Help: Analyze and Index

This is the task that does the actual AI work: it sends your photos to a model, gets keywords, titles, captions and alt text back, and builds the search index the other features rely on.

Start the task

In Lightroom Classic:

Library → Plug-in Extras → Analyze & Index Photos...

The dialog has three tabs.

General

Scope

Selected photos only
Current view — the folder or collection currently open
All photos in catalog
New or unprocessed photos — only photos the backend has no data for

AI Model

Picks the provider and model, plus Temperature, Max Tokens, output Language, and Replace ß with ss for German output. Which providers appear depends on what is configured in the Plug-in Manager — see Choosing an AI Model.

Primary Tasks

Generate AI metadata (Keywords, Title, Caption) — the LLM pass. Which of the four fields are actually produced is set on the Metadata Options tab.
Enable smart photo search — computes the SigLIP embedding that Advanced Search and Find Similar need. Greyed out until the search model has been downloaded in the Plug-in Manager.
Enable face detection — detects and embeds faces for People.
Identify animal and plant species — runs BioCLIP 2 and writes a taxonomic identification. Greyed out until the models have been downloaded in the Plug-in Manager, under On-device AI models — one Download AI models button fetches everything that is missing, so there is nothing to pick per feature. Two sub-options:
- Only where an animal or plant is detected (on by default) — checks the already-computed search embedding first and skips photos with no organism in them. BioCLIP is a large model and most of a general photo library is portraits, architecture and landscapes, so leaving this on is typically the difference between minutes and hours. Turn it off if you know the selection is all wildlife.
- Also write the taxonomy as keywords (off by default) — see below.

Turning off AI metadata makes the rest of the run faster

Unchecking Generate AI metadata does more than skip the language model. With no model in the loop there is nothing that has to see one photo at a time, so — when Submit original files is on in the Plug-in Manager and the backend runs on this machine — the plugin hands over a group of photos per request and the backend reads, decodes and measures them across several CPU cores. See Performance Tips §5.

That makes “embeddings and faces now, metadata later” a genuinely cheaper way to work through a large import than doing everything in one pass.

If a model is not downloaded yet

Pressing Start checks that the on-device models the selected tasks need are actually on disk, and asks before spending the indexing time if one is not:

Download now — starts the download (it reports its own progress) and stops this run. Start it again once the download finishes and the photos are processed with everything you selected.
Continue without it — indexes anyway. The photos are processed without that model’s contribution, and the completion dialog reports it as a warning naming the Download AI models button. Running Analyze & Index over the same photos later fills in what was skipped.
Cancel — nothing runs.

Only the models the run actually needs are checked: a search-only run is not stopped by a missing species model, and tasks that run on an LLM or in the cloud need nothing downloaded. If the check itself cannot reach the backend the run proceeds, and any real failure is reported as usual.

Species identification

Everything runs on your machine; nothing is uploaded.

What gets written

Five fields in the LrGeniusAI metadata panel, on every photo that was checked:

Field	Example
Species (common name)	`Great Tit`
Species (scientific name)	`Parus major`
Species rank	`species`, `genus`, … or `none`
Species confidence	`0.91`
Species taxonomy	`Animalia>Chordata>Aves>Passeriformes>Paridae>Parus>Parus major`

All five are searchable, so you can build a Smart Collection on e.g. Species rank is species, or on a family name in Species taxonomy.

Looking the species up on the web

Two more fields — Species on iNaturalist and Species on Wikipedia — hold a link, and Lightroom draws them with a button that opens the page in your browser. iNaturalist is the one to click when you want to check the identification against photos, range maps and similar-looking species; Wikipedia is the one for what the organism actually is.

The links point at the taxon’s own page, not at a search: the backend resolves the scientific name against GBIF and iNaturalist — both free, and no account is needed — and remembers the answer, so each taxon is looked up once per machine and never again. If the lookup cannot be made (no internet at the time, or a name neither database carries), the fields hold a search link on the same site instead, which lands you one click away.

Look-up links in picks the Wikipedia edition and the language iNaturalist reports common names in. Automatic follows Lightroom’s own interface language. The identification itself is language-independent — this only changes which article opens.

Both fields are filled in as photos are analyzed, so photos identified before this feature existed have them empty; re-running Analyze & Index over them, or Retrieve Metadata from Backend, fills them in.

The AI model is told what the classifier found

When species identification and AI metadata run together, the identification is put into the prompt before the LLM writes anything, and the prompt says to use it. A general vision model is confident and often wrong about animals and plants — a photo of two rabbits came back titled “Sheep in Green Pasture” while the species field on the same photo read European Rabbit — and BioCLIP is a specialist trained on the tree of life, so where the two disagree the classifier wins.

A coarse answer is passed on as a limit rather than a name: told family: Leporidae, the model is asked to stay at that level instead of inventing a species. Photos with no identification are unaffected, and a photo identified by an earlier run contributes its stored answer to a later metadata-only run.

This needs no separate switch — ticking Identify animal and plant species is what turns it on.

Why the answer is sometimes just “Aves”

The model reports the deepest rank it is actually confident about. A clear frame of a garden bird gets a species; a distant silhouette gets an order or a class. That is deliberate — a coarse answer that is right is more useful than a binomial that is wrong, and the two are easy to tell apart because the rank is written alongside the name.

Photos where nothing cleared the confidence floor (including everything the pre-filter skipped) get Species rank = none and empty name fields. They still count as checked and are not re-examined on the next run.

Common names are not always in English

The common name comes from GBIF’s vernacular-name data, which has no reliable per-language coverage. Where no English name exists, whatever language GBIF lists first is used — so a run over European wildlife will produce Swedish and Danish names among the English ones (honungslök for Allium siculum, Almindelig kuglebærerflue for a fly). This is upstream data, not a setting.

The scientific name is always correct and always present, so use Species (scientific name) when you need something dependable, and treat Species (common name) as a convenience. The keyword branch below has the same property: the leaf uses the common name when there is one, with the scientific name attached as a synonym, so a search for the binomial always finds the photo.

Keywords

With Also write the taxonomy as keywords on, the identification is additionally written as a keyword hierarchy under a Species root, separate from the LrGeniusAI root the AI keywords use:

Species > Animalia > Chordata > Aves > Passeriformes > Paridae > Parus > Great Tit

Only the leaf is marked for export, so a JPEG exported from Lightroom carries Great Tit in its IPTC keywords and not the six ranks above it. The scientific name is stored as a keyword synonym, so searching for Parus major finds the photo too.

This is off by default because it changes your catalog’s keyword tree and travels with exported files, while the metadata fields stay inside the catalog. The fields are written either way.

Coverage

The bundled model covers the taxa a photographer is realistically pointing a camera at — birds, mammals, insects, reptiles, amphibians, fish, flowering plants, conifers, ferns, mushrooms — rather than all 867,000 taxa BioCLIP knows, which would be a multi-gigabyte download. Outside that set the model does not say “unknown”; it picks the closest thing it does know. The confidence floor is what keeps those cases at a coarse rank instead of a confidently wrong species.

Metadata Options

Metadata Tasks

One checkbox each for Keywords, Title, Caption, Alt Text.

Hierarchy & Language

Keyword Hierarchy → Enable — produces category-based keywords instead of a flat list. Edit categories opens the category editor.
Use existing catalog structure — takes the categories from the keyword hierarchy already in your catalog rather than the built-in list. Be careful with large or messy hierarchies: everything you have gets offered to the model.
Top-level Keyword — nests everything generated under one keyword of your choosing, so AI keywords stay separable from your own.
Bilingual Keywords — additionally produces keywords in a second language.
Keyword aliases — reuses keywords already in your catalog instead of creating near-duplicates (“bicycle” vs “bike”). See Keyword Dedup & Declutter.

Instructions / Prompt

Prompt templates. The text box holds the selected template’s instructions, and what is in it replaces the backend’s built-in persona for that run rather than being appended to it.

A template is a system prompt: it decides which expert is looking at the photo, which vocabulary they reach for, how specific they are allowed to be, and what they must never invent. It does not decide the output fields, the language or the keyword structure — those come from the request itself, from the switches on this dialog. So picking a template changes the voice and the precision of what you get, not its shape.

Default is the analytical voice: species, landmarks, vehicle makes, described precisely. Unchanged, and still what every catalog uses unless you pick otherwise.

Family & Everyday Photos is the album voice: who is in the picture, what they are doing, where, and what the occasion appears to be. It asks for the people by the names your catalog already has on them and for the place in the title, and it is explicitly forbidden to invent what it was not given — no relationships, no birthdays or weddings, no landmark the photo’s own data does not support.

The genre templates each bring a working vocabulary and a specific line they will not cross:

Wildlife & Nature — field-biologist naming: common name plus the scientific name where it is certain, genus or family where it is not, with behaviour, life stage, habitat and season. It will not anthropomorphise or guess a species.
Landscape & Travel — landform and terrain in the words a map would use, plus light, weather and season. It will not invent a named peak, lake, trail or landmark.
Architecture & Urban — building type, period and style, materials and the elements carrying the composition. It names a specific building or architect only when it is certain.
Events & Weddings — the part of the day a frame belongs to, people by name or by role, and the details that make a frame findable months later. It will not guess relationships, traditions or the meaning of a ritual.
Sports & Action — the sport, the discipline and the phase of play, filed the way a wire caption is. It will not invent a team, a competition or a score.
Street & Documentary — observational restraint: the gesture, the light, the geometry, the setting, with dignity toward everyone in the frame and no inferred story.
Portrait & Studio — the kind of portrait and the craft behind it, lighting pattern included. It will not speculate about the person.
Product & Stock — commercial keywording a buyer would actually type: shot type, material, finish, negative space, concept. Brands only when legibly visible.
Food & Drink — the dish, its components, the preparation and the styling, without inventing an ingredient or a dietary claim.
Night & Astro — what is actually in the sky and how the frame was made, without inventing a catalogue designation or an event.

The people-facing templates pair with People’s names from face recognition and Location below; without those switches they have nothing to name.

Both are ordinary templates once they appear: edit them, add your own, delete what you do not want. A template you delete stays deleted, and one you rewrite stays rewritten — the plug-in only ever offers each built-in once. “Reset to defaults” in the prompt dialog brings both back.

Clearing a template’s text is allowed and is saved: an empty template means “no persona of my own”, and the run uses the backend’s built-in one instead of being sent an empty instruction. The template itself stays in the menu — it is empty, not deleted.

Default is the one template the Delete button refuses, because it is the fallback every other selection resolves to. If it is missing from an older install it is put back the next time the plug-in loads; that is a repair, not a re-offer, and it leaves the templates you deleted deleted.

Context & Save

AI Context

Extra information sent alongside the photo to improve accuracy:

Existing Keywords — the photo’s keywords go into the prompt.
People’s names from face recognition — the names Lightroom put on the faces are sent, separately from the keywords and labelled as the people in the picture. Two things follow: a person called Ivo is no longer read as scenery and turned into “Ivo Beach”, and the model is asked to use the names, so a photo of two tagged people gets “Ivo and Myriam on the beach” rather than “a man and a woman on a beach”. With several people named and no way to tell which face is which, it names them together rather than guessing who is who.

Its own switch since it is its own decision: a name identifies a person, and a household happy to send “beach, sunset” to a cloud provider may well not want to send “Ivo, Myriam” with it. Untick it and no name from your catalog leaves the computer — for AI edits either, which never had a checkbox of their own and follow this one. On upgrade it inherits whatever Existing Keywords was set to, so nothing starts being sent that was not being sent before.
Folder Names
Location — on by default, and the place reaches the prompt from whichever of these knows it:
1. Lightroom’s own location fields (Sublocation, City, State/Province, Country) as the catalog holds them, which is also the only version that includes a place you corrected after import;
2. failing that, what the file itself carries — its IPTC/EXIF, and the XMP sidecar beside a raw original;
3. failing that, the photo’s GPS coordinates, looked up against an offline list of ~145,000 places (GeoNames cities1000, bundled — nothing is asked of the network). This is what covers the common case where Lightroom’s address lookup only ever suggested a city and you never confirmed it: the suggestions are not stored anywhere the plug-in can read, but the coordinates are. The prompt marks a looked-up place as looked up, and says “near” when the nearest known place is more than 5 km away, so the model names the area rather than inventing a landmark.
Raw coordinates only leave your machine if no place name can be worked out at all. Untick the option and nothing about where the photo was taken is sent.

Two configurations used to lose the location silently, and no longer do: Submit originals instead of exports (Plug-in Manager) and an export size above 2048 px. Both make the backend re-encode the image, and the JPEG it produces has no metadata left to read.
Ask for context before each batch — opens a dialog where you can type context for the upcoming photos by hand

Catalog Integration

Write generated data to Lightroom catalog — off means results are stored in the backend only, and you apply them later with Retrieve Metadata from Backend.
Review/Edit each photo before saving — opens a validation dialog per photo. A species identification appears there too, read-only (there is nothing to reword in a taxonomic name) with its own Save species tickbox. Discard and Cancel now drop the species along with everything else — before, the species fields, the look-up links and the species keywords were written whichever button you pressed.
Import metadata from catalog before indexing — pushes the metadata you already have into the backend first, so the AI does not overwrite it blindly.

Data Handling

Mode: Regenerate all (overwrite existing AI data) vs Skip photos with existing data (default).
Write: Append to existing values instead of replacing — adds the new text below what is already in the field rather than overwriting it.

Retrieve Metadata from Backend — apply stored results to the catalog later, or after a failed write.
Import Metadata from Catalog — one-way sync of your Lightroom metadata into the backend database.