AI Tags
Clipthesis can look at every video and photo in your library and suggest tags for what is in the frame: the subject, the setting, the kind of shot, what is happening, the light and the mood. A vision model runs on your Mac to do it. Frames never leave your machine. Automatic tagging is off until you turn it on.
Suggestions stay apart from the tags you write yourself. They show as dashed violet chips with a sparkle, and they only become real tags when you accept them.
Platform and hardware
AI tags run on macOS 13.3 or later with at least 8 GB of memory. On Apple Silicon the model runs on the GPU (Metal) and tags a clip in a few seconds. On Intel Macs it runs on the CPU, which takes 20 to 60 seconds per clip, so you may prefer Tag now on individual clips over automatic tagging. The model is a one-time 3.2 GB download. Windows support is not in this version.
Getting started
1. Turn it on
Automatic tagging is off by default. Once your library holds some media, an AI tags chip appears at the right of the bottom bar. Click it to read what the feature does, then choose Turn on. Not now hides the chip for good.
Either way, the switch lives in Settings → Library → AI Tags as Automatically tag new imports, and Tag now on any clip runs the model for that clip alone without turning automatic tagging on.
Turning it on queues every video and photo already in your library once, and downloads the model if you have not done so yet.
2. Download the model
Go to Settings → Library. The AI Tags section shows the model, Qwen3.5 4B (vision), with a Download button. If you skip this step, the first tagging job downloads the model for you and the bottom bar shows Downloading AI model 42% while it does.
3. Let the queue run
Tagging runs in the background. New videos and photos are queued as they are imported or indexed. The bottom-bar chip reads Tagging 12 of 480 while it works, and the queue panel in Settings → Library lists what is running and what is waiting.
Tagging waits while a transcription or an index scan is running, so your Mac is never running Whisper, ffmpeg and the vision model at the same time.
4. Review the suggestions
Open a clip. The AI tags section sits under Tags in the right sidebar and shows one of these states:
- Not tagged, with Tag now
- Queued or Tagging…, with Cancel
- Tagged, with Regenerate
- Failed, with Retry (hover the status for the error)
- Skipped when neither the file nor a thumbnail can be reached
- AI tagging off while automatic tagging is off, with a link to turn it on in Settings
Hover a chip to see its category and how confident the model was. Accepted chips leave the AI section and appear under Tags like any tag you typed.
How AI tags differ from your tags
| Your tags | AI suggestions | |
|---|---|---|
| Where they live | Tags page, groups, colours | A separate list per clip |
| Look | Solid pill in the group colour | Dashed violet chip with a sparkle |
| Cloud Sync | Synced | Stay on this Mac |
| Autocomplete and Tags page | Included | Not included |
| Required groups | Count towards them | Do not count until accepted |
| Resolve keywords and exports | Included | Only once accepted |
Accepting and dismissing
Accept (+) creates a normal tag with that name on the clip. If a tag with that name already exists it is reused. Clipthesis remembers the suggestion as accepted, so Regenerate never proposes it again for that clip.
Dismiss (×) hides the suggestion on that clip for good. Regenerate does not bring it back. To undo an accepted suggestion, remove the tag from the clip as you would any other tag.
Labels are lowercase, at most three words, and never contain a colon, so accepting one never creates a tag group by accident.
Searching and filtering
- The search bar matches AI labels along with filenames, tags and transcripts. Type
lighthouseand press Enter to find every clip the model labelled that way. - The AI tags section in the filter sidebar lists the most common labels with counts. Tick several to narrow the grid to clips carrying all of them. Active AI filters show as dashed chips above the grid, and saved presets remember them.
- Cards with pending suggestions show a violet sparkle badge. Hover the tag icon on a card to see the first few labels.
See Search & Filters for how these combine with the other filters.
Turning it on and off
Settings → Library → AI Tags holds the Automatically tag new imports switch. It starts off. Switching it on queues your existing library once and tags new imports from then on. Switching it off cancels everything still queued (the job in progress finishes), and new imports are no longer queued. Tag now and Regenerate keep working per clip either way, and Tag existing library still queues the backlog when you ask for it.
Privacy and offline
The model runs inside Clipthesis through llama.cpp, bound to 127.0.0.1 on your Mac. For a video, four frames are laid out as a 2×2 contact sheet in a scratch folder inside the app's data directory and deleted after the job. Photos use their existing thumbnail. The only network access is the one-time model download from Hugging Face. The model unloads after 90 seconds without work, so the roughly 4 GB it uses while tagging is free the rest of the time.
Troubleshooting
"The AI runtime is missing from this install" means the bundled llama.cpp files were not found. Reinstall Clipthesis. In a development checkout, run npm run download:llama.
A job is stuck on "Queued" while the bottom bar shows a download: the model is still coming down. If nothing is downloading, quit and relaunch Clipthesis. Jobs that were running when the app closed are reset and picked up again.
Suggestions are generic or wrong. Click Regenerate. Fast cuts, very dark footage and heavy motion blur give weaker results. Dismiss what is wrong so it never comes back.
"The AI runtime failed to start 3 times in a row" pauses tagging for five minutes, then tries again (or sooner if you relaunch). The message includes the last lines the runtime printed.
Nothing is being tagged. Automatic tagging is off until you turn it on. Use the AI tags chip in the bottom bar, the switch in Settings → Library → AI Tags, or Tag now on a clip. On Intel Macs CPU inference is slow, so per-clip Tag now may suit you better than leaving it running.