Immich 3.3 adds people management and TensorRT acceleration for its AI models
The first release candidate of Immich 3.3 introduces people management, birthday memories, and TensorRT inference on NVIDIA RTX GPUs. If you self-host your photos, plan the migration and test the RC on a copy before the stable release lands.
September 29, 2026. The Immich team publishes v3.3.0-rc.0, the first release candidate of version 3.3 of the self-hosted photo manager. September 28, 2026. Version v3.2.4 had just patched a memory leak introduced by a dependency, promising a v3.3 that was “right around the corner.” September 2026. The RC ships two headline items: people management and TensorRT inference on NVIDIA RTX GPUs. Why it matters: Immich, at 115,300 GitHub stars, is closing the last two functional gaps separating it from Google Photos — face organization and the speed of its AI models.
Immich’s weight in self-hosting
Immich is not a niche project. At 115,300 GitHub stars, it has become one of the most popular self-hosted programs in the world, carried by an active community and backed by FUTO, the organization that has funded its development since 2024. Its positioning is clear: reproduce the Google Photos experience — automatic mobile backup, search, memories, face recognition — without a single photo leaving your server.
That success rests on a fast release cadence. v3.0, in July 2026, brought mobile editing and workflows; v3.2, in September, reworked search with Search v2; and v3.3 now tackles people management and inference speed. Each iteration closes the gap with the Mountain View giant on one specific point, and each point closed is one more reason to leave the cloud.
People management, finally native
Since its early days, Immich has detected faces and grouped them into “people” through its recognition models. What was missing was the ability to administer those people properly: rename, merge, hide, or reassign a misclassified face without hopping between several screens. v3.3.0-rc.0 introduces a dedicated people management feature, led by Daniel Dietzler, that gathers these operations in one place.
This is a deeper ergonomic change than it looks. On a library of tens of thousands of photos, a classification mistake — two merged faces, a stranger mislabeled — used to be fixed case by case. The new management aligns Immich with how families actually use it: a few minutes of face cleanup instead of a settings tunnel. The accompanying fix, which prevents moving the faces of a user other than their owner, shows the team designed the feature for shared accounts, not just the primary one.
Memories that know a birthday
The second visible addition is birthday memories. Immich already generates “memories” from photos taken on the same day in previous years; the RC adds a category dedicated to birthdays, fed by the birth dates attached to people. In practice, the app can resurface, every year, a loved one’s photos on their birthday — the kind of automated magic Google Photos has offered for years and that keeps users loyal.
That detail is not trivial. It illustrates Immich’s strategy: compete with the cloud giants on their own turf — serendipity and automation — while keeping the data on your own server. Every feature like this erodes the “I stick with Google Photos for the memories” argument and strengthens self-hosting as a complete alternative, not a consolation prize.
TensorRT: faster AI on RTX GPUs
Under the hood, the most technical contribution is trt-rtx, by Mert Alev. It enables TensorRT acceleration for machine learning model inference on NVIDIA RTX GPUs. Concretely: face recognition, object detection, and the other AI tasks run locally become markedly faster on machines with an RTX GPU, without hammering the CPU.
The RC also ships two inference optimizations — improved model loading and optimized model inference — plus linear-light resampling. That last point, subtler, improves the color fidelity of image transforms. For family use, the AI benefits most from the change: the initial indexing of a large library, a long operation, is precisely where TensorRT acceleration pays off immediately.
The RC finally adds a few foundation-level improvements useful in enterprise deployments: OAuth claim synchronization at login, which keeps roles and groups from an identity provider up to date, and support for .jfif files. Album editors can now update the title and description of a shared album.
The machine-learning pipeline behind the scenes
For self-hosters, the real headline is less the feature list than the inference cost it implies. Immich runs its face detection, object detection, and CLIP-based search locally, which means every uploaded photo triggers a small ML job on your own hardware. On a CPU-only server, indexing a large library is the single slowest operation Immich performs — often the difference between a smooth migration and a weekend-long wait.
That is why TensorRT matters. By moving inference onto the GPU’s dedicated accelerators and optimizing model loading, v3.3 targets the bottleneck self-hosters complain about most. The payoff compounds on the initial index: a library of 100,000 photos represents 100,000 small inference runs, and shaving milliseconds off each one turns hours into minutes. For anyone who has watched the CPU pegged at 100% during a first import, this is the change that makes Immich feel fast on modest hardware.
A release candidate to test, not to deploy
v3.3.0-rc.0 remains a release candidate. The Immich team repeats it at the top of the release notes: the RC “is subject to change and may contain bugs or breaking changes,” and you must back up the database and library before any upgrade. The upgrade path is the classic one for Docker Compose deployments: change the IMMICH_VERSION variable in the .env file, then restart the image.
# In .env
IMMICH_VERSION=v3.3.0-rc.0
# or, to follow the release-candidate channel:
IMMICH_VERSION=v3-rc
docker compose pull && docker compose up -dThe backup the team urges first goes through Immich’s two usual operations: a dump of the PostgreSQL database and a copy of the library folder. On the database side, the canonical command is:
docker exec -t immich_postgres pg_dumpall -c -U postgres | gzip > "dump.sql.gz"On mobile, the RC ships through the Play Store beta on Android and TestFlight on iOS. The team explicitly invites testers to report regressions on GitHub before the final release.
Immich against its alternatives
In the self-hosted ecosystem, Immich is not alone. PhotoPrism offers a library-management approach geared toward archiving, Nextcloud Memories integrates into the collaboration suite, and Piwigo remains a reference for shared albums. Immich’s strength is its focus: it does one thing — the personal photo experience — and does it at a Google Photos level, where generalist alternatives spread their effort thin.
v3.3’s people management widens that gap further. PhotoPrism has face recognition, but its correction of misclassified faces is less fluid than what Immich promises here. For a household switching from Google Photos, the first criterion is not feature count but familiarity: getting back your faces, memories, and search without relearning a logic. That is exactly what this RC aims for.
Verdict
If you already run Immich in production, stay on the stable v3.2.x branch and install the RC only on a test clone — people management and TensorRT are worth evaluating, but not at the cost of a family library. If you are standing up a new instance or comparing Google Photos alternatives, this is the moment to plan on v3.3: native people management and birthday memories close the two most visible usage gaps, and TensorRT acceleration makes AI indexing bearable even on a small server with an RTX GPU. And if you administer a multi-account deployment, watch the v3.3 OAuth synchronization — it will decide whether this release genuinely simplifies your identity management.