Skip to main content

Local AI Model Management

This page covers the local vision models in Settings → Local Content Recognition. Once downloaded and activated, a model can power import-time content label generation, label filtering, and semantic search; inference runs on your machine.

To configure an API provider for photo descriptions, ratings, and rating reasons, see API provider configuration instead.

Opening local AI settings​

Click Settings on the toolbar and choose Local Content Recognition on the left; or, when the Set up an AI model to analyze images prompt appears during import, click Set up model and the app jumps straight to this section.

The page contains three main sections:

  • AI content recognition: label statistics for the current album, and the generate / regenerate flows.
  • Model: choose, download, and activate a local model.
  • Import: whether to auto-generate labels when importing new images.

Choosing a model​

Pick a model from the dropdown at the top of the Model section. The models in the list differ in size, accuracy, and language support; after selection, the status card shows the display name and total size. If you are unsure, start with the default item — you can switch and re-download later.

Setting the download directory​

Download directory shows where the model files are stored. Click the folder icon on the right and choose a directory in Select Model Download Folder. Model files are large, so pick a location with plenty of space that you can keep long-term, and do not delete the files in it casually.

Choosing a mirror source​

The Source dropdown offers four options:

  • HF Mirror: a Hugging Face mirror for networks with poor access to the official site; prefer it when the official site is slow or unstable.
  • Hugging Face: the official source; use it with a good connection — the most complete and up-to-date files.
  • Sufy CDN: a CDN-distributed mirror, useful as a backup to the first two.
  • Custom: a custom download address. Selecting it shows an input field for the full download root address, e.g. https://example.com.

Recommendations: use Hugging Face when you can reach the official source at a decent speed; use HF Mirror when access from China is difficult; if both are slow, try Sufy CDN or a custom address. When a download fails or stalls, switching the source and re-running Download is usually the most direct fix.

Downloading and activating​

With model, directory, and source set:

  1. Click Check to refresh the local status of the current model; a not-yet-downloaded model shows Not downloaded.
  2. Click Download to start. The status card shows a percentage, downloaded / total bytes, the current file name, download speed, and estimated time remaining; the button becomes Cancel while downloading.
  3. When the download finishes, the status card shows Activate to generate labels — the files are in place but not loaded into the runtime.
  4. Click Activate. When done, the status card shows a green Active; label generation and semantic search are now available.

Model status card​

The status card is the primary source of truth for model state:

  • Red / Not downloaded: no complete model files locally.
  • Accent color / Downloading %1% or Activating…: download or activation in progress.
  • Gray / Activate to generate labels: files downloaded but not activated.
  • Green / Active: the local model is ready.

When idle, the card also shows controller-reported information such as errors, install paths, or current runtime state.

Other operations​

  • Check: re-reads the local status of the selected model.
  • Download / Cancel: starts or interrupts the download.
  • Delete: removes the local files for this model; the status returns to Not downloaded.
  • Activate: loads a downloaded model.

Import-time label preference​

The label generation options in the Import section control what happens when importing new images:

  • Always: generate on every import without asking.
  • Always Ask: show the Use AI to analyze image content? dialog on every import.
  • Always Skip: never generate labels at import and never show the model setup prompt.

See AI label generation for the import dialogs and the manual generation flow.

Troubleshooting​

  • Download fails or stalls with no progress: switch the Source and re-run Download; for a custom source, confirm the address is reachable and the directory structure matches the official one.
  • Download completes but Activate fails: check that the directory is intact and disk space is sufficient; Delete and re-download if needed.
  • Status shows Active but label generation is still unavailable: confirm the current project is open, and click Check in the local AI section to refresh the status.
  • Semantic search still unavailable: confirm the local model is Active; an API provider profile does not replace the local semantic search model.