Immich v3.3.0 and the shift in self-hosted photo team decisions

The Immich v3.3.0 Release

Immich v3.3.0 introduces people sharing and machine learning optimizations that change the requirements for self-hosted photo teams. This update allows users to share people within a cluster group, which causes the shared person to appear in both users’ lists on mobile and web interfaces. Users can edit the name and birth date for a shared person, and these changes apply to everyone in the group by default. Web users can select the "Only me" option in sharing settings to keep name or date of birth changes local. The release also includes auto-stacked mobile edits, which places phone-based modifications on top of the original files to prevent timeline clutter. Machine learning performance improvements provide faster speeds and lower memory usage for users who set the MACHINE_LEARNING_MODEL_REVISION variable to v2. This change specifically utilizes TensorRT for NVIDIA Ampere and newer GPUs to increase efficiency. The PP-OCRv6_tiny model is efficient for modest hardware, while PP-OCRv6_medium improves on the previous PP-OCRv6_small as the highest quality option.

Resource Demands and Deployment

Hardware requirements create a massive divide between Immich and PhotoPrism. Immich requires at least 6 GB of RAM to support its machine learning container and settles at 1.9 GB of idle usage before a user imports a single photo into the system. Most users need 8 GB of RAM to handle the heavy machine learning worker during large library imports. PhotoPrism provides a much lighter footprint with a 2 GB minimum requirement and an idle usage of only 171 MB. Because PhotoPrism uses a compiled Go binary and MariaDB, it functions easily on a small VPS where Immich cannot operate without significant headroom. PhotoPrism focuses on indexing existing folders, whereas Immich focuses on the continuous ingestion of new mobile data.

The complexity of deployment also differs between the two. Immich relies on a multi-container stack including a server, machine-learning, Redis, and Postgres. This setup earns it a self-hosting difficulty rating of 3 out of 5. PhotoPrism requires a single main container plus a database, which leads to a lower difficulty rating of 2 out of 5. Users who manage large libraries using Google Takeout have distinct paths for both. Immich recommends the immich-go CLI tool to read JSON sidecar files for capture dates and GPS locations. PhotoPrism reads those same JSON sidecars natively, though users must manually check dates because Takeout provides raw UNIX timestamps without timezones.

The Decision

Teams choosing between PhotoPrism and LibrePhotos must prioritize either continuous phone synchronization or the organization of large, pre-existing image archives. Immich provides a polished mobile app with reliable background upload, a feature that PhotoPrism lacks since it relies on a progressive web app or WebDAV sync. PhotoPrism treats files as the source of truth and works well for users who prefer a stable, slow-moving tool for managing decades of camera dumps. LibrePhotos remains an option for those looking at the OwnPhotos fork. PhotoPrism suits users who manage a NAS share or an old Lightroom export and want a low-maintenance tool.

Users who look at alternative privacy models may prefer Ente, which offers end-to-end encryption where the server cannot read the plaintext photos. Ente requires an S3-compatible object store like Backblaze B2, Wasabi, Cloudflare R2, or AWS instead of a plain folder. Because encryption and machine learning happen on the user’s device, the Ente server is the lightest to run and idles in the low hundreds of MB. However, Ente is a newer, less-documented project compared to the maturity of Immich or PhotoPrism. If your team needs to validate candidate photo reuse or investigate impersonation online, you might look at PimEyes, which matches uploaded faces against public images. PimEyes matches uploaded faces against its index using facial geometry rather than pixel-matching. This approach ensures cropped or re-edited photos still surface matches.

Security analysts who look for automated facial recognition may prefer tools like Paravision or Face++. Paravision focuses on camera-driven entry decisions with an API that links identities to schedules and event handling. It provides decision traceability so staff can review why a person received or was refused access. Face++ builds around programmatic matching and search workflows that integrate into back ends through documented APIs. It includes liveness detection support to reduce the risk of simple presentation attacks during onboarding. How do teams manage the frequent update cycles of fast-moving projects like Immich without risking data loss? Immich is the recommendation for teams that need a Google Photos-style experience with automatic background backup from various mobile devices, which you likely already prioritize for your household.