-
Notifications
You must be signed in to change notification settings - Fork 14
FAQ
LrGeniusAI is an AI extension for Adobe Lightroom Classic. It adds AI-powered metadata generation (keywords, titles, captions), semantic free-text search, automatic develop edits, image culling, face/person management, and more — all running locally as a background server without freezing Lightroom.
Only if you choose a cloud provider (ChatGPT/OpenAI, Google Gemini; Vertex AI was removed in August 2026). With a local provider — the built-in llama.cpp and MLX engines, or an external Ollama / LM Studio server — your photos never leave your machine. For cloud providers, images are sent to the respective API for analysis. Embeddings and all generated metadata are always stored locally.
Adobe Lightroom Classic only. Lightroom CC (cloud) and other Lightroom versions are not supported, as the plugin relies on the Lightroom Classic SDK.
Yes, LrGeniusAI is open source (AGPL-3.0). Cloud API usage (Gemini, OpenAI) incurs costs at the respective provider's standard rates. Local models are completely free to run — including the ones the backend downloads and runs itself, see Local AI Models.
Yes. LrGeniusTagAI is our earlier, lightweight Lightroom Classic plugin that focuses solely on AI-generated tags and descriptions, without the full server, semantic search, culling, or editing features of LrGeniusAI. It's a good fit if you just want quick AI keywording with a simpler setup.
Yes. The installers are not code-signed (cost and complexity for an open-source project), which causes operating system warnings. This is expected behavior:
- Windows (SmartScreen): Click More info → Run anyway.
-
macOS (Gatekeeper): Right-click the
.pkg→ Open → Open anyway. Or go to System Settings → Privacy & Security → Open Anyway.
- Check that you completed the installer for the backend (separate from the plugin
.lrdevplugin). - On macOS, make sure you allowed the binary in System Settings → Privacy & Security.
- Try starting the server manually:
lrgenius-server/lrgenius-server(macOS) orlrgenius-server/lrgenius-server.cmd(Windows). - Check the Plugin Manager → Backend Server section — it shows the server status and URL.
Yes. Change the Backend Server URL in Plug-in Manager → LrGeniusAI to point to the remote host, e.g. http://192.168.1.100:19819. Docker Compose files are included in the repository.
See Help: Choosing AI Model for a full breakdown. Quick summary:
| Goal | Recommendation |
|---|---|
| Cheap bulk keywording |
gemini-2.5-flash-lite or gpt-5-nano
|
| Balanced default | gemini-2.5-flash |
| Best quality |
gemini-2.5-pro or gpt-5.4-pro
|
| Privacy / no API cost | Built-in llamacpp with Gemma 4 E4B |
| Apple Silicon, local | Built-in mlx with Gemma 4 E4B |
Yes. The backend has two built-in engines — llama.cpp (GGUF; macOS, Windows, Linux) and MLX (Apple silicon). In Plug-in Manager → LrGeniusAI find the Local AI Model sections, pick a model, and click Download. Nothing else to install or keep running. See Local AI Models.
The model list is loaded from the backend at runtime. If it's empty:
- Make sure the backend server is running and reachable (Plugin Manager → Status).
- Check that the relevant API key or local server URL is configured.
- For Ollama/LM Studio: the local server must be started before Lightroom is opened (or before running a task).
- For the built-in local engines: a model must be downloaded first — the Installed line in the Plug-in Manager tells you whether one is present.
MLX needs the lrgenius-mlx helper next to the server binary. It ships in the official macOS
installer, so a greyed-out section usually means a source build without the xcodebuild step —
the status line names the exact reason. On Windows the section is llama.cpp rather than MLX.
File → Plug-in Manager → LrGeniusAI → scroll to the API Keys section. Enter your Gemini or OpenAI key there.
Up to four things in one pass:
- Sends each photo to the configured AI model to generate keywords, title, caption, and alt text.
- Creates a semantic embedding (using SigLIP2 locally) so the photo can be found by Advanced Search.
- Detects and embeds faces for the People workflows.
- Identifies animal, plant and fungus species (using BioCLIP 2 locally) and writes the taxonomy to the plugin's metadata fields.
Every step is optional and each has its own checkbox — only step 1 involves a language model or a cloud account at all.
No. It runs BioCLIP 2 on the machine running the backend, the same way search embeddings and face detection do. Nothing is uploaded, and it works with no API key and no internet connection once the model is downloaded. See Help: Analyze and Index.
Because that is the deepest rank the model was confident about. A clear frame of a garden bird gets a binomial; a distant silhouette gets an order or a class. The rank is written into its own metadata field so you can always tell which you got. Full explanation, including why some common names come back in Swedish, is in Help: Analyze and Index.
Yes. Advanced Search only works on photos that have been indexed (embeddings created). Unindexed photos will not appear in search results.
- Try a more capable model (e.g. move from a small local model to
gemini-2.5-flash). - With a local model, turn off keyword aliases and bilingual keywords — both make the model emit structured keyword objects, which small models handle badly (often returning no keywords at all).
- Add Photo Context (folder names, capture date, GPS coordinates) to give the AI more information.
- Write a custom System Prompt in Plug-in Manager → Prompts to guide the output style.
- Adjust the Temperature slider — lower values produce more consistent output.
Enable Regenerate all data (overwrite existing) in the Analyze & Index dialog.
- Make sure photos were indexed with Create search embeddings enabled.
- Set the Lightroom collection sort order to Custom Order — otherwise results appear in random order and the best matches are not at the top.
- Try a broader query — semantic search understands concepts, not just exact keywords.
Results are ranked by combining visual semantic embeddings with a text search over AI-generated metadata (keywords, caption, title, alt text). The final score reflects both visual and textual similarity to your query.
A structured Lightroom develop recipe of global adjustments (exposure, white balance, tone curve, contrast, presence, sharpening, and so on), built by matching the photo against your own saved edits. The recipe is applied via the Lightroom SDK — no raw pixel editing happens outside Lightroom.
None. AI Edit does not call a language model at all — it interpolates the develop settings of the training examples closest to the photo. Model choice only affects Analyze & Index.
It needs at least five saved training examples before it can produce anything. Edit some photos the way you like them and run Save Edits as AI Training Examples (Library → Plug-in Extras). See Help: Train from Edits.
Enable Review each proposed edit before applying it in the AI Edit dialog. You will see the proposed develop values, the confidence of the style match, and any guardrail explanations, and can apply or skip each photo. (There is no rendered before/after preview yet — use Lightroom's History panel to judge the result.)
Enable Apply the edit to a new virtual copy in the AI Edit dialog. Each edited photo gets a virtual copy named AI Edit and the recipe lands there.
No. Culling is completely non-destructive. It creates Lightroom collections (Picks, Alternates, Reject Candidates, optional Duplicates) — your photos are never moved or deleted.
Photos need to be indexed with Analyze & Index Photos first. Face-aware ranking (eye openness, sharpness) requires face detection to have been run during indexing.
It lists all detected face clusters (persons) from your indexed photos, lets you assign names, and creates Lightroom collections per person. See Help: People & Faces.
In the Analyze & Index Photos dialog, make sure face detection is enabled. The backend uses YuNet for detection and FaceNet for the embeddings it clusters.
-
Auto De-Clutter runs automatically during indexing and prevents the AI from creating near-duplicate keywords of ones already in your catalog (e.g. if
Carexists,AutomobilebecomesCar). - Deduplicate Keyword Synonyms is an interactive workflow you run manually to clean up synonym sprawl that already exists in your catalog.
See Help: Keyword Deduplication and De-Clutter.
- In Lightroom: after a batch task finishes, a Task Completion Dialog shows per-photo errors.
- In Plugin Manager: click Show logfile or Copy logs to desktop.
- On the server: check the terminal window where
geniusai-serveris running, or the log files in the server's working directory.
The backend is not reachable:
- Check Plug-in Manager → Backend Server — the status indicator shows whether the server responded.
- Verify the server URL (default
http://127.0.0.1:19819). - Restart the backend manually if it crashed.
For more detailed solutions see Troubleshooting.
Versions before file-based photo_id values stored Lightroom catalog UUIDs as primary IDs,
and the backend cannot match those against your photos any more.
There is no migration — the one the plugin used to offer never worked, and has been removed. Run Analyze & Index Photos over the catalog again instead. Photos already indexed under the current IDs are skipped, so only the ones that genuinely need it are re-processed.