immich-folio/docs/deployment.md

Deployment

Operational detail for running Immich Folio: what the setup wizard writes, the Docker options beyond the basic Compose recipe, and what happens when Immich goes away.

For the quick path — clone, npm run dev, open /install — see the README.

Contents:

• What the Setup Wizard Writes
• Environment Variables Always Win
• Docker Compose
• Standalone Docker
• Health Check
• Config Doctor
• Behaviour when Immich is Unreachable
• Reverse Proxy
• CDN Mode
• Edits made in Immich
• Colour Profiles

What the Setup Wizard Writes

The wizard at /install writes three files into content/:

File	Contents
content/gallery.yaml	The albums you picked
content/settings.yaml	Site title, subtitle, theme — only if it does not exist yet
content/install.json	Immich URL and API key, a generated site secret, admin password hash

Important

content/install.json holds credentials. It is written with mode 0600, and it is the reason a backup of your content/ directory is also a backup of your Immich API key — treat it accordingly. The admin password is stored as an scrypt hash, not as you typed it; the API key and site secret have to stay readable to be usable.

Nothing is written until your credentials have been verified against your Immich server, so a typo cannot leave you with an "installed" site that loads no photos.

Environment Variables Always Win

IMMICH_API_URL, IMMICH_API_KEY, AUTH_SECRET and ADMIN_PASSWORD override anything in install.json, so any of them can be rotated by setting the variable — no need to touch the file. SITE_PASSWORD works the same way over sitePassword in settings.yaml, and so does CONTACT_NOTIFY_URL over contact.notifyUrl.

Set all of them up front and the wizard never appears, which is the usual choice for an infrastructure-as-code deployment.

Docker Compose

services:
  lightbox:
    build: .
    container_name: immich-folio
    restart: unless-stopped
    ports:
      - '7211:7211'
    env_file:
      - .env.local
    volumes:
      - ./content:/app/content

docker compose up -d

The gallery will be available at http://localhost:7211.

Standalone Docker

# Build
docker build -t immich-folio .

# Run
docker run -d \
  --name immich-folio \
  --restart unless-stopped \
  -p 7211:7211 \
  --env-file .env.local \
  -v ./content:/app/content \
  immich-folio

Important

The content/ volume mount lets you update gallery.yaml and about.md without rebuilding the image. It must be read-write: the setup wizard, the admin panel, the journal, the favicon upload and the backup rotation all write into it. A :ro mount leaves the wizard unable to complete and the admin panel unable to save.

Health Check

The container includes a built-in health check at /api/health:

curl http://localhost:7211/api/health

Config Doctor

npm run doctor

Checks an installation from the terminal and prints what it finds: whether AUTH_SECRET is set and long enough, whether gallery.yaml and settings.yaml parse, whether Immich answers the three calls Folio depends on, whether every published album ID still exists in Immich (and is shared there), whether any password is still stored in plaintext or as an unusable bcrypt hash, whether an enabled Impressum has a name, address, email and a second contact channel, whether the contact form notifies anyone, and whether content/ can be written to.

This is the same set of checks as the Diagnostics page in /admin, but it needs neither a running app nor an admin password — which is the point. It is the tool for the case where the site will not come up at all. (The page adds an alt-text report on top, which needs the albums loaded and has no CLI counterpart.)

It runs in the shipped image too:

docker compose exec folio npm run doctor

No secret is ever printed — only whether something is set, and where to look. The exit code is the worst thing it found, so it can gate a deployment script:

Code	Meaning
0	every check passed
1	warnings, no errors
2	at least one error
3	the doctor itself could not run

Two findings are reported rather than judged. content/ writability is tested as the user who typed the command, and on a Docker deployment that is not the user the app writes as — so when the paths belong to someone else, the CLI says whose they are and leaves it at that. Run it through the container to have it checked as the app itself:

docker compose exec folio npm run doctor

The other is TRUSTED_PROXY_HOPS, judged against the X-Forwarded-For chain of a live request, and a CLI has none. It reports the configured value and says so rather than guessing — use the Diagnostics page, reached over your public URL, to have that one measured.

Both are printed under NOTES rather than among the passed checks, since nothing was established, and counted separately in the summary line for the same reason. Notes never affect the exit code. An unwritable path does stay a plain error when the CLI runs as the owner, which is the case that is really about the app.

Run from a checkout, it reads .env.local and .env the way npm run dev does; real environment variables and content/install.json are resolved in the same order the app resolves them.

It needs Node 22.18 or newer, which is what runs the TypeScript directly and is why the CLI adds no dependency of its own. The shipped image is already on it; an older local Node is told so instead of crashing.

Behaviour when Immich is Unreachable

Immich Folio buffers your gallery rather than merely proxying it:

• Album and asset pages keep serving the last known good data for up to STALE_MAX_AGE, so a restarting or briefly unreachable Immich does not take the public site down with it.
• Once nothing cached is left, they return 503, never 404 — a 404 would tell search engines to drop a URL for content that still exists.
• Outages are never cached, so the gallery recovers as soon as Immich does.

The cache lives in the process, so it is empty right after a container restart.

Reverse Proxy

Put Immich Folio behind nginx / Caddy / Traefik with TLS. Example Caddy config:

photos.example.com {
    reverse_proxy localhost:7211
}

Important

Set TRUSTED_PROXY_HOPS to the number of proxies in front of the app, and make sure they forward X-Forwarded-For correctly — otherwise the client IP comes from a header the client itself can set, which defeats the brute-force limits on the password endpoints. See Trusted Proxies for the setting and a matching nginx config.

Login cookies (admin session and gallery passwords) are marked Secure when the request arrived over HTTPS, which the app learns from X-Forwarded-Proto. Caddy and Traefik send it on their own; nginx needs the proxy_set_header X-Forwarded-Proto $scheme; line from the config linked above. Opening the app over plain HTTP, e.g. http://server:7211 on a home network, works too: the cookies are then set without Secure, since a browser would discard them otherwise.

CDN Mode

With CDN_URL set, every photo and video URL on the site points at a pull CDN instead of this server:

CDN_URL=https://cdn.example.com

https://cdn.example.com/api/image/v2:…?size=preview&w=1440

The CDN uses Immich Folio as its origin. The first request for a photo is a cache miss and goes to /api/image as usual; every request after that is answered by the CDN and never reaches your server or the Immich box behind it. On a home connection with a slow uplink, that is the difference between a gallery that crawls and one that does not.

Setting up the CDN. Create a pull zone (Bunny), a distribution (CloudFront) or a CNAME'd subdomain (Cloudflare) whose origin is your public Folio URL, then:

• Cache /api/image/* and /api/video/*, and nothing else. Pages are rendered per request, and other routes check cookies.
• Include the full query string in the cache key. size, w, the edit marker e and the IMAGE_CACHE_VERSION buster v all select a different file.
• Respect the origin's Cache-Control. Photos come back public, max-age=31536000, immutable; errors (429, 503, 404) come back no-store and must not be cached.
• Forward Range requests for videos, so seeking works.

Rate limiting. A cache miss arrives from a CDN edge, not from the visitor. Set TRUSTED_PROXY_HOPS to count the CDN as a proxy (Cloudflare → Caddy = 2) so the limiter reads the visitor's IP from X-Forwarded-For; otherwise a handful of edges share one bucket. The Config Doctor warns when CDN mode is on with TRUSTED_PROXY_HOPS=0.

What stays on this server: pages, /_next assets, EXIF, the map, downloads, the admin panel and the Open Graph images. These are either small or carry a cookie check a CDN would bypass.

Site passwords switch it off. On a password-protected site, the image route checks the visitor's unlock cookie, and a CDN would answer from its cache without asking. So while a site password is set, photos stay on this server whatever CDN_URL says, and the diagnostics page says so. Gallery (per-page) passwords are unaffected: they already protect the page that holds the photo links, not the image URLs themselves.

Changing a photo. A CDN keeps immutable responses for as long as a browser does. Bump IMAGE_CACHE_VERSION after regenerating thumbnails in Immich, which changes every URL at once, or purge the CDN. Edits made in Immich's editor need neither: see Edits made in Immich.

The Content-Security-Policy allows the CDN's origin for img-src and media-src automatically.

Edits made in Immich

Photos edited in Immich's own editor (crop, rotate) appear as edited everywhere: in the grid, the lightbox, the zoom, share images, the map and downloads. Immich keeps an edit apart from the file and applies it only when asked to, so Folio always asks for the edited rendition (for a photo that is not edited, Immich returns the same file either way; the parameter exists since Immich 2.5, below the 3.0 that Folio requires), and lays the grid out with the edited proportions.

The image URL of an edited photo carries an extra e parameter, derived from when the photo last changed in Immich, to the second. Image responses are cached as immutable for a year and the asset ID stays the same through an edit, so without it browsers and CDNs would keep showing the photo as it was before. Only edited photos get a new URL; all other URLs stay as they are. A later edit changes the parameter again, once Folio's album cache has picked it up (CACHE_TTL).

Downloads of an edited photo are Immich's rendition of the edit: a full-resolution JPEG without EXIF, named .jpg. Should Immich hand out the edit in a format the location scrubber does not handle (WebP, PNG), the photo is not downloaded rather than sent unchecked; a ZIP goes on without it and the server log names it.

Colour Profiles

Visitors only ever get Immich's own renditions: the thumbnail for grid tiles and the preview for everything larger. Folio streams those bytes unchanged. It never decodes or re-encodes a photo, so whatever colour profile Immich embeds is what the browser sees, and colour fidelity is decided by how Immich builds its renditions.

Immich's relevant setting is image.colorspace (Administration → Settings → Image Settings), p3 by default. Measured on Immich 3.2.0 with p3, thumbnail as WebP and preview as JPEG:

Original	Thumbnail	Preview	Colours
JPEG, sRGB ICC profile	sRGB	sRGB	correct
HEIC from an iPhone, Display P3 ICC profile	Display P3	Display P3	correct, wide gamut kept
AVIF, no ICC profile, CICP (nclx) BT.2020 tags	sRGB (P3 for an iPhone shot)	sRGB (P3 for an iPhone shot)	13–37% less saturated

The last row is a loss in Immich, not in Folio. Immich ignores the CICP colour description of these AVIF files and labels the BT.2020 pixel values as sRGB (or Display P3) without converting them, so the browser shows them flatter than the original (mean chroma, compared with the decoded original). The files measured were AVIF exports, two of them confirmably from Lightroom. Exporting with an embedded ICC profile, or as JPEG or HEIC, should avoid it, since ICC-tagged originals came through correctly, but that has not been tested. Adobe RGB originals were not part of the measurement.

After changing image.colorspace or replacing originals, regenerate thumbnails in Immich and bump IMAGE_CACHE_VERSION, since browsers and CDNs keep the old renditions as immutable.