* fix(mobile): don't block app open on slow validateAccessToken
AuthGuard.onNavigation was async so auto_route awaited the body through validateAccessToken's OS timeout. now it's sync and the validate runs in bg. kicks to login on 401.
* fix(mobile): handle re-login race in AuthGuard validate
if user logs out + logs back in during a slow validate, the old 401 was logging them out again. now we check the token hasn't changed before redirecting, and dedupe in-flight calls.
---------
Co-authored-by: Alex <alex.tran1502@gmail.com>
* fix(mobile): clear linkedRemoteAlbumId in reset() so FK refs dont dangle
reset() runs with foreign_keys off before wiping remote_* tables, so the ON DELETE SET NULL cascade on linkedRemoteAlbumId doesnt fire. local rows keep pointing at deleted remote ids.
affects logout (clearLocalData calls reset()) and the server SyncResetV1 path (30 day idle, etc). after re-login, syncLinkedAlbum either silently warns or fires 400s (those are covered by #28299).
null the column manually inside the same transaction. cascade still works for normal SyncAlbumDeleteV1.
verified on pixel 9a with this branch built locally: logged out, deleted album from web, logged back in. without fix linkedRemoteAlbumId stayed dangling. with fix all three local rows have linkedRemoteAlbumId = NULL after the logout reset, and recovery is clean once manageLinkedAlbums runs again.
* fix(mobile): always re-enable foreign_keys in reset() + simplify the update
re-enable foreign_keys inside a try/finally so it always runs even if the transaction throws. without this, a failed reset would leave the connection with foreign_keys = OFF and silently disable cascades for everything after (per copilot review).
also drop the where filter on the linkedRemoteAlbumId update, unconditional update-all is simpler and we wipe everything in reset anyway (per ganka review).
server removed both fields from AssetMediaCreateDto in #27818. zod silently strips unknown fields so uploads still work, but we send dead weight on every request.
drop from foreground + background upload paths + share intent path. deviceAssetId stays as the internal background_downloader taskId, just not in the multipart form fields anymore.
* feat: add slideshow metadata overlay and settings
* Introduced a new SlideshowMetadataOverlay component to display image information during slideshows.
* Updated slideshow settings modal to include options for showing the metadata overlay and selecting its display mode (Description Only or Full).
* Added corresponding translations and store management for the new overlay features.
* remove noisy log
* constant opacity
* 2nd pass
* more
* use text components
* lint
* feat(ci): publish PR Android APK to R2 with installable links
Adds a universal debug APK to PR builds and uploads it to a public
R2 bucket alongside the existing GitHub Actions artifact. Posts a
sticky PR comment with tap-to-install links and a QR code so testers
can install directly on their device without unzipping artifacts.
Required setup:
- Secrets: R2_APK_ACCESS_KEY_ID, R2_APK_SECRET_ACCESS_KEY,
R2_APK_ACCOUNT_ID, R2_APK_BUCKET
- Optional repo variable: APK_PUBLIC_HOST (defaults to apk.immich.app)
- R2 bucket configured with a public custom domain matching APK_PUBLIC_HOST
* chore(ci): drop R2 upload, link directly to GitHub artifact
Surfaces the existing release-apk-signed artifact in a sticky PR
comment with a QR code. Avoids new infra and secrets — the trade-off
is GitHub login and a zip wrapper instead of tap-to-install.
* feat(ci): build PR APK as release and publish to GitHub Release
PR builds now produce a release APK signed with the release keystore.
The universal APK is published as a GitHub Release asset under tag
'pr-<num>' (prerelease), giving testers a direct, unzipped, tap-to-
install URL plus a QR code in the PR comment. The release-apk-signed
artifact is unchanged.
* chore(ci): drop GitHub Release, publish universal APK as own artifact
Reverts the prerelease publish. Uploads the universal release APK as
a separate single-file artifact so its download URL gives a zip
containing only that APK — no extra files to dig through. The QR in
the PR comment points at this universal-only artifact.
* chore(ci): build only universal APK for PR, drop split artifact
PR builds skip the arm64-only split — release-apk-signed now contains
just the universal app-release.apk, so the download zip is a single
file. Removes the redundant separate universal artifact and points
the PR comment QR at the main artifact URL.
* feat(mobile): suffix PR APK applicationId so it installs alongside production
Each PR build now becomes app.alextran.immich.pr<num> via PR_NUMBER env
read in build.gradle, so testers can install multiple PR builds and the
Play Store version on the same device without uninstalling. Also tags
the version with -pr<num> for visibility.
* feat(ci): allow PR APK build to run on forks
Forks can now run the Android build job. Steps that need repo secrets
(create-workflow-token, Create Keystore) are skipped when the PR is
from a fork, the checkout falls back to GITHUB_TOKEN, and build.gradle
falls back to debug signing if the release keystore isn't materialised.
The PR comment still requires write access, so it's gated to non-fork
PRs — fork APKs are reachable from the workflow run's artifact tab.
fix(mobile): deduplicate assets in person view timeline
Previously, assets with multiple face records for the same person (e.g.,
manual Digikam imports and Immich ML detections) appeared multiple times
in the person timeline. This was caused by an inner join on the
assetFaceEntity without proper deduplication.
This commit refactors the timeline queries to use a subquery approach
instead of joins and grouping. This ensures:
- _getPersonBucketAssets: Only unique assets are fetched, even if
multiple face records exist for a single asset.
- _watchPersonBucket: Asset counts in timeline headers are accurate
and represent unique assets.
- Performance: Database overhead is reduced by avoiding complex joins
and explicit groupBy operations on large result sets.
Signed-off-by: thowdev <12428285+thowdev@users.noreply.github.com>
_manualSyncAlbums fires a setState 1s after sync via Future.delayed
with no mounted check. if the widget is gone by then, setState throws
null check and the global error logger logs it severe.
#27666 removed LocalNotificationService with the legacy stack, which
was the only place calling FlutterLocalNotificationsPlugin().initialize().
without it, ios never prompts for the notification perm on fresh
installs so background_downloader notifications get dropped silently.
restores the init in the same spot the deleted call used to live.
* hide hidden person from memories
* clean up
* fix united test
* clean up
* moved sql to inline, rebased
* clean up
* clean up again
* chore: sync sql
---------
Co-authored-by: Jason Rasmussen <jason@rasm.me>
* feat(server)!: add owned filter to albums API
BREAKING CHANGE: GET /albums with no parameters now returns all accessible albums (owned + shared-with-me) instead of only owned albums.
* document tri-state matrix
* web impl
* collapse to single method and handover branching to sql
* dedupe
* verify that owned, shared, and notShared counts are mapped independently from their respective queries
* refactor(server): add select:['id'] overload to albumRepository.getAll
Avoid fetching full album rows (with albumUsers/sharedLinks subqueries) in map.service where only album IDs are needed.
* focus relevant test filters
* fmt
* Revert "verify that owned, shared, and notShared counts are mapped independently from their respective queries"
This reverts commit 47aab458192c766de4662aada5a6841b091d2a80.
* sync sql
* Revert "document tri-state matrix"
This reverts commit a5b2355d0c.
* address review comments
* inline shared condition and return as ternary
* sync sql
* use [...albums].sort
Array.toSorted() is not supported in Chrome 109
* use isShared and isOwned nomenclature
* fix e2e tests
* add params to sql query
* fix(mobile): view similar defaults to images only
* fix(mobile): reset filter chips when pre-filter is applied
---------
Co-authored-by: shenlong <139912620+shenlong-tanwen@users.noreply.github.com>
* refactor: app metadata
* refactor to per row store
* cleanup
* more test
* review changes
* more refactor
* refactor
---------
Co-authored-by: shenlong-tanwen <139912620+shalong-tanwen@users.noreply.github.com>
* fix(web): fix shared link navigation after password login
* use regex after all
* chore: use special case for shared link with slug route
* dont use onMount
* fix lint
---------
Co-authored-by: Alex <alex.tran1502@gmail.com>
* feat(web): custom video player controls
* add seek & rate buttons
* wrap memory viewer in media-controller for muted/volume store
* fix memories
* disable video shortcut keys
* re-add playsinline for safari iphone playback
* fix black screen issue
* always display time range
* remove seek buttons and center controls, and put time range above controls
* change ui
* update memory viewer
* fix full width on video player on safari
* enhance video player layout by ensuring full width and maintaining aspect ratio
* layout: don't shrink buttons, tabular time text
---------
Co-authored-by: timonrieger <mail@timonrieger.de>
* refactor(server)!: sanitize error messages to avoid leaking resource and permission details
* fix e2e tests
* fix(server): prevent login timing oracle by always running bcrypt
Always call compareBcrypt in the login path regardless of whether the
email is registered. When no user is found, a dummy hash is used so the
bcrypt KDF still runs and response latency is constant, making it
impossible to enumerate valid email addresses by measuring response time.
* fix(server): collapse OAuth callback messages to prevent email-existence oracle
Two distinct error messages in the OAuth callback endpoint revealed
whether an email address was already registered in the database.
An attacker controlling the OAuth provider's email claim could probe
the user table without authentication. Both cases now return the same
generic message.
* fix(server): replace email-in-use messages to prevent user-existence oracle
Error messages on registration and profile-update that named whether an
email address was already taken allowed callers to enumerate registered
accounts. All three sites now return the same generic message regardless
of whether the address is in use.
* fix(server): hide slug uniqueness constraint to prevent shared-link probe
Surfacing the Postgres unique-constraint name in the error response let
any authenticated user brute-force whether a custom slug was already in
use by another user's shared link, leaking the existence of other links.
* fix(server): unify profile image errors to prevent user-existence oracle via status code
GET /users/:id/profile-image returned HTTP 400 for an unknown user ID
but HTTP 404 when the user existed without a photo, letting callers
distinguish the two cases. Both now return 404 so the response is
identical regardless of whether the UUID maps to an account.
* fix(server): replace album user-not-found message to prevent UUID-existence oracle
Album owners could probe arbitrary UUIDs via the add-user endpoint and
determine whether they belonged to registered accounts by receiving
'User not found'. The message is now ambiguous about whether the ID was
unrecognised or the user is inactive.
* Revert "fix e2e tests"
This reverts commit c1bd7a116b.
* Revert "refactor(server)!: sanitize error messages to avoid leaking resource and permission details"
This reverts commit b96421a083.
* fix(server): use 403 instead of 400 for access-denied errors
requireAccess threw BadRequestException which is incorrect HTTP semantics.
Access denial is a client authorization problem (403 Forbidden), not a
malformed request (400 Bad Request). Keep the descriptive permission name
in the message since the full permission set is public API surface.
* Revert "fix(server): use 403 instead of 400 for access-denied errors"
This reverts commit bb06990957.
* shorten comment
* add log messages
* format
* one more
* docs: Update Tailscale free tier user and device limits
* chore: generalize
Updated the description of the Tailscale free tier for clarity.
---------
Co-authored-by: Alex <alex.tran1502@gmail.com>
* fix(ml): handle empty/corrupt images in face detection
When a corrupt or degenerate image with zero-dimension (0 width or 0 height)
reaches the face detection pipeline, insightface's RetinaFace.detect() calls
cv2.resize() with a target size of 0, triggering an OpenCV assertion failure:
error: (-215:Assertion failed) inv_scale_x > 0 in function 'resize'
This crashes the ML worker and returns a 500 error to the server.
Add an early return in FaceDetector._predict() that checks for zero-dimension
images after decoding and returns empty detection results instead of passing
them to the insightface model.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(ml): move empty image validation to request level
Per review feedback, validate image dimensions in the predict endpoint
(returning 400) rather than in each model's _predict method. This
catches all zero-dimension images before they reach any model task.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(ml): resolve mypy strict type error in predict endpoint
Use intermediate `decoded` variable so mypy knows `.width` and `.height`
are accessed on `Image`, not on `Image | str`.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat(server): add OIDC logout URL override option
- Added toggle and field consistent with existing mobile redirect URI override.
- Existing auto-discovery remains default.
- Update tests and docs for new feature.
* fix(server): changes from review for OIDC logout URL override
- Rename 'logoutUri' to 'endSessionEndpoint'
- Remove toggle, just use override if provided
- Moved field in settings UI
* invalidate album data on album update to fix stale page load
* invalidate album data on album update to fix stale page load
* factor out callback, make async and await invalidate
* chore: formatting
---------
Co-authored-by: Jason Rasmussen <jason@rasm.me>
* feat(server): add configurable OAuth prompt parameter
Add a `prompt` field to the OAuth system config, allowing admins to
configure the OIDC `prompt` parameter (e.g. `select_account`, `login`,
`consent`). Defaults to empty string (no prompt sent), preserving
backward compatibility.
This is useful for providers like Google where users want to be prompted
to select an account when multiple accounts are signed in.
Discussed in #20762
* chore: regenerate OpenAPI spec and clients for OAuth prompt field
* Adding e2e test cases
* feat: web setting
* feat: docs
---------
Co-authored-by: Jason Rasmussen <jason@rasm.me>
* fix(oauth): normalize email claim to lowercase before account lookup and registration
* test(auth): add test for OAuth email case normalization
* chore: clean up
---------
Co-authored-by: Jason Rasmussen <jason@rasm.me>
* fix!: do not allow insecure oauth requests by default
* fix: format
* fix: make open-api
* fix: tests
* nit: casing
* chore: migration to allow insecure if current oauth uses http
Move `showEditFaces` state to `assetViewerManager` so the edit faces
panel open/close state is globally accessible. Add Escape key handling
to `PersonSidePanel` that closes the assign-face sub-panel first, then
the edit faces panel. Guard the asset viewer's global Escape-to-close
action so it doesn't fire while either face panel is open.
Change-Id: I0c947fe0758aef0eac473f4cc72f6a3b6a6a6964
AdaptiveImage stacks quality layers (thumbnail → preview → original) as they load. Without compositor layer promotion on the container, the browser could render a stale frame when the original-quality layer was overlaid on top of the preview-quality layer.
Add `will-change: transform` as a class on AdaptiveImage's root element so it gets a dedicated compositor layer from first paint. This also subsumes the imperative `style.willChange = 'transform'` that zoomImageAction was applying to the same element (the zoom target from photo-viewer is the AdaptiveImage root), so drop that now-redundant code.
Change-Id: Icd866a2bb5a5fce299c36404547fa0546a6a6964
fix(web): cancel video preview fetch when bind:this is cleared early
In Svelte 5.53.9, `bind:this` is now cleared earlier in the unmount
sequence ("better bind:this cleanup timing"). The video thumbnail's
$effect was relying on the old order to read the bound `player` element
and clear its `src` to abort the in-flight `/api/assets/{id}/video/playback`
range request — but the bind is now `undefined` by the time the effect
runs, so the cleanup is silently skipped. The detached <video> keeps
its src, and Firefox does not abort an in-flight media fetch when the
element is detached/GC'd. Long-lived 206 range requests then saturate
Firefox's 6-connection HTTP/1.1 per-host limit and freeze the timeline
(see #27585).
Capture the player reference inside the effect and tear down via the
effect cleanup return — Svelte runs the prior cleanup (with the captured
ref) before `bind:this` is cleared. Use the canonical
`pause() / removeAttribute('src') / load()` sequence which actually aborts
the fetch in Firefox, even on a detached element.
Fixes#27585
Change-Id: I4d9fba859955f5c54f603c345e61d4206a6a6964
* fix: add markFinished parameter to loadRequest and loadCodecRequest methods
* update loadRequest and loadCodecRequest methods to use isFinal
* Apply suggestions from code review
Co-authored-by: Mert <101130780+mertalev@users.noreply.github.com>
* remove redundant check
* fix: ensure isFinished is set correctly during cache eviction
* formatting
---------
Co-authored-by: Mert <101130780+mertalev@users.noreply.github.com>
* fix(server): increase restore health check timeout and reject with Error
* chore: clean up
---------
Co-authored-by: André Erasmus <25480506+NoBadDays@users.noreply.github.com>
Co-authored-by: Jason Rasmussen <jason@rasm.me>
When a concurrent caller awaits `this.complete` inside `execute()` and
`cancel()` is called, the promise rejects with `undefined` outside of any
try/catch, causing "Uncaught (in promise) undefined" console spam during
rapid timeline scrolling.
- Wrap the `await this.complete` path in try/catch, returning 'CANCELED'
- Guard the `finally` block to only null `cancelToken` if it still belongs
to this call, preventing a race condition with `cancel()` to `init()`
Change-Id: I65764dd664eb408433fc6e5fc2be4df56a6a6964
* Fix#26502: Fix timestamp handling for database backup in Web UI
Frontend parsed backup timestamps as UTC, but they were in the
server's local timezone, causing wrong relative times.
Add `timezone` field to DatabaseBackupDto to expose server timezone.
Update frontend to parse timestamps using this timezone.
Convert timestamps to user's local timezone before rendering.
Fallback to browser timezone if server timezone is missing.
Ensures correct relative time display in Web UI.
* fix: regenerate open-api types and remove custom backup type
- Ran `make open-api` to update types based on backend changes
- Removed custom BackupWithTimezone type in MaintenanceBackupsList
- Updated timezone props to use the newly generated native type
* fix: simplify timezone handling for database backups
- Updated DatabaseBackupDto to make timezone a required property
- Removed manual DateTime.local().zoneName fallbacks
- Cleaned up type casts after regenerating OpenAPI types
* fix: Add missing newline at end of spec file
We are generally looking to move away from hooks as they are hard to
reason about and have weird bugs. In particular, the timer did not
properly capture the ref of the callback, and so it would execute on old
state. A standard stateful widget does not have this problem, and is
easier to organise.
Co-authored-by: Alex <alex.tran1502@gmail.com>
* disable bottom safe area on trash bottom bar so that it extends below the system nav bar
* remove manual padding calculations
* re-add static vertical padding to maintain previous bottom bar height
* refactor: replace DispatchQueue + DispatchSemaphore with OperationQueue for image processing
* implement RequestRegistry and UnfairLock for managing cancellable requests
* implement requests registry for local and remote image processing
* remove Cancellable protocol and cancel method from request registry
* use mutex
---------
Co-authored-by: mertalev <101130780+mertalev@users.noreply.github.com>
The widget is not recreated correctly when videoPlayName changes, which
can cause weird behaviour with hooks and timers (like timers not firing
when they should do). Using a key forces flutter to recreate the widget
properly and allow the assumptions in build to work correctly.
* fix(mobile): don't update search filters in-place
Search filters are currently modified in-place, which can feel quite
janky. The chips behind the bottom sheet update instantly, and the
search page gets confused because filters have been applied but no
search has been initiated. Filters should keep their own copy of the
filter when they're opened, and the commit + search on apply.
The previous filter and pre-filter concepts were also cleaned up. They
added complexity, and `search()` now owns the full life cycle of the
filter.
This now also reverts the changes from #27155, as this solution should
be cleaner.
* refactor & color tweak
---------
Co-authored-by: Alex <alex.tran1502@gmail.com>
3 seconds is too long for some people, and it can be confusing to see
the video not playing and no indication that it's buffering. Reducing
the duration of the timer should show the spinner faster and prevent
confusion.
Co-authored-by: Alex <alex.tran1502@gmail.com>
fix(mobile): reset video controls hide timer when showing controls changes
The hide timer currently only resets when the status of the video
changes, but does not account for when the controls change. This means
that two things happen:
1. The hide timer does not reset when the controls become visible
again, the controls will stay visible forever or until the playback
status changes.
2. The hide timer will fire too quickly, and will hide the controls
much sooner than 5 seconds if the controls are hidden and then shown
again before the hide timer fires.
At current, the controls for videos are always hidden when opening an
asset from the timeline, and when swiping between assets. The latter is
actually quite annoying, so it would be better UX if video controls were
hidden when opening from the timeline like before, but visibility of the
controls was retained when swiping between assets.
These toasts can sometimes cover UI elements and make them impossible to
interact with until they are dismissed. Specifically, deleting an asset
will show a toast over the video controls and prevent seeking.
Rename MonthGroup class to TimelineMonth to better convey that it represents a single month within the timeline. Updates the file, class, and all references across 16 files.
Change-Id: Id50fd6d4b7d0e431571b67c0f81c0e316a6a6964
Rename DayGroup class to TimelineDay to better convey that it represents
a single day within the timeline. Updates the file, class, and all
references across 13 files.
Change-Id: I9faef9bad73cd5b11f40daaf5eb140dd6a6a6964
- Dim non-hovered person thumbnails to 40% opacity when any face is active
- Add ring highlight on the active person's thumbnail
- Add focus-visible outline styling for keyboard navigation
- Apply same treatment to both detail panel people section and edit faces side panel
Change-Id: I4ac10fe4568b95f3e0e8d9104133180f6a6a6964
Co-authored-by: Alex <alex.tran1502@gmail.com>
- Add subtle drop shadow to the asset viewer nav bar for better visual
separation from the image behind it
- Add drop shadow to the OCR text recognition button in the lower right
- Prevent nav bar action buttons from shrinking to nothing by adding
*:shrink-0 to the flex container, with p-1/-m-1 to avoid clipping
focus outlines
Change-Id: I61cdc0ec66a65cde1c95b40c2c5428006a6a6964
When hovering over a detected face in the photo viewer, an SVG mask overlay
dims the rest of the image (40% black) while leaving the hovered face fully
visible. The overlay fades in/out smoothly via CSS opacity transition by
freezing the last highlighted box positions in state, preventing the overlay
from popping off instantly when the mouse leaves.
Change-Id: I07e2eb2b297820ec89812785fe7943846a6a6964
* refactor listener tracking for image stream completers and fix early cancel call
* fix: improve cache listener identification in image stream tracking
* add documentation and test cases for listener tracking in ImageStreamCompleter
* fix: remove unnecessary image provision flag from listener tracking
* fix: override setImage method in cache aware listener tracker mixin
* fix: rename test file
The spacing was required for the old slider, but the new one has its own
spacing and makes it redundant. There is too much now, and we've
received feedback that it should be less sparse. The default track
height of 16px is an improvement over the old track height, but it is
very thick. A middleground of 12px might be better.
When downloading a live photo, Safari overwrites the image file with
the motion video because both share the same base filename. Append
'-motion' suffix to the video filename to prevent collision.
For example, IMG_1234.heic and IMG_1234.mov become IMG_1234.heic
and IMG_1234-motion.mov.
Fixes#23055
* feat(web): Synchronize information from deduplicated images
* Added new settings menu to the the deduplication tab.
* The toggable options in the settings are synchronization of: albums, favorites, ratings, description, visibility and location.
* When synchronizing the albums, the resolved images will be added to all albums of the duplicates.
* When synchronizing the favorite status, the resolved images will be marked as favorite, if at least one selectable image is marked as favorite.
* When synchronizing the ratings, the highest rating from the selectable images will be applied to the resolved image.
* When synchronizing the description, all descriptions from the selectable images will be merged into one description for the resolved image.
* When synchronizing the visibility, the most restrictive visibility setting from the selectable images will be applied to the resolved image.
* When synchronizing the location, if exactly one unique location exists among the selectable images, this location will be applied to the resolved image.
* There is no additional UI for these settings to keep the visual clutter minimal. The settings are applied automatically based on the user's preferences.
* Replace addAssetToAlbums with copyAsset
* fix linter
* feat(web): add duplicate sync fields and fix typo
* feat(web): add tag sync and enhance duplicate resolution
This update introduces tag synchronization for duplicate resolution,
ensuring all unique tag IDs from duplicates are applied to kept assets.
The visibility sync logic is updated to use a simplified ordering, as the hidden status items will never show up in a duplicate set.
Album synchronization now merges albums directly via addAssetsToAlbums; as the approach with copyAsset API endpoint was ineffiecient.
Description, rating, and location sync logic is improved for correctness.
and deduplication. i18n strings were added / updated.
* feat(server): move duplicate resolution to backend with sync and stacking
Moves duplicate metadata synchronization from frontend to backend, enabling robust
batch operations and proper validation. This is an improved refactor of PR #13851.
New endpoints:
- POST /duplicates/resolve - batch resolve with configurable metadata sync
- POST /duplicates/stack - create stacks from duplicate groups
- GET /duplicates - now includes suggestedKeepAssetIds based on file size and EXIF
Key changes:
- Move sync logic (albums, tags, favorites, ratings, descriptions, location, visibility) to server
- Add server-side metadata merge policies with proper conflict resolution
- Replace client-side resolution logic with new backend endpoints
- Add comprehensive E2E tests (70+ test cases) and unit tests
- Update OpenAPI specs and TypeScript SDK
No breaking changes - only additions to existing API.
* feat(preferences): enable all duplicate sync settings by default
* chore: clean up
* chore: clean up
* refactor: rename & clean up
* fix: preference upgrade
* chore: linting
* refactor(e2e): use updateAssets API for setAssetDuplicateId
* fix: visibility sync logic in duplicate resolution
* fix(duplicate): write description to exifUpdate
Previously the duplicate resolution populated assetUpdate.description even
though description belongs to exif info.
* fix(duplicate): remove redundant updateLockedColumns wrapper
updateAllExif already computes lockedProperties via distinctLocked
using Object.keys(options). The wrapper added a lockedProperties key
to the options object, causing the spurious string 'lockedProperties'
to be stored in the lockedProperties array.
* fix(duplicate): write merged tags to asset_exif to survive metadata re-extraction
During duplicate resolution, replaceAssetTags correctly wrote merged tag
IDs to the tag_asset table, but never updated asset_exif.tags or locked
the tags property. The subsequent SidecarWrite → AssetExtractMetadata
chain calls applyTagList, which destructively replaces tag_asset rows
with whatever is in asset_exif.tags — still the original per-asset tags,
not the merged set.
Write merged tag values to asset_exif.tags via updateAllExif (which also
locks the property via distinctLocked), and queue SidecarWrite when tags
change so they persist to the sidecar file.
* docs(duplicates): clarify location and tag sync behavior
* refactor(duplicate): remove sync settings, always sync all metadata on resolve
Remove DuplicateSyncSettingsDto and the per-field sync toggles
(albums, favorites, rating, description, visibility, location, tags).
Duplicate resolution now unconditionally syncs all metadata from
trashed assets to kept assets.
- Remove DuplicateSyncSettingsDto and settings field from DuplicateResolveDto
- Update DuplicateService to always run all sync logic without conditionals
- Delete DuplicateSettingsModal.svelte and settings gear button from UI
- Remove DuplicateSettings type and duplicateSettings persisted store
- Update unit and e2e tests to remove settings from resolve requests
* docs: update duplicates utility to reflect automatic metadata sync
* docs(web): replace duplicates info modal with link to documentation
* chore: clean up
* fix: add missing type cast to jsonAgg in duplicate repository getAll
* fix: skip persisting rating=0 in duplicate merge to avoid unnecessary sidecar write
---------
Co-authored-by: Toni <51962051+EinToni@users.noreply.github.com>
Co-authored-by: Jason Rasmussen <jason@rasm.me>
Co-authored-by: Jason Rasmussen <jrasm91@gmail.com>
* fix(server): prevent album shared link from breaking after uploads
* update test
* add withSharedAssets helper
* remove options
* add more helpers
* update selects
* fix(mobile): option padding on search dropdowns
* chore: prevent height fill up screen and block the bottom menu entry
---------
Co-authored-by: Alex <alex.tran1502@gmail.com>
* fix(web): allow pasting PIN code from clipboard or password manager
The keydown handler was blocking Ctrl+V/Cmd+V since it called
preventDefault() on all non-numeric keys. Also adds an onpaste
handler to distribute pasted digits across the individual inputs.
* refactor: handle paste in handleInput, remove maxlength
* cleanup + fix digit focus
---------
Co-authored-by: Preslav Penchev <preslav.penchev@acronis.com>
Co-authored-by: Michel Heusschen <59014050+michelheusschen@users.noreply.github.com>
Ensure that all files are flushed after they've been written.
At current, files are not explicitly flushed to disk, which can cause
data corruption. In extreme circumstances, it's possible that uploaded
files may not ever be persisted at all.
* feat(mobile): open in browser
* chore: open in browser instead of webview
* chore: allow archived asset
* fix: moved openinbrowser above unstack
* feat: deeplink into favorites, trash & archived
* fix: use remoteId (for tests to succeed)
---------
Co-authored-by: Alex <alex.tran1502@gmail.com>
- Precise font sizing using canvas measureText instead of character-count heuristic
- Fix overlay repositioning on viewport resize by computing metrics from reactive state instead of DOM reads
- Fix animation delay on resize by using transition-colors instead of transition-all
- Add keyboard accessibility: OCR boxes are focusable via Tab with reading-order sort
- Show text on focus (same styling as hover) with proper ARIA attributes
Final step on #22833
PReq #22833 is about adding support for SMTP-over-TLS rather than just STARTTLS when sending emails. That PReq adds almost everything; it just forgot to actually pass the flag to Nodemailer at the end.
This adds that last line of code and makes it work correctly (for me, anyways!).
Co-authored-by: Nathaniel <I@nathaniel.land>
* use cookiejar
* cookie duping hook
* remove old pref
* handle network switching on logout
* remove bootstrapCookies
* dead code
* fix cast
* use constants
* use new event name
* update api
* Add support for showing animated images in AssetViewer with AnimatedImageStreamCompleter
* Add GIF overlay to thumbnail tile for animated assets
* formatting
* require isAnimated parameter in image providers for better asset handling
* feat: refactor AnimatedImageStreamCompleter to use streams for codec loading and initial image handling
* formatting
* add isAnimatedImage property to BaseAsset
* remove ApiService.getRequestHeaders() usage
* feat(web): adaptive progressive image loading for photo viewer
Replace ImageManager with a new AdaptiveImageLoader that progressively
loads images through quality tiers (thumbnail → preview → original).
New components and utilities:
- AdaptiveImage: layered image renderer with thumbhash, thumbnail,
preview, and original layers with visibility managed by load state
- AdaptiveImageLoader: state machine driving the quality progression
with per-quality callbacks and error handling
- ImageLayer/Image: low-level image elements with load/error lifecycle
- PreloadManager: preloads adjacent assets for instant navigation
- AlphaBackground/DelayedLoadingSpinner: loading state UI
Zoom is handled via a derived CSS transform applied to the content
wrapper in AdaptiveImage, with the zoom library (zoomTarget: null)
only tracking state without manipulating the DOM directly.
Also adds scaleToCover to container-utils and getAssetUrls to utils.
* fix: don't partially render images in firefox
* add passive loading indicator to asset-viewer
---------
Co-authored-by: Alex <alex.tran1502@gmail.com>
* feat(mobile): use material design 3 slider
The new slider is easier to use, and looks more modern.
* chore: add shadow to button and text for better visibility
---------
Co-authored-by: Alex <alex.tran1502@gmail.com>
The numbers in the backup page are not monospace, and so changes cause
the layout to shift. Using tabular figures (monospace) will prevent
that.
Refs: #25021
Videos have recently been changed to support zooming, but this can make
the controls in the centre of the screen unergonomic as they will either
stay in the centre when dismissing, or stick to the video when zooming.
Neither is great. We should align the behaviour with other apps which
has the play/pause toggle at the bottom of the screen with the seeker
bar instead.
Co-authored-by: Alex <alex.tran1502@gmail.com>
* Enable OpenVINO CPU acceleration in Immich
* Remove unnecessary debug log
* Removing checking for device_ids for openvino since cpu will always be available
* Find OpenVINOExecutionProvider index instead of assuming index 0
* Fix openvino tests
* Fix failing test mock. OpenVINO expects provider options, but cuda provide doesn't so use that for mocked tests.
* Support empty provider options in OrtSessions in which case ONNXRuntime will use its own defaults
* Use OpenVINOExecutionProvider for test_sets_provider_options_kwarg
* fix mock
* simplify
* unused variable
---------
Co-authored-by: Aleksander <pejcic@adobe.com>
Co-authored-by: mertalev <101130780+mertalev@users.noreply.github.com>
* Fix opus handling as accepted audio codec in transcode policy
Fix the issue when opus is among accepted audio codecs in transcode policy
(which is default) but it still triggers transcoding because the codec name
from ffprobe (opus) does not match `libopus` literal in Immich.
Make a distinction between a codec name and encoder:
- codec name: switch to `opus` as the audio codec name. This matches what ffprobe
returns for a media file with opus audio.
- encoder: continue using the `libopus` encoder in ffmpeg.
* Add unit tests for accepted audio codecs and for libopus encoder
* Add db migration for ffmpeg.targetAudioCodec opus
* backward compatibility
* tweak
* noisy logs
* full mapping
* make check happy
* mark deprecated
* update api
* indexOf
---------
Co-authored-by: mertalev <101130780+mertalev@users.noreply.github.com>
* fix(mobile): correct local asset dimensions
We are constraining the size of videos so that they play nicely with
hero animations, and don't stretch in weird ways. This however caused a
regression as we are not account for local assets on Android which have
un-oriented dimensions.
* post-orientation width and height in local sync
* migration
* no need to handle it in asset viewer
---------
Co-authored-by: mertalev <101130780+mertalev@users.noreply.github.com>
Co-authored-by: Alex <alex.tran1502@gmail.com>
* feat(android): enhance playback style detection using MIME type
* feat(android): improve playback style detection for GIF and WebP formats
* fix(android): make playback style detection faster
* refactor(android): simplify XMP reading logic for API 29 and below
* update playback style detection documentation
* use DefaultImageHeaderParser instead of all available ones for webp playbackStyle type detection
Search results use a different provider than the main timeline, and they
appear appear to have diverged a bit. This means that assets can
sometimes look wrong or different in search compared to the main
timeline or albums.
The background of the photo view does not extend below the height of the
viewport, and so the asset details fade in over black with the photo
view, and the standard surface colour scheme of the scaffold for the
rest. This leads to a janky animation. We can't change the background of
the scaffold to black, as it in turn makes the iOS bouncing scroll
physics cut off incorrectly. The best fix is to remove background
decoration from the photo view, and defer to the parent to colour the
background.
Co-authored-by: Alex <alex.tran1502@gmail.com>
Search results are replaced with a spinner when loading the next page,
which is quite jarring. Search results now remain visible when loading
the next page with a spinner at the bottom. The next page also loads
sooner, which makes it feel a lot smoother.
Co-authored-by: Alex <alex.tran1502@gmail.com>
In order to scroll smoothly without interfering with the gesture
detector on the photo view, we have an offstate scroll view which we
defer all drags to, and then forward scroll offsets to the real scroll
controller. This works well, but it can be simpler. Instead, we can
create a custom scroll controller on a scroll view with never scrollable
physics, and then forward drag events to that, bypassing the need for a
proxy scroll controller.
Co-authored-by: Alex <alex.tran1502@gmail.com>
* feat(web): when hovering over a face already deteced, display the bounding box also shown when hovering over the person in the details-pane.
* prevent lint error
* fix unused var
Add `name` to all vitest configs matching CI job buckets (server:unit,
server:medium, cli:unit, web:unit, e2e:server, e2e:maintenance) so they
appear as filterable @tags in the Vitest VSCode extension.
Fix `root` in server vitest configs to use an absolute path derived from
`import.meta.url` instead of `'./'`, which resolved relative to the config
file directory (`server/test/`) rather than `server/`, causing test
discovery to fail in the Vitest VSCode extension.
Consolidate video state into a single asset-scoped provider, and reduce
dependency on global state generally. Overall this should fix a few
timing issues and race conditions with videos specifically, and make
future changes in this area easier.
The image in the photo view has no height, and is therefore entirely
unconstrained. This causes the image to take up the full height of the
viewport during the hero animation, which can make look out of sync. In
some other cases, it can stretch or resize the image to fill the entire
viewport.
* fix(android): detect supported version for special format column
* fix(android): remove unnecessary suppression for new API in special format check
* fix(android): change visibility of hasSpecialFormatColumn method to private
* feat: add playbackStyle to local asset entity and related database schema
* implement conversion function for playbackStyle in local sync service
* implement conversion function for playbackStyle in local sync service
* refactor: remove deducedPlaybackStyle from TrashedLocalAssetEntityData
* add playbackStyle column to trashed local asset entity
* make playbackStyle non-nullable across the mobile codebase
* Streamline playbackStyle backfill:
- only backfill local assets playbackStyle in flutter/dart code
- only update trashed local assets in db migration
* bump target database version to 23 and update migration logic for playbackStyle
* set playback_style to 0 in merged_asset.drift as its a getter in base asset
* run make pigeon
* Populate playbackStyle for trashed assets during native migration
* feat(mobile): add support for encoded image requests in local and remote image APIs
* fix(mobile): handle memory cleanup for cancelled image requests
* refactor(mobile): simplify memory management and response handling for encoded image requests
* fix(mobile): correct formatting in cancellation check for image requests
* Apply suggestion from @mertalev
Co-authored-by: Mert <101130780+mertalev@users.noreply.github.com>
* refactor(mobile): rename 'encoded' parameter to 'preferEncoded' for clarity in image request APIs
* fix(mobile): ensure proper resource cleanup for cancelled image requests
* refactor(mobile): streamline codec handling by removing unnecessary descriptor disposal in loadCodec request
---------
Co-authored-by: Mert <101130780+mertalev@users.noreply.github.com>
* feat(mobile): add playbackStyle to native sync API
Adds a `playbackStyle` field to `PlatformAsset` in the pigeon sync API so
native platforms can communicate the asset's playback style (image, video,
animated, livePhoto) to Flutter during sync.
- Add `playbackStyleValue` computed property to `PHAsset` extension (iOS)
- Populate `playbackStyle` in `toPlatformAsset()` and the full-sync path
- Update generated Dart/Kotlin/Swift files
* fix(tests): add playbackStyle to local asset test cases
* fix(tests): update playbackStyle to use integer values in local sync tests
* feat(mobile): extend playbackStyle enum to include videoLooping
* Update PHAssetExtensions.swift
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* fix(playback): simplify playbackStyleValue implementation by removing iOS version check
* feat(android): implement proper playbackStyle detection
* add PlatformAssetPlaybackStyle enum
* linting
---------
Co-authored-by: Mert <101130780+mertalev@users.noreply.github.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
The changes in #25952 inadvertently removed an optimisation which
prevents the video player from being recreated when the tree changed.
This happens surprisingly often, namely when the hero animation
finishes. The widget is particularly expensive, so recreating it 2-3 in
a short period not only feels sluggish, but also causes the video to
hitch and restart.
The solution is to bring the global key back for the native video
player. Unlike before, we are using a custom global key which compares
the values of hero tags directly. This means we don't need to maintain a
map of hero tags to global keys in the state, and also means we don't
have to pass the global key down multiple layers.
This also fixes#25981.
Asset details are prematurely hidden when a drag ends if the simulation
shows that it will close given its current velocity. It makes for a much
more responsible feeling UI. However, this behaviour conflicts with the
logic which determines whether details are showing based on the current
offset. The result is that the details are hidden, then immediately
shown again, and then hidden once it passes the min snap distance
threshold.
This can be fixed by only evaluating the position based logic when a
drag is active, and then inferring upcoming state with a simulation.
-f body="This PR has been automatically closed as the description doesn't follow [our template](https://github.com/immich-app/immich/blob/main/.github/pull_request_template.md). After you edit it to match the template, the PR will automatically be reopened." \
-f body="Thank you for your interest in contributing to Immich! Unfortunately this PR looks like it was generated using an LLM. As noted in our [CONTRIBUTING.md](https://github.com/immich-app/immich/blob/main/CONTRIBUTING.md#use-of-generative-ai), we request that you don't use LLMs to generate PRs as those are not a good use of maintainer time." \
Installs as a separate app (applicationId `app.alextran.immich.pr${{ github.event.pull_request.number }}`), so it coexists with the Play Store version and any other PR builds.
-f body="Thank you for your interest in contributing to Immich! Unfortunately this PR looks like it was generated using an LLM. As noted in our [CONTRIBUTING.md](https://github.com/immich-app/immich/blob/main/CONTRIBUTING.md#use-of-generative-ai), we request that you don't use LLMs to generate PRs as those are not a good use of maintainer time." \
@@ -15,6 +15,8 @@ Please try to keep pull requests as focused as possible. A PR should do exactly
If you are looking for something to work on, there are discussions and issues with a `good-first-issue` label on them. These are always a good starting point. If none of them sound interesting or fit your skill set, feel free to reach out on our Discord. We're happy to help you find something to work on!
We usually do not assign issues to new contributors, since it happens often that a PR is never even opened. Again, reach out on Discord if you fear putting a lot of time into fixing an issue, but ending up with a duplicate PR.
## Use of generative AI
We ask you not to open PRs generated with an LLM. We find that code generated like this tends to need a large amount of back-and-forth, which is a very inefficient use of our time. If we want LLM-generated code, it's much faster for us to use an LLM ourselves than to go through an intermediary via a pull request.
@@ -210,7 +210,7 @@ The provided restore process ensures your database is never in a broken state by
## Filesystem
Immich stores two types of content in the filesystem: (a) original, unmodified assets (photos and videos), and (b) generated content. We recommend backing up the entire contents of `UPLOAD_LOCATION`, but only the original content is critical, which is stored in the following folders:
Immich does not handle filesystem backups for you. You have to arrange these yourself! Immich stores two types of content in the filesystem: (a) original, unmodified assets (photos and videos), and (b) generated content. We recommend backing up the entire contents of `UPLOAD_LOCATION`, but only the original content is critical, which is stored in the following folders:
@@ -49,6 +50,10 @@ Before enabling OAuth in Immich, a new client application needs to be configured
-`https://immich.example.com/auth/login`
-`https://immich.example.com/user-settings`
3. Configure Backchannel logout URL
If the authentication server supports it, the **Backchannel logout URL** can be specified, and it is of the form: `http://DOMAIN:PORT/api/oauth/backchannel-logout`.
## Enable OAuth
Once you have a new OAuth client application configured, Immich can be configured using the Administration Settings page, available on the web (Administration -> Settings).
@@ -62,6 +67,8 @@ Once you have a new OAuth client application configured, Immich can be configure
| `scope` | string | openid email profile | Full list of scopes to send with the request (space delimited) |
| `id_token_signed_response_alg` | string | RS256 | The algorithm used to sign the id token (examples: RS256, HS256) |
| `userinfo_signed_response_alg` | string | none | The algorithm used to sign the userinfo response (examples: RS256, HS256) |
@@ -81,7 +81,7 @@ VectorChord is the successor extension to pgvecto.rs, allowing for higher perfor
### Migrating from pgvecto.rs
Support for pgvecto.rs will be dropped in a later release, hence we recommend all users currently using pgvecto.rs to migrate to VectorChord at their convenience. There are two primary approaches to do so.
Support for pgvecto.rs has been dropped as of 3.0, hence all users currently using pgvecto.rs should migrate to VectorChord. There are two primary approaches to do so.
The easiest option is to have both extensions installed during the migration:
Immich uses the [OpenAPI](https://swagger.io/specification/) standard to generate API documentation. To view the published docs see [here](https://api.immich.app/).
@@ -10,4 +10,4 @@ OpenAPI is used to generate the client (Typescript, Dart) SDK. `openapi-generato
make open-api
```
You can find the generated client SDK in the `open-api/typescript-sdk/client` for Typescript SDK and `mobile/openapi` for Dart SDK.
You can find the generated client SDK in the `packages/sdk/client` for Typescript SDK and `mobile/openapi` for Dart SDK.
All three clients use [OpenAPI](./open-api.md) to auto-generate rest clients for easy integration. For more information about this process, see [OpenAPI](./open-api.md).
All three clients use [OpenAPI](/api.md) to auto-generate rest clients for easy integration. For more information about this process, see [OpenAPI](/api.md).
:::
### Mobile App
@@ -71,7 +71,7 @@ An incoming HTTP request is mapped to a controller (`src/controllers`). Controll
### Domain Transfer Objects (DTOs)
The server uses [Domain Transfer Objects](https://en.wikipedia.org/wiki/Data_transfer_object) as public interfaces for the inputs (query, params, and body) and outputs (response) for each endpoint. DTOs translate to [OpenAPI](./open-api.md) schemas and control the generated code used by each client.
The server uses [Domain Transfer Objects](https://en.wikipedia.org/wiki/Data_transfer_object) as public interfaces for the inputs (query, params, and body) and outputs (response) for each endpoint. DTOs translate to [OpenAPI](/api.md) schemas and control the generated code used by each client.
@@ -34,26 +34,28 @@ Run all web checks with `pnpm run check:all`
Run all server checks with `pnpm run check:all`
:::
:::info Auto Fix
:::tip Auto Fix
You can use `pnpm run __:fix` to potentially correct some issues automatically for `pnpm run format` and `lint`.
:::
## Mobile Checks
## Mobile Checklist
The following commands must be executed from within the mobile app directory of the codebase.
- [ ]`mise //mobile:codegen` (auto-generate files using build_runner)
- [ ]`mise //mobile:lint` (static analysis via Dart Analyzer and DCM)
- [ ]`mise //mobile:format` (formatting via Dart Formatter)
- [ ]`mise //mobile:test` (unit tests)
- [ ]`make build` (auto-generate files using build_runner)
- [ ]`make analyze` (static analysis via Dart Analyzer and DCM)
- [ ]`make format` (formatting via Dart Formatter)
- [ ]`make test` (unit tests)
:::tip
Run all these commands at once with `mise //mobile:checklist`
:::
:::info Auto Fix
You can use `dart fix --apply` and `dcm fix lib` to potentially correct some issues automatically for `make analyze`.
:::tip Auto Fix
You can use `mise //mobile:lint-fix` to potentially correct some issues automatically for `mise //mobile:lint`.
:::
## OpenAPI
The OpenAPI client libraries need to be regenerated whenever there are changes to the `immich-openapi-specs.json` file. Note that you should not modify this file directly as it is auto-generated. See [OpenAPI](/developer/open-api.md) for more details.
The OpenAPI client libraries need to be regenerated whenever there are changes to the `immich-openapi-specs.json` file. Note that you should not modify this file directly as it is auto-generated. See [OpenAPI](/api.md) for more details.
@@ -80,9 +80,9 @@ To see local changes to `@immich/ui` in Immich, do the following:
1. Install `@immich/ui` as a sibling to `immich/`, for example `/home/user/immich` and `/home/user/ui`
2. Build the `@immich/ui` project via `pnpm run build`
3. Uncomment the corresponding volume in web service of the `docker/docker-compose.dev.yaml` file (`../../ui:/usr/ui`)
4. Uncomment the corresponding alias in the `web/vite.config.js` file (`'@immich/ui': path.resolve(\_\_dirname, '../../ui')`)
5. Uncomment the import statement in `web/src/app.css` file `@import '/usr/ui/dist/theme/default.css';` and comment out `@import '@immich/ui/theme/default.css';`
3. Uncomment the corresponding volume in web service of the `docker/docker-compose.dev.yml` file (`../../ui:/usr/src/ui`)
4. Uncomment the corresponding alias in the `web/vite.config.ts` file (`'@immich/ui': path.resolve(\_\_dirname, '../../ui/packages/ui')`)
5. Uncomment the import statement in `web/src/app.css` file `@import '../../../ui/packages/ui/dist/theme/default.css';` and comment out `@import '@immich/ui/theme/default.css';`
6. Start up the stack via `make dev`
7. After making changes in `@immich/ui`, rebuild it (`pnpm run build`)
Immich comes with a duplicates utility to help you detect assets that look visually similar. The duplicate detection feature relies on machine learning and is enabled by default. For more information about when the duplicate detection job runs, see [Jobs and Workers](/administration/jobs-workers). Once an asset has been processed and added to a duplicate group, it becomes available to review in the "Review duplicates" utility, which can be found [here](https://my.immich.app/utilities/duplicates).
## Reviewing duplicates
The review duplicates page allows the user to individually select which assets should be kept and which ones should be trashed. When more than one asset is kept, there is an option to automatically put the kept assets into a stack.
### Automatic preselection
When using "Deduplicate All" or viewing suggestions, Immich automatically preselects which assets to keep based on:
1.**Image size in bytes** — larger files are preferred as they typically have higher quality.
2.**Count of EXIF data** — assets with more metadata are preferred.
### Synchronizing metadata
When resolving duplicates, metadata from trashed assets is automatically synchronized to the kept assets. The following metadata is synchronized:
-`**/Raw/**` will exclude all files in any directory named `Raw`
-`**/*.{tif,jpg}` will exclude all files with the extension `.tif` or `.jpg`
Note that `*` is a wildcard matching zero or more characters (i.e., withinin a filename or single directory name). `**` matches zero or more subdirectories, recursively. It also includes any/all files within a subdirectory, i.e., when used at the end of a pattern. For example, `**/exclude_me/**` will exclude all files in any directory named `exclude_me`, as well as all files in any subdirectories of `exclude_me`, recursively.
Special characters such as @ should be escaped, for instance:
-`**/\@eaDir/**` will exclude all files in any directory named `@eaDir`
@@ -47,9 +47,11 @@ You do not need to redo any machine learning jobs after enabling hardware accele
#### ROCm
- On Linux, The [AMDGPU driver module](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/how-to/docker.html) needs to be installed on the server and, if secure boot is used, the signing key of DKMS [needs to be enrolled in UEFI BIOS](https://wiki.debian.org/SecureBoot)
- The GPU must be supported by ROCm. If it isn't officially supported, you can attempt to use the `HSA_OVERRIDE_GFX_VERSION` environmental variable: `HSA_OVERRIDE_GFX_VERSION=<a supported version, e.g. 10.3.0>`. If this doesn't work, you might need to also set `HSA_USE_SVM=0`.
- The ROCm image is quite large and requires at least 35GiB of free disk space. However, pulling later updates to the service through Docker will generally only amount to a few hundred megabytes as the rest will be cached.
- This backend is new and may experience some issues. For example, GPU power consumption can be higher than usual after running inference, even if the machine learning service is idle. In this case, it will only go back to normal after being idle for 5 minutes (configurable with the [MACHINE_LEARNING_MODEL_TTL](/install/environment-variables) setting).
- MIGraphX is a new backend for AMD cards, which compiles models at runtime. As such, the first few inferences will be slow.
@@ -39,7 +39,7 @@ You can learn how to set up Tailscale together with Immich with the [tutorial vi
### Cons
- The Tailscale client usually needs to run as root on your devices and it increases the attack surface slightly compared to a minimal Wireguard server. e.g., an [RCE vulnerability](https://github.com/tailscale/tailscale/security/advisories/GHSA-vqp6-rc3h-83cp) was discovered in the Windows Tailscale client in November 2022.
- Tailscale is a paid service. However, there is a generous [free tier](https://tailscale.com/pricing/) that permits up to 3 users and up to 100 devices.
- Tailscale is a paid service. However, there is a generous [free tier](https://tailscale.com/pricing/) suitable for personal use.
- Tailscale needs to be installed and running on both server-side and client-side.
| `IMMICH_LOG_FORMAT` | Log output format (`console`, `json`) | `console` | server | api, microservices |
| `IMMICH_MEDIA_LOCATION` | Media location inside the container ⚠️**You probably shouldn't set this**<sup>\*2</sup>⚠️ | `/data` | server | api, microservices |
| `IMMICH_CONFIG_FILE` | Path to config file | | server | api, microservices |
| `IMMICH_HELMET_FILE` | Path to a json file with [helmet](https://www.npmjs.com/package/helmet) options. Set to `false` to disable. Set to `true` to use `server/helmet.json`<sup>\*3</sup>. | `false` | server | api |
| `NO_COLOR` | Set to `true` to disable color-coded log output | `false` | server, machine learning | |
| `CPU_CORES` | Number of cores available to the Immich server | auto-detected CPU core count | server | |
| `IMMICH_API_METRICS_PORT` | Port for the OTEL metrics | `8081` | server | api |
| `IMMICH_MICROSERVICES_METRICS_PORT` | Port for the OTEL metrics | `8082` | server | microservices |
| `IMMICH_PROCESS_INVALID_IMAGES` | When `true`, generate thumbnails for invalid images | | server | microservices |
| `IMMICH_TRUSTED_PROXIES` | List of comma-separated IPs set as trusted proxies | | server | api |
| `IMMICH_IGNORE_MOUNT_CHECK_ERRORS` | See [System Integrity](/administration/system-integrity) | | server | api, microservices |
| `IMMICH_ALLOW_SETUP` | When `false` disables the `/auth/admin-sign-up` endpoint | `true` | server | api |
\*1: `TZ` should be set to a `TZ identifier` from [this list][tz-list]. For example, `TZ="Etc/UTC"`.
`TZ` is used by `exiftool` as a fallback in case the timezone cannot be determined from the image metadata. It is also used for logfile timestamps and cron job execution.
\*2: This path is where the Immich code looks for the files, which is internal to the docker container. Setting it to a path on your host will certainly break things, you should use the `UPLOAD_LOCATION` variable instead.
\*3: The [default configuration](https://helmetjs.github.io/#content-security-policy) sets `upgrade-insecure-requests`, which tells the browser to upgrade all requests to HTTPS. This breaks on HTTP-only deployments. If you cannot use HTTPS, you should use a custom helmet config file with `"upgrade-insecure-requests": null`.
## Workers
| Variable | Description | Default | Containers |
@@ -80,7 +83,7 @@ Information on the current workers can be found [here](/administration/jobs-work
| `DB_VECTOR_EXTENSION`<sup>\*2</sup> | Database vector extension (one of [`vectorchord`, `pgvector`, `pgvecto.rs`]) | | server |
| `DB_VECTOR_EXTENSION`<sup>\*2</sup> | Database vector extension (one of [`vectorchord`, `pgvector`]) | | server |
| `DB_SKIP_MIGRATIONS` | Whether to skip running migrations on startup (one of [`true`, `false`]) | `false` | server |
| `DB_STORAGE_TYPE` | Optimize concurrent IO on SSDs or sequential IO on HDDs ([`SSD`, `HDD`])<sup>\*3</sup> | `SSD` | database |
@@ -166,6 +169,8 @@ Redis (Sentinel) URL example JSON before encoding:
| `MACHINE_LEARNING_PRELOAD__CLIP__VISUAL` | Comma-separated list of (visual) CLIP model(s) to preload and cache | | machine learning |
| `MACHINE_LEARNING_PRELOAD__FACIAL_RECOGNITION__RECOGNITION` | Comma-separated list of (recognition) facial recognition model(s) to preload and cache | | machine learning |
| `MACHINE_LEARNING_PRELOAD__FACIAL_RECOGNITION__DETECTION` | Comma-separated list of (detection) facial recognition model(s) to preload and cache | | machine learning |
| `MACHINE_LEARNING_PRELOAD__OCR__RECOGNITION` | Comma-separated list of (recognition) OCR model(s) to preload and cache | | machine learning |
| `MACHINE_LEARNING_PRELOAD__OCR__DETECTION` | Comma-separated list of (detection) OCR model(s) to preload and cache | | machine learning |
@@ -8,7 +8,7 @@ Hardware and software requirements for Immich:
## Hardware
- **OS**: Recommended Linux or \*nix operating system (Ubuntu, Debian, etc).
- **OS**: Recommended Linux or \*nix 64-bit operating system (Ubuntu, Debian, etc).
- Non-Linux OSes tend to provide a poor Docker experience and are strongly discouraged.
Our ability to assist with setup or troubleshooting on non-Linux OSes will be severely reduced.
If you still want to try to use a non-Linux OS, you can set it up as follows:
@@ -19,6 +19,10 @@ Hardware and software requirements for Immich:
If you have issues, we recommend that you switch to a supported VM deployment.
- **RAM**: Minimum 6GB, recommended 8GB.
- **CPU**: Minimum 2 cores, recommended 4 cores.
- Immich runs on the `amd64` and `arm64` platforms.
Since `v2.6`, the machine learning container on `amd64` requires the `>= x86-64-v2` [microarchitecture level](https://en.wikipedia.org/wiki/X86-64#Microarchitecture_levels).
Most CPUs released since ~2012 support this microarchitecture.
If you are using a virtual machine, ensure you have selected a [supported microarchitecture](https://pve.proxmox.com/pve-docs/chapter-qm.html#_qemu_cpu_types).
- **Storage**: Recommended Unix-compatible filesystem (EXT4, ZFS, APFS, etc.) with support for user/group ownership and permissions.
- The generation of thumbnails and transcoded video can increase the size of the photo library by 10-20% on average.
@@ -45,7 +49,7 @@ Immich requires [**Docker**](https://docs.docker.com/get-started/get-docker/) wi
The Compose plugin will be installed by both Docker Engine and Desktop by following the linked installation guides; it can also be [separately installed](https://docs.docker.com/compose/install/).
:::note
Immich requires the command `docker compose`; the similarly named `docker-compose` is [deprecated](https://docs.docker.com/compose/migrate/) and is no longer supported by Immich.
Immich requires the command `docker compose`; the similarly named `docker-compose` is [deprecated](https://docs.docker.com/retired/#docker-compose-v1-replaced-by-compose-v2) and is no longer supported by Immich.
@@ -130,7 +130,3 @@ These storage mediums have different performance characteristics. As a result, t
#### Can I use the new database image as a general PostgreSQL image outside of Immich?
It’s a standard PostgreSQL container image that additionally contains the VectorChord, pgvector, and (optionally) pgvecto.rs extensions. If you were using the previous pgvecto.rs image for other purposes, you can similarly do so with this image.
#### If pgvecto.rs and pgvector still work, why should I switch to VectorChord?
VectorChord is faster, more stable, uses less RAM, and (with the settings Immich uses) offers higher-quality results than pgvector and pgvecto.rs. This translates to better search and facial recognition experiences. In addition, pgvecto.rs support will be dropped in the future, so changing it sooner will avoid disruption.
@@ -6,6 +6,8 @@ You can read more about the differences between storage template engine on and o
The admin user can set the template by using the template builder in the `Administration -> Settings -> Storage Template`. Immich provides a set of variables that you can use in constructing the template, along with additional custom text. If the template produces [multiple files with the same filename, they won't be overwritten](https://github.com/immich-app/immich/discussions/3324) as a sequence number is appended to the filename.
Date and time variables in storage templates are rendered in the server's local timezone.
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.