Immich Machine Learning

Machine learning service for Immich — handles facial recognition, image classification, and semantic search using ONNX models.

Port	3003
Registry	ghcr.io/daemonless/immich-ml
Source	https://github.com/immich-app/immich
Website	https://immich.app/

Version Tags

Tag	Description	Best For
latest	Upstream Binary. Built from official release.	Most users. Matches Linux Docker behavior.
beta	Beta release built from upstream v3.0.0-rc.2.	Alternative build.

Prerequisites

Before deploying, ensure your host environment is ready. See the Quick Start Guide for host setup instructions.

Deployment

Podman Compose

services:
  immich-ml:
    image: "ghcr.io/daemonless/immich-ml:latest"
    container_name: immich-ml
    environment:
      - MACHINE_LEARNING_HOST=0.0.0.0  # Host to bind to (0.0.0.0)
      - MACHINE_LEARNING_PORT=3003  # Port to bind to (3003)
      - MACHINE_LEARNING_CACHE_FOLDER=/cache  # Path to cache folder (/cache)
      - PUID=1000  # User ID for the application process
      - PGID=1000  # Group ID for the application process
      - TZ=UTC  # Timezone for the container
      - MACHINE_LEARNING_WORKERS=1  # Number of ML worker processes. Keep at 1 on weak / CPU-only hosts.
      - MACHINE_LEARNING_WORKER_TIMEOUT=300  # Gunicorn worker timeout in seconds. Raise on slow CPUs so model loading doesn't time out and cycle the worker.
      - SKIP_CHOWN=true  # Skip the one-time recursive chown of /cache and /config once ownership is recorded in /config/.chown_done (default true). Set false to force a chown on every start.
    volumes:
      - "/path/to/containers/immich-ml/cache:/cache"
      - "/path/to/containers/immich-ml:/config"
    ports:
      - "3003:3003"
    restart: unless-stopped

AppJail Director

.env:

# .env

DIRECTOR_PROJECT=immich-ml
MACHINE_LEARNING_HOST=0.0.0.0
MACHINE_LEARNING_PORT=3003
MACHINE_LEARNING_CACHE_FOLDER=/cache
PUID=1000
PGID=1000
TZ=UTC
MACHINE_LEARNING_WORKERS=1
MACHINE_LEARNING_WORKER_TIMEOUT=300
SKIP_CHOWN=true

appjail-director.yml:

# appjail-director.yml

options:
  - virtualnet: ':<random> default'
  - nat:
services:
  immich-ml:
    name: immich_ml
    options:
      - container: 'boot args:--pull'
      - expose: '3003:3003 proto:tcp' \
    oci:
      user: root
      environment:
        - MACHINE_LEARNING_HOST: !ENV '${MACHINE_LEARNING_HOST}'
        - MACHINE_LEARNING_PORT: !ENV '${MACHINE_LEARNING_PORT}'
        - MACHINE_LEARNING_CACHE_FOLDER: !ENV '${MACHINE_LEARNING_CACHE_FOLDER}'
        - PUID: !ENV '${PUID}'
        - PGID: !ENV '${PGID}'
        - TZ: !ENV '${TZ}'
        - MACHINE_LEARNING_WORKERS: !ENV '${MACHINE_LEARNING_WORKERS}'
        - MACHINE_LEARNING_WORKER_TIMEOUT: !ENV '${MACHINE_LEARNING_WORKER_TIMEOUT}'
        - SKIP_CHOWN: !ENV '${SKIP_CHOWN}'
    volumes:
      - immich-ml_cache: /cache
      - immich-ml: /config
volumes:
  immich-ml_cache:
    device: '/path/to/containers/immich-ml/cache'
  immich-ml:
    device: '/path/to/containers/immich-ml'

Makejail:

# Makejail

ARG tag=latest

OPTION overwrite=force
OPTION from=ghcr.io/daemonless/immich-ml:${tag}

Note: Exposing ports in AppJail means that your service can be reached from remote hosts. If that is not your intention, do not expose the ports and communicate with the service using the IPv4 address assigned by the virtual network.

Podman CLI

podman run -d --name immich-ml \
  -p 3003:3003 \
  -e MACHINE_LEARNING_HOST=0.0.0.0 \
  -e MACHINE_LEARNING_PORT=3003 \
  -e MACHINE_LEARNING_CACHE_FOLDER=/cache \
  -e PUID=1000 \
  -e PGID=1000 \
  -e TZ=UTC \
  -e MACHINE_LEARNING_WORKERS=1 \
  -e MACHINE_LEARNING_WORKER_TIMEOUT=300 \
  -e SKIP_CHOWN=true \
  -v /path/to/containers/immich-ml/cache:/cache \
  -v /path/to/containers/immich-ml:/config \
  ghcr.io/daemonless/immich-ml:latest

AppJail

appjail oci run -Pd \
  -o overwrite=force \
  -o container="args:--pull" \
  -o virtualnet=":<random> default" \
  -o nat \
  -o expose="3003:3003 proto:tcp" \
  -e MACHINE_LEARNING_HOST=0.0.0.0 \
  -e MACHINE_LEARNING_PORT=3003 \
  -e MACHINE_LEARNING_CACHE_FOLDER=/cache \
  -e PUID=1000 \
  -e PGID=1000 \
  -e TZ=UTC \
  -e MACHINE_LEARNING_WORKERS=1 \
  -e MACHINE_LEARNING_WORKER_TIMEOUT=300 \
  -e SKIP_CHOWN=true \
  -o fstab="/path/to/containers/immich-ml/cache /cache <pseudofs>" \
  -o fstab="/path/to/containers/immich-ml /config <pseudofs>" \
  ghcr.io/daemonless/immich-ml:latest immich-ml

Note: Exposing ports in AppJail means that your service can be reached from remote hosts. If that is not your intention, do not expose the ports and communicate with the service using the IPv4 address assigned by the virtual network.

Ansible

- name: Deploy immich-ml
  containers.podman.podman_container:
    name: immich-ml
    image: "ghcr.io/daemonless/immich-ml:latest"
    state: started
    restart_policy: always
    env:
      MACHINE_LEARNING_HOST: "0.0.0.0"
      MACHINE_LEARNING_PORT: "3003"
      MACHINE_LEARNING_CACHE_FOLDER: "/cache"
      PUID: "1000"
      PGID: "1000"
      TZ: "UTC"
      MACHINE_LEARNING_WORKERS: "1"
      MACHINE_LEARNING_WORKER_TIMEOUT: "300"
      SKIP_CHOWN: "true"
    ports:
      - "3003:3003"
    volumes:
      - "/path/to/containers/immich-ml/cache:/cache"
      - "/path/to/containers/immich-ml:/config"

Parameters

Environment Variables

Variable	Default	Description
MACHINE_LEARNING_HOST	0.0.0.0	Host to bind to (0.0.0.0)
MACHINE_LEARNING_PORT	3003	Port to bind to (3003)
MACHINE_LEARNING_CACHE_FOLDER	/cache	Path to cache folder (/cache)
PUID	1000	User ID for the application process
PGID	1000	Group ID for the application process
TZ	UTC	Timezone for the container
MACHINE_LEARNING_WORKERS	1	Number of ML worker processes. Keep at 1 on weak / CPU-only hosts.
MACHINE_LEARNING_WORKER_TIMEOUT	300	Gunicorn worker timeout in seconds. Raise on slow CPUs so model loading doesn't time out and cycle the worker.
SKIP_CHOWN	true	Skip the one-time recursive chown of /cache and /config once ownership is recorded in /config/.chown_done (default true). Set false to force a chown on every start.

Volumes

Path	Description
/cache	Model cache (HuggingFace ONNX models). Use a persistent volume to avoid re-downloading models.
/config	Gunicorn HOME and the .chown_done ownership marker. Must be a persistent volume (the chown-skip relies on it).

Ports

Port	Protocol	Description
3003	TCP	ML API

This image is part of the Immich Stack.

Low-power / CPU-only hosts

On weak or CPU-only machines, model loading can be slow and the worker may time out during startup (repeated healthy/unhealthy cycles). Raise the timeout and pin the numeric libraries to a single thread to avoid thrashing:

environment:
  MACHINE_LEARNING_WORKERS: "1"
  MACHINE_LEARNING_WORKER_TIMEOUT: "300"
  MACHINE_LEARNING_MODEL_INTER_OP_THREADS: "1"
  MACHINE_LEARNING_MODEL_INTRA_OP_THREADS: "1"
  MACHINE_LEARNING_REQUEST_THREADS: "1"
  OMP_NUM_THREADS: "1"
  OPENBLAS_NUM_THREADS: "1"
  MKL_NUM_THREADS: "1"

Persistent volumes

Mount both /config and /cache as persistent volumes. The one-time recursive chown is skipped on later starts via a /config/.chown_done marker — without a persistent /config, the marker is lost and the chown re-runs every start, so SKIP_CHOWN=true only takes effect once /config persists the marker.

Architectures: amd64 User: bsd (UID/GID via PUID/PGID, defaults to 1000:1000) Base: FreeBSD 15.1

---

Need help? Join our Discord community.