Immich 3.3.0 is out, released on October 7, 2026. The team posted the full changelog on the official blog and on GitHub, and this one carries real feature work rather than just bug fixes.
What’s new in Immich 3.3.0
The headline addition is shared people management. If you have set up a group in Immich so a partner or family member has their own login, recognized faces can now be shared between accounts. A person tagged as “Mom” or “Dad” in your library can show up in someone else’s people list too, and both accounts can edit that person independently. Sharing direction matters here: you choose who sees which people, and merging duplicate entries across users is supported, which should cut down on the situation where the same face ends up tagged twice under two different names.
Birthday memories are the second big addition. Immich already builds “on this day” style memories, and it will now also surface a memory when a recognized person’s birthday comes around, pulling in photos from previous years. It is a small feature, but anyone who has used Google Photos will recognize the gap it fills for a self-hosted Google Photos alternative like Immich.
Mobile users get a quieter but useful change: edits made in other apps on your phone, such as a crop from a photo editor or a filter from another app, now get auto-stacked on top of the original photo in your timeline instead of showing up as a separate, unrelated image. That keeps the timeline from filling up with near-duplicates.
Smaller changes worth knowing about
- OAuth claims (storage quota, user role, preferred name) now sync on every login, not just at account creation, so changes made through an identity provider apply without recreating accounts.
- Thumbnails and previews use a more accurate resampling method for detail and brightness. This applies going forward only; existing thumbnails stay as they are unless you regenerate them manually.
- New OCR models from the PP-OCRv6 family (tiny, small, medium) for recognizing text in photos.
- A macOS storage-reporting bug is fixed, so disk space numbers, including inside Docker, now reflect actual usage.
Machine learning improvements, and how to turn them on
This release also brings faster inference and lower memory use for machine learning, plus two hardware-specific gains: RKNPU now supports every model Immich ships, and TensorRT RTX is used for NVIDIA GPUs from the Ampere generation (RTX 30-series) onward.
None of this is on by default. To enable it, set MACHINE_LEARNING_MODEL_REVISION=v2 on your machine learning container. The models were re-exported rather than retrained, so outputs are numerically equivalent to what you had before, meaning there is no need to reprocess your library after switching. The only catch is that the first run after enabling it takes longer while the new models get cached. The release notes say this revision becomes the default in a future version, so this is early access for anyone who wants it now.
Beyond that, there is a stack of smaller fixes: a timezone bug that could show the wrong birth date, machine learning errors on Intel Alchemist and Battlemage GPUs, mobile uploads that were re-sending files unnecessarily, a server crash tied to connection resets, and hidden files no longer showing up in “large files” search results.
No breaking changes
Immich lists no breaking changes for 3.3.0, so a normal update should be straightforward: docker compose pull, then docker compose up -d, as usual. If you want the machine learning speedups, that is the one extra step: add the MACHINE_LEARNING_MODEL_REVISION=v2 variable to your machine learning service before restarting it.
For self-hosters running Immich as their Google Photos replacement, this is a solid update. Shared people management is the kind of multi-user feature that has been missing, and the machine learning changes matter if you are running face and object recognition on your own hardware instead of offloading it to the cloud. Full details are in the official release notes and the GitHub release page.