Skip to main content

Titles, maps and music

These are render choices. You can change them after reviewing the cut, without asking the app to choose the shots again.

Titles and language​

Under Render, type a Title and Subtitle, or leave them empty for the app's own. The app titles a film by what it is and when, in the film's language. No model writes these:

FilmTitle
A year, month or seasonits period: "2025", "July 2025", "Summer 2025"
One personfirst name and year: "Kit · 2025"; with no dates, the first and last year their pictures show: "Kit · 2003–2026"
Several peopleevery first name, whether you picked them by name, group or expression: "Kit, Pat & Lou · 2025", with the same span of years when you gave no dates
A birthdaywhose it is and the year being celebrated: "Kit's birthday · 2026"
A holidayits own name and the years the cut shows: "Assomption · 2015–2026"
On this daythe day and its span of years: "October 15 · 2012–2025"
A special day the scan never cataloguedits place: "A day in Lisbon"

Nothing goes under the title unless you type a subtitle. With title_screens.llm_titles: true (experimental, off by default) an optional text reader names special days instead, from their stories, people and places. A reader title keeps the year the template title would show; one that drops it falls back to the template. A reader subtitle may only carry dates, places, people and words the film's facts use: a mood line the facts don't back ("salt air and golden light") is dropped and the title stands.

The film's language is separate from the interface language. Save Settings > Title screens > locale (including in Docker), or set it in your configuration. A special day found by the automatic scan is catalogued with an English name, but its opening title is always written in this configured language, not in English.

title_screens:
locale: fr

The default, auto, follows the host's language: LC_ALL, LC_MESSAGES, LANG and LANGUAGE are checked in that order, and on a Mac with none of those set, the macOS language setting is used instead. A Docker container usually starts with none of them set, so it renders English titles until you set title_screens.locale or the container's LANG.

A trip is named at one scale (a town, a region, a country, whichever fits how far its pictures spread) and every part of that name is shown in this same configured language, in the opening title, the trip picker in the web UI, and the discovery table a generate --memory-type trip run prints before cutting one. A French film never shows a region translated next to its country in English, or the other way round.

Opening and closing fades​

Choose Opening and closing fade under Render: white, black, or As configured. This changes the fade at both ends of the title sequence for this film. Title screens must be enabled.

Set the default in Settings → title screens → fade_color, or in your configuration:

title_screens:
fade_color: black # white is the default

For one CLI render, add --fade-color black to generate or runs render. The override leaves your saved default alone.

Styles​

Choose Title style under Render for one film. As configured keeps the saved default. The choices are Auto (match the mood), Random, Modern warm, Elegant minimal, Vintage charm, Playful bright and Soft romantic. For a lasting default:

title_screens:
style_mode: elegant_minimal

auto and random are also valid defaults. For a saved CLI cut, use runs render RUN_ID --title-style elegant_minimal. Preview a card before rendering the whole film:

immich-memories titles test --year 2025 --style elegant_minimal

In Docker, write the preview to the mounted output folder:

docker compose exec immich-memories immich-memories titles test --year 2025 --style elegant_minimal -o /app/output/title-test.mp4

The host file is ./output/title-test.mp4.

The style and font reference lists the palettes and supported alphabets. Docker includes the fonts; native installs may need immich-memories titles fonts --install.

On a CPU or NAS, title cards keep the font, layout and palette. FFmpeg slides, scales and fades the text over a still background; the text is drawn once. Moving backgrounds, bokeh, animated deblur and separate title/subtitle timing still need a rendering GPU.

Date and place captions​

In Render, untick Add date overlay or Caption clips with their place if you want a cleaner frame. Frequently visited places are not labelled over and over. The lasting defaults are defaults.add_date and defaults.add_place.

Place captions show the city when it changes. The home country stays hidden. Abroad, the country appears when you enter it, unless the opening title already named it; the next city doesn't repeat it. Crossing another border shows the new country.

With network.geocoding: true, Nominatim resolves rounded coordinates at zoom 16 in your chosen title/caption language. If a translation is missing, it tries the base language (for example, Portuguese for Brazilian Portuguese), then falls back to the place's own native name rather than English. A Greek village with no French name stays in Greek script in a French film, not translated or anglicised. A town keeps its name even when no translation exists. Administrative words come off: "Limassol District" reads Limassol, "Pärnu linn" reads Pärnu. A Latvian, Lithuanian or Finnish municipality name comes in the genitive ("Helsingin kaupunki"), so the caption uses the city Immich stored for the picture instead ("Helsinki"). Some places write their name in several alphabets at once, as many Moroccan ones do ("Medina ⵎⴷⵉⵏⴰ المدينة"). The caption keeps one: the part in your film's alphabet ("Medina" in a Spanish film), or the first part when none matches. A long place name shrinks to fit the frame, portrait included, and only gets cut short with "…" when it still doesn't fit at the smaller size.

Near the configured home base, captions can name the district, such as Montmartre in Paris; further out, a district needs most of a stay's pictures behind it to keep its name, and excursions keep their own labels. Visits spread across districts use their shared locality, and towns separated by more than trips.max_gap_days do not rename one another. Streets and points of interest stay out of captions. A country disagreement keeps the source label; failed lookups are retried on a later run. A few distant excursions do not turn a local stay into a regional trip, and a town already named by its own pictures is not replaced by a broader label from the trip's centre point. The output reference has the exact distances and shares.

The map fly-over​

Trip maps are off by default. Enable them when you want an animated route and are comfortable requesting satellite tiles from an outside provider:

In Docker, save Settings > Network > map_tiles, or use YAML:

network:
map_tiles: true

With maps off, a trip uses ordinary title and location cards. Smooth maps fly between stops; the fast preset uses three views joined by short fades. Maps add time to the finished film beyond the picture budget. Map rendering covers providers, route grouping and timing.

Music​

In Render, choose Automatic, No music, or upload your own MP3, M4A or WAV. Music volume changes the mix; the music drops under the clips' own sound. When a clip's own sound is music or singing, the soundtrack steps aside for its whole window instead of competing with it.

Untick Keep original audio to remove all sound recorded in videos and Live Photos, including speech. The film keeps its motion and plays only the soundtrack. With No music selected too, it is silent. Original audio stays on by default.

For the same choice on the CLI, add --no-original-audio to generate or runs render:

immich-memories runs render RUN_ID --no-original-audio --music ~/Music/track.mp3

Automatic uses a bundled track by default: 76 generated tracks, each about 90 seconds, in five moods (calm, tender, nostalgic, happy, energetic) and styles from piano and lo-fi to downtempo, neo-soul, post-rock, funk and city pop. The Docker image has them built in; on a pip install, immich-memories models fetch downloads them once (118 MB, checked against a pinned SHA-256), and a film never downloads music. Without them a film plays the smaller set of 28 shorter tracks that ships with the music extra. models fetch skips the download when audio.no_music is on. A configured music generator running the XL model makes an original track instead and enables Preview a track so you can listen first. A generator that fails, or can't run XL, falls back to a bundled track and leaves a warning.

Your own music folder​

Put your own tracks in audio.local_music_dir (~/Music/Memories unless you change it), one subfolder per mood, the way the bundled tracks are laid out:

~/Music/Memories/
calm/ slow-piano.mp3, rain-guitar.flac
happy/ ukulele-morning.m4a
tender/ lullaby.opus

Automatic then picks from your folder instead of the bundled tracks, by the film's mood, the same way: a beat that fits the photo cadence first, and one track per film, looped onto itself when the film is longer. Folder names are moods: calm, peaceful, tender, nostalgic, playful, happy, exciting, energetic, and the rarer ones (sad, romantic, dramatic, mysterious, inspiring, melancholic, uplifting). A playful film without a playful folder takes happy; peaceful takes calm, exciting energetic and romantic tender, as the bundled music does.

A mood your folder lacks plays the bundled tracks for it. A mood neither has (a year in review starts as inspiring) picks from all of your folder. An empty or missing folder leaves the bundled music in charge, and the log says so without printing the path. A configured generator still comes first; your folder replaces the bundled fallback, not the generator. In Docker, mount the folder and point audio.local_music_dir at the mount.

No music by default​

audio:
no_music: true

Films then render without music on every surface: generate, runs render, the daily timer and triggered runs, and the web Render panel, which opens on No music. A run can still ask for music: --music FILE, --music auto, or picking Automatic or a track in the panel. --no-music turns it off for one film when the config leaves it on.

Temporary files used to loop and master bundled tracks are removed after mixing, including when the mix fails.

On the CLI:

immich-memories generate --year 2025 --music ~/Music/track.mp3 --music-volume 0.4

When a model names the film​

Model-written titles and summaries are experimental and off by default: they miss too often. Turn them on with title_screens.llm_titles: true, or for one run with --llm-title. Then a text reader names special days (and trips, with --llm-title) and writes the film's summary. A reader still chooses the music mood on the Basic tier, whatever this setting says. Those calls use text, not pictures. Years, months, seasons, people, birthdays, holidays and "on this day" films keep the app's own title whatever the reader could say. Your typed title wins over all of it. Title provenance explains the source shown on a run.

A model title can use a first name, a place, an occasion name or a relationship word your family record actually states between the people it names, such as "her grandmother". It never calls someone family, or states a relationship, that your record doesn't back, and it never names a year that isn't true for the span. It also never narrates a parent or grandparent from a child's point of view ("maman", "papy", "mum", "oma") even when your record backs the relation; standard words such as "abuela" or "mormor", which have no separate child's-eye form in their own language, are unaffected. A title that breaks a rule falls back to the plain template instead.

Generated music​

Set up generated music when you want it. The output reference has language catalogues, typography, map timing and audio mixing details.