know which model to install or how to update them, head straight to the Users portal.

If you're here to contribute – adding a new model, fixing the pipeline, helping cut releases – start with the Developers portal.

What this repo actually is

darktable-ai is a conversion pipeline. It takes published AI model checkpoints (PyTorch, etc.), converts them to ONNX with the input shapes and metadata darktable expects, packages each one as a .dtmodel archive, and publishes them as GitHub releases. darktable downloads those archives from inside its preferences UI – preferences → AI.

The whole thing is driven by a small Python CLI called dtai. One model = one directory under models/, with a model.yaml describing it and per-model convert.py / demo.py scripts. CI runs the same commands on every model whenever master moves.

Quick links

• Getting started – install, clone, first pipeline run
• Users portal – picking, installing, updating models
• Developers portal – contributing
• Model catalog – what we ship today
• AI model policy – which models we accept and why
• Adding a new model – step-by-step
• Pipeline reference – model.yaml, CLI, schemas
• Release cycle – nightly vs release, tag conventions
• Creating a release – maintainer checklist
• Roadmap

Reporting bugs

Two repos, two trackers – pick the one closest to where the problem actually lives:

• The model misbehaves inside darktable (bad mask, denoise artefacts, won't load, "update available" never clears, GPU provider issues) → darktable issues. That's where the AI-tab UI, model loading, and inference pipeline live.
• The pipeline itself fails (conversion errors, packaging issues, CI failures, wrong versions.json) → darktable-ai issues. That's this repo.

When you open an issue, please include:

1. darktable version (and whether it's a release or nightly build)
2. the model id and version you're using
3. your platform – OS, CPU, GPU, execution provider in preferences → AI
4. exact steps to reproduce
5. one sample image that triggers the problem, if possible
6. for crashes: a backtrace (see darktable bug reporting for how to produce one)

Logs from darktable -d ai are extremely useful for AI-related issues.