The Server guide.

Photos per year, and a week-by-week "where you were" ribbon with automatically detected trips. Filter the whole picture through a lens — one person, one place, one tag — and watch decades re-draw themselves.

A constellation of the people in your photos, linked by how often they appear together, plus leaderboards. Click a person for their card: their years, their places, the people they're photographed with — and keep clicking from card to card.

Cities and countries ranked by your photos, each with its own card. If the map data splits one spot into several names, Settings → Location Groups lets you merge them.

Which cameras and phones shot your library, your top tags, and the records: busiest day, longest quiet spell, home base, furthest photo from home.

Start a run whenever you like and give it a budget — say, a few hundred photos — then watch the progress strip. Runs can be stopped and resumed; history is kept. Photos that keep failing gather in a "Stuck photos" strip with a Retry button.

Under Settings → Enrich, turn on Daily Enrich and choose a local daily time and photo budget. Saving captures this browser's time zone. Each automatic run uses the provider last selected on Enrich, analyzes only photos that have never been successfully enriched, sends successful results to Curate, and appears in Recent Runs. A manual or queued run keeps priority. Daily Enrich runs only once per day; if today's time has already passed when you enable it, today's catch-up starts promptly. If an automatic run starts but fails, it waits until the next day rather than repeatedly retrying.

A review list shows what the AI wrote for each photo. Agree, fix, or send a photo onward to Curate for a human decision.

Connection details and model identifiers live under Settings → AI Providers. Choose the provider for new runs on Enrich; that choice is remembered across visits and server restarts. Local Ollama or LM Studio keeps photos on your hardware, or you can use your own OpenAI, OpenRouter, Ollama Cloud, or Venice key. Optionally, captions can be written back to Immich descriptions — off unless someone turns it on.

A cloud vision model with structured output. The current Venice catalog marks it Private; verify that status when you set it up.

The same cloud model also accepts the multiple images a Stack comparison needs. The reference setup leaves Curate on Follow Enrich provider.

A cloud model with image input for spoken answers about the photo currently on screen.

A small, fast cloud text model for general questions whose answers are spoken while someone waits.

The speech model; voice, speed, and output format remain separate choices under Settings → Voice TTS.

Photos taken in the same moment — bursts, retakes, near-duplicates — collapse into a stack instead of cluttering the list. Open a stack to compare side by side.

Each stack comes with a suggested keeper: a silver ★ picked by photo-quality scores, or a gold ★ when the optional AI referee has compared the whole burst side by side (with a short note on why). One action keeps the best and skips the rest — or pick a different winner yourself.

Everything has a keystroke, so a rainy afternoon genuinely puts a dent in a 30,000-photo backlog. And it all works with AI switched off — the stars are suggestions, never decisions.

Compose filters: people, places (multiple cities work), dates, cameras, tags, or a free-text search. A preview with a result count shows exactly what you'll get before anything is created.

For free-text rules, "best of" double-checks every search hit against your own enrichment data and ranks keepers using your Curate decisions — a "Top 50" that is your best 50, not a search engine's first 50. Photos not yet enriched are reported honestly instead of guessed at.

Every album the server maintains, with its rule and schedule. Edit the rule, run it on demand, or stop managing an album at any time — removing a rule never deletes the Immich album.

How Pictaria reaches your Immich library — the URL and API key (the key's exact permission checklist is in Getting started) — plus its browser-facing URL when that differs. Changes apply immediately, no restart. The Immich connection supports HTTP and HTTPS. Include https:// explicitly for TLS and use the Immich root address without /api; an address without a scheme defaults to HTTP. Public certificates work normally. A private or home CA must be mounted into the container and named with NODE_EXTRA_CA_CERTS, as shown in the Configuration guide. Certificate validation stays enabled.

Reusable provider connections, credentials, endpoints, and provider-default model identifiers, grouped into Local Models and Cloud Models. Choose the provider for new enrichment runs on Enrich, where the choice is remembered. The ElevenLabs key also lives here, but it is used only for text to speech; its model, voice, and output choices stay under Voice TTS.

Every frame that has ever connected, by name. Rename or remove retired devices; this is also where a re-onboarded frame's old entry gets cleaned up.

Merge nearby map labels into one place ("Brooklyn" and "Kings County" become one), and rename places the way your family actually says them.

Custom wake phrases can be uploaded here as openWakeWord models. Each frame installs its selected phrase from the server, and detection stays on the frame. Create your own phrase at openwakeword.com; details in the wake-words doc.

The feature-specific choices for spoken answers: provider, Interesting photo model, Ask text model, TTS provider, model, voice, and output. Connections and keys come from AI Providers; try-it-now testers here let you hear an answer before using it on a frame.

Schedules — including Daily Enrich's local time and photo budget — image size and run limits, taxonomy and prompts, grouping, and the optional AI referee. Enrich's active provider is chosen on the Enrich page itself; Curate can follow it or use its own provider and model override.

Scheduled backups of everything Pictaria knows (decisions, enrichment, albums' rules). Your photos are not in these — they never leave Immich.

Light or dark dashboard, and the place to paste a supporter key if you have one — which adds a badge and our gratitude, and changes nothing else.

Read that release's notes, open Settings → Backups, select Back up now, and confirm the new snapshot is complete. Note both its name and the version shown under Settings → Server. Do not continue with a failed or incomplete backup.

Docker installations need the release-matched Compose file as well as its image; pulling an image alone can leave an old deployment definition behind. Source installations should switch to the named version tag rather than pull the moving main branch. The exact commands and customization safeguards are in the technical upgrade procedure.

When a release changes persisted state, an established install creates and verifies a specially retained pre-migration snapshot before any store opens. An older install adopting Pictaria's safety metadata for the first time establishes a baseline instead, which is why the fresh manual backup is required for every upgrade.

Confirm Settings → Server shows the new version, Home reports Immich connected, and a frame still displays photos and responds to one Remote or voice command. If a rollback is needed, stop the server first and follow the release notes and technical procedure: a migration requires restoring its complete pre-migration snapshot, while a release with no persisted-state change can be returned to the prior code without discarding new work.