For most people, Google Photos or iCloud is where photos go to live—and where privacy goes to die. Between storage paywalls, creeping compression algorithms, and the uncomfortable reality of handing over your entire visual biography to advertising giants, the motivation to own your photo library has never been higher.
Enter Immich. Over the last couple of years, Immich has evolved into the indisputable gold standard of self-hosted photo management. It doesn’t just mimic Google Photos; in many ways, it outclasses it. You get lightning-fast timeline scrolling, beautiful native mobile apps with automatic background backup, multi-user partner sharing, and best of all, local AI: semantic CLIP search (“dog in snow”, “red bicycle”) and facial clustering, all running entirely on your own infrastructure.
However, once you jump down the self-hosting rabbit hole, you immediately hit a hard reality: photo indexing and computer vision are ridiculously resource-intensive.
THE HIDDEN BOTTLENECK OF HOME PHOTO HOSTING
If you run a typical homelab or low-power home server—think a Raspberry Pi 5, an Intel N100 mini-PC, or an old repurposed office desktop—your hardware is tuned for quiet, low-wattage, 24/7 reliability. It sips 10 to 25 watts and hums along peacefully in a corner.
Now drop a 100,000-photo personal archive and 4K family videos onto that little box.
Immich manages three primary workloads:
- Immich Server (API): Serves web and mobile client queries.
- Immich Microservices: Chews through raw EXIF parsing, video transcoding, and multiple thumbnail resolutions (small, large, preview).
- Immich Machine Learning: Loads gigabyte-heavy neural network models into RAM to run facial recognition and vector embeddings.
On modest edge hardware, your CPU instantly pins to 100%, fans start screaming, temperatures soar, and RAM runs dry. Worst of all, because the CPU is choked by background encoding, the primary API crawls. You open the app on your phone, and the timeline stutters or fails to load thumbnails. That completely defeats the purpose of having a seamless, family-friendly photo service.
THE SOLUTION: SPLITTING BRAINS FROM STORAGE
The key insight is realizing that storage sovereignty and compute muscle do not need to live inside the exact same physical chassis.
I wanted my original high-resolution photos and primary database to live physically at home on my own disks—safe under my own roof. But I also wanted machine learning and video transcoding to run at blistering speeds without melting my home server or hiking up my home electricity bill.
So I split Immich into a distributed hybrid architecture across my low-power home server and a generous, free-tier ARM cloud instance (such as Oracle Cloud’s always-free 4-core Ampere VM):
1. The Home Server (The Vault & Front Door)
My home server acts as the authoritative store of record and user gateway:
- Physical Storage: Holds the raw master library on dedicated local drives.
- Database: Runs PostgreSQL 14 with
pgvector(or VectorChord) for blazing-fast vector queries, paired with a Valkey/Redis cache. - Immich Server (API Only): Configured with
IMMICH_WORKERS_INCLUDE=api. It does zero heavy lifting. Its sole job is to accept mobile uploads and serve API responses to clients. Because it skips ML and transcoding, client responsiveness remains instantaneous.
2. The Cloud Worker (The Heavy Lifter)
For the muscle, I take advantage of an always-free cloud ARM virtual machine (4 Ampere cores, 24 GB of RAM):
- Immich Microservices: Configured with
IMMICH_WORKERS_EXCLUDE=api. It connects directly to the home database and Redis queue over the VPN, handling thumbnail generation, video transcoding, and metadata processing. - Immich Machine Learning: Runs the ML container with generous model caches for facial detection and CLIP search.
- Encrypted Mesh Transport: Both nodes communicate across a private, WireGuard-powered Tailscale/Headscale mesh network. No ports are opened to the public internet.
- Storage Access via CIFS/SMB: The cloud node mounts the library, uploads, and thumbnail cache directories directly from the home server over the VPN using authenticated CIFS driver mounts in Docker.
A Note on Network Bandwidth: Because the cloud worker reads incoming raw media from your home server to transcode it and writes thumbnails back, your home internet upload speed matters. Symmetrical fiber or a broadband upload connection of at least 25–50 Mbps will provide the smoothest experience for large 4K video uploads.
THE ARCHITECTURE IN CODE
Immich natively supports worker splitting out of the box through environment flags. (Note: Do not wrap the worker values in quotes—use IMMICH_WORKERS_INCLUDE=api, not ='api').
On the Home Server (API Only):
# docker-compose.yml on Home Server
services:
immich_server:
container_name: immich_server
image: ghcr.io/immich-app/immich-server:release
environment:
- IMMICH_WORKERS_INCLUDE=api # <-- Only serve web and mobile clients!
- DB_HOSTNAME=immich_postgres
- REDIS_HOSTNAME=immich_redis
- TZ=Etc/UTC # Change to your timezone (e.g. America/New_York)
volumes:
- /mnt/storage/photos:/data/library
- /etc/container_configs/immich/cache:/data/thumbs
- /etc/container_configs/immich/upload:/data/upload
ports:
- "2283:2283"
networks:
- immich_net
restart: unless-stopped
immich_postgres:
container_name: immich_postgres
image: ghcr.io/immich-app/postgres:14-vectorchord-pgvectors
ports:
# Bind to your private Tailscale/VPN IP so the cloud worker can reach the DB
- "100.x.y.z:5432:5432"
networks:
- immich_net
restart: unless-stopped
immich_redis:
container_name: immich_redis
image: valkey/valkey:9
ports:
# Bind to your private Tailscale/VPN IP
- "100.x.y.z:6379:6379"
networks:
- immich_net
restart: unless-stopped
networks:
immich_net:
On the Cloud Worker Node (Microservices + Machine Learning):
# docker-compose.yml on Cloud Node
services:
immich_microservices:
container_name: immich_microservices
image: ghcr.io/immich-app/immich-server:release
environment:
- IMMICH_WORKERS_EXCLUDE=api # <-- Run all background queues EXCEPT the API!
- DB_HOSTNAME=homeserver.vpn # Reaches home DB over Tailscale/Headscale
- REDIS_HOSTNAME=homeserver.vpn
- TZ=Etc/UTC
volumes:
- immich-data:/data/library
- immich-thumbs:/data/thumbs
- immich-upload:/data/upload
restart: unless-stopped
immich_machine_learning:
container_name: immich_machine_learning
image: ghcr.io/immich-app/immich-machine-learning:release
ports:
# Expose port 3003 over VPN so the Home API can query Smart Search / CLIP embeddings
- "3003:3003"
volumes:
- model-cache:/cache
restart: unless-stopped
volumes:
model-cache:
immich-data:
driver: local
driver_opts:
type: cifs
device: "//homeserver.vpn/media/photos"
o: "addr=homeserver.vpn,username=backupuser,password=${SMB_PW},uid=1000,gid=100"
immich-thumbs:
driver: local
driver_opts:
type: cifs
device: "//homeserver.vpn/cache/thumbs"
o: "addr=homeserver.vpn,username=backupuser,password=${SMB_PW},uid=1000,gid=100"
immich-upload:
driver: local
driver_opts:
type: cifs
device: "//homeserver.vpn/cache/upload"
o: "addr=homeserver.vpn,username=backupuser,password=${SMB_PW},uid=1000,gid=100"
Important Configuration Step: Once deployed, log in to your Immich Web UI as Admin, navigate to Administration → Settings → Machine Learning Settings, and set the ML URL to
http://cloudnode.vpn:3003. This ensures that when you run a live Smart Search query from your phone, the home API knows how to reach the cloud machine learning model to vectorize your query!
WHAT THIS FEELS LIKE IN PRACTICE
The operational difference is night and day:
- Zero Home Server Strain: When you dump 500 photos from a vacation, the home server simply ingests the incoming files and enqueues the tasks in Redis. The 4-core cloud instance immediately picks up the jobs, churning through face detection, facial clustering, and CLIP search indexing with 24 GB of memory headroom. The home server’s CPU graph barely registers a ripple.
- Instant Mobile UI: Because the home API worker is completely freed from transcoding and matrix math, pulling up albums or scrolling years back on the mobile app is fluid and immediate.
- True Data Sovereignty: At no point are photos stored long-term in the cloud. The cloud instance is strictly a stateless compute engine. The files reside on my own physical disks, protected by automated database snapshots and multi-tier backup routines.
- Completely Free Compute: By utilizing always-free ARM cloud tiers that would otherwise sit idle, you get enterprise-grade compute throughput for zero dollars in extra hardware or electricity bills.
CLOSING THOUGHTS
Self-hosting doesn’t mean you have to buy a rack full of noisy enterprise servers to get modern AI features. By combining lightweight local edge storage with cloud compute workers over WireGuard, you get the best of both worlds: uncompromised privacy, physical ownership of your memories, and the speed of modern cloud processing.
If you’ve been putting off self-hosting Immich because your home hardware felt too weak for machine learning, this hybrid setup is the blueprint you need.