Self-Host PhotoPrism: AI-Powered Photo Management 2026

TL;DR

PhotoPrism (AGPL 3.0, ~34K GitHub stars, Go) is a self-hosted AI photo management app that automatically classifies, tags, and organizes your photo library. It detects scenes ("beach", "mountains", "food"), recognizes faces, geo-tags from EXIF, and creates smart albums — all running locally without sending data to any cloud. Compare: Google Photos charges $2.99/month for 100GB and trains on your photos. PhotoPrism is free, your data stays on-prem, and CPU inference runs on any server.

Key Takeaways

PhotoPrism: AGPL 3.0, ~34K stars, Go — AI classification, face recognition, smart albums
Scene detection: Classifies 1000+ scene types (beach, food, wedding, architecture) automatically
Face recognition: Clusters faces, lets you name people — works fully offline
No cloud required: All AI inference runs locally; optional GPU acceleration via TensorFlow
Import-friendly: Reads from any directory, preserves folder structure, handles duplicates
Immich comparison: PhotoPrism = mature AI + albums; Immich = faster UI + mobile backup focus

PhotoPrism vs Immich vs Google Photos

Feature	PhotoPrism	Immich	Google Photos
License	AGPL 3.0	AGPL 3.0	Proprietary
Cost	Free	Free	$2.99/mo (100GB)
AI Scene detection	Yes (1000+ labels)	Basic	Yes
Face recognition	Yes	Yes	Yes
Mobile backup app	Limited	Yes (iOS + Android)	Yes
Smart albums	Yes	Yes	Yes
Video support	Yes	Yes	Yes
RAW support	Yes	Yes	Yes
Maps / geo	Yes	Yes	Yes
Duplicate detection	Yes	Limited	Yes
GPU acceleration	Optional (TF)	Yes (ML)	—
GitHub Stars	~34K	~55K	—

Part 1: Docker Setup

# docker-compose.yml
services:
  photoprism:
    image: photoprism/photoprism:latest
    container_name: photoprism
    restart: unless-stopped
    ports:
      - "2342:2342"
    security_opt:
      - seccomp:unconfined
      - apparmor:unconfined
    environment:
      PHOTOPRISM_AUTH_MODE: "password"
      PHOTOPRISM_SITE_URL: "https://photos.yourdomain.com"
      PHOTOPRISM_ORIGINALS_LIMIT: 5000          # MB, -1 = unlimited
      PHOTOPRISM_HTTP_COMPRESSION: "gzip"
      PHOTOPRISM_LOG_LEVEL: "info"
      PHOTOPRISM_READONLY: "false"
      PHOTOPRISM_EXPERIMENTAL: "false"
      PHOTOPRISM_DISABLE_CHOWN: "false"
      PHOTOPRISM_DISABLE_WEBDAV: "false"
      PHOTOPRISM_DISABLE_SETTINGS: "false"
      PHOTOPRISM_DISABLE_TENSORFLOW: "false"    # Set true to skip AI (faster startup)
      PHOTOPRISM_DISABLE_FACES: "false"
      PHOTOPRISM_DISABLE_CLASSIFICATION: "false"
      PHOTOPRISM_DISABLE_VECTORS: "false"
      PHOTOPRISM_DISABLE_RAW: "false"
      PHOTOPRISM_RAW_PRESETS: "false"
      PHOTOPRISM_JPEG_QUALITY: 85
      PHOTOPRISM_DETECT_NSFW: "false"
      PHOTOPRISM_UPLOAD_NSFW: "true"
      PHOTOPRISM_DATABASE_DRIVER: "mysql"
      PHOTOPRISM_DATABASE_SERVER: "mariadb:3306"
      PHOTOPRISM_DATABASE_NAME: "photoprism"
      PHOTOPRISM_DATABASE_USER: "photoprism"
      PHOTOPRISM_DATABASE_PASSWORD: "${PHOTOPRISM_DB_PASSWORD}"
      PHOTOPRISM_ADMIN_USER: "admin"
      PHOTOPRISM_ADMIN_PASSWORD: "${PHOTOPRISM_ADMIN_PASSWORD}"
      PHOTOPRISM_UID: 1000
      PHOTOPRISM_GID: 1000
    volumes:
      - /path/to/your/photos:/photoprism/originals   # Your photo library (read-write)
      - photoprism_storage:/photoprism/storage        # AI models, thumbnails, sidecar files
    depends_on:
      mariadb:
        condition: service_healthy

  mariadb:
    image: mariadb:11
    container_name: photoprism_db
    restart: unless-stopped
    command: --innodb-buffer-pool-size=512M --transaction-isolation=READ-COMMITTED
    healthcheck:
      test: mysqladmin ping -h localhost -u photoprism --password="${PHOTOPRISM_DB_PASSWORD}"
      interval: 10s
      start_period: 20s
    environment:
      MARIADB_AUTO_UPGRADE: "1"
      MARIADB_INITDB_SKIP_TZINFO: "1"
      MARIADB_DATABASE: "photoprism"
      MARIADB_USER: "photoprism"
      MARIADB_PASSWORD: "${PHOTOPRISM_DB_PASSWORD}"
      MARIADB_ROOT_PASSWORD: "${MARIADB_ROOT_PASSWORD}"
    volumes:
      - mariadb_data:/var/lib/mysql

volumes:
  photoprism_storage:
  mariadb_data:

# .env
PHOTOPRISM_ADMIN_PASSWORD=your-secure-admin-password
PHOTOPRISM_DB_PASSWORD=your-db-password
MARIADB_ROOT_PASSWORD=your-root-password

docker compose up -d

Visit http://your-server:2342 — login with admin / your password.

Part 2: HTTPS with Caddy

photos.yourdomain.com {
    reverse_proxy localhost:2342
}

Part 3: Initial Photo Import

Import from an Existing Directory

If your photos are already on the server at /path/to/your/photos:

That path is already mounted as /photoprism/originals
Go to Library → Index → Start
PhotoPrism indexes all photos, generates thumbnails, runs AI classification

Import via Web Upload

Library → Upload — drag and drop from browser
Files go into /photoprism/originals/[year]/[month]/

Import via WebDAV

PhotoPrism exposes a WebDAV endpoint for remote file copying:

# Mount with macOS Finder or any WebDAV client:
# URL: https://photos.yourdomain.com/originals/

# Or use rclone:
rclone copy /local/photos "photoprism:originals" --transfers 4

CLI Import

docker exec photoprism photoprism import /photoprism/originals

Part 4: AI Features

Scene Classification

After indexing, PhotoPrism automatically adds labels like:

beach, sunset, mountains, forest, city, architecture
food, drink, restaurant
birthday, wedding, graduation
dog, cat, bird, horse
car, bicycle, airplane

Browse by label: Search → Labels → [label name]

Face Recognition

PhotoPrism clusters faces and lets you assign names:

People → Unmatched Faces — review auto-detected faces
Click a face cluster → Name → type a person's name
PhotoPrism learns and matches that person across the library
People → [Name] — all photos of that person

Face recognition runs locally via TensorFlow — no data leaves your server.

Smart Albums

PhotoPrism auto-creates albums based on:

Moments: Auto-grouped by date + location ("Paris, June 2024")
States/Countries: Photos grouped by geo-location
Calendar: Browse by year/month
Colors: Filter by dominant color

Manual albums: Albums → + New Album → add photos manually or via search.

Part 5: Search

PhotoPrism's search is powerful:

# Search by label:
label:beach

# Search by date range:
after:2024-01-01 before:2024-12-31

# Search by camera:
camera:iphone

# Search by person:
person:Alice

# Search by geo:
near:"San Francisco"

# Combine:
label:sunset after:2024-06-01 before:2024-09-01

# Only videos:
type:video

# Favorites:
is:favorite

# Unreviewed:
quality:0

Part 6: GPU Acceleration (Optional)

For faster AI classification and video transcoding:

NVIDIA GPU

services:
  photoprism:
    image: photoprism/photoprism:latest
    # Add GPU access:
    runtime: nvidia
    environment:
      NVIDIA_VISIBLE_DEVICES: all
      PHOTOPRISM_INIT: "tensorflow-gpu intel-graphics"

Requires nvidia-container-toolkit on the host.

Apple Silicon / ARM

PhotoPrism has native ARM64 builds — runs well on Apple Silicon Macs or ARM servers (Raspberry Pi 4+ with 4GB RAM for small libraries).

CPU-Only Mode (No AI)

If you want a fast photo browser without AI overhead:

environment:
  PHOTOPRISM_DISABLE_TENSORFLOW: "true"
  PHOTOPRISM_DISABLE_FACES: "true"
  PHOTOPRISM_DISABLE_CLASSIFICATION: "true"

Indexing is 10-20x faster — you lose auto-labeling and face recognition but keep geo, EXIF, and manual organization.

Part 7: Storage Layout

/photoprism/originals/    ← Your original photos (never modified)
  2024/
    01/
      IMG_1234.jpg
      IMG_1235.jpg
    06/
      vacation/
        IMG_2000.jpg

/photoprism/storage/      ← PhotoPrism's working directory
  cache/
    thumbnails/           ← Auto-generated thumbnails
  sidecar/                ← XMP/JSON metadata sidecar files
  albums/                 ← Album definitions
  tensorflow/             ← AI model files (~300MB, auto-downloaded on first run)

Originals are never modified. Metadata is stored in sidecar files and the database.

Click any photo → share icon
Generate a public link with optional password + expiry
Recipients can view without creating an account

WebDAV Access

Access your photos from Lightroom, Finder, or any WebDAV client:

URL: https://photos.yourdomain.com/originals/
Username: admin
Password: your-password

API

# Get session token:
TOKEN=$(curl -s -X POST https://photos.yourdomain.com/api/v1/session \
  -d '{"username":"admin","password":"your-password"}' | jq -r .id)

# Search photos:
curl "https://photos.yourdomain.com/api/v1/photos?q=beach&count=10" \
  -H "X-Session-ID: $TOKEN" | jq '.[].FileName'

# Download a photo:
curl "https://photos.yourdomain.com/api/v1/photo/UUID/dl" \
  -H "X-Session-ID: $TOKEN" -O

Part 9: Mobile Access

PhotoPrism doesn't have a dedicated mobile backup app (unlike Immich). Workarounds:

Progressive Web App (PWA): Visit the URL in Safari/Chrome → "Add to Home Screen" → works like a native app
Syncthing: Auto-sync phone photos to your server's originals folder → PhotoPrism indexes automatically
Rclone: Scheduled sync from phone (via Termux on Android or Scriptable on iOS)

Maintenance

# Update PhotoPrism:
docker compose pull
docker compose up -d

# Reindex after adding photos:
docker exec photoprism photoprism index --cleanup

# Convert RAW to JPEG (for web preview):
docker exec photoprism photoprism convert

# Rebuild faces index:
docker exec photoprism photoprism faces index

# Backup database:
docker exec photoprism_db mysqldump -u photoprism -p"${PHOTOPRISM_DB_PASSWORD}" photoprism \
  | gzip > photoprism-db-$(date +%Y%m%d).sql.gz

# Logs:
docker compose logs -f photoprism

# Check storage size:
docker exec photoprism du -sh /photoprism/storage/cache/thumbnails/

Why Self-Host PhotoPrism?

The case for self-hosting PhotoPrism comes down to three practical factors: data ownership, cost at scale, and operational control.

Data ownership is the fundamental argument. When you use a SaaS version of any tool, your data lives on someone else's infrastructure subject to their terms of service, their security practices, and their business continuity. If the vendor raises prices, gets acquired, changes API limits, or shuts down, you're left scrambling. Self-hosting PhotoPrism means your data and configuration stay on infrastructure you control — whether that's a VPS, a bare metal server, or a home lab.

Cost at scale matters once you move beyond individual use. Most SaaS equivalents charge per user or per data volume. A self-hosted instance on a $10-20/month VPS typically costs less than per-user SaaS pricing for teams of five or more — and the cost doesn't scale linearly with usage. One well-configured server handles dozens of users for a flat monthly fee.

Operational control is the third factor. The Docker Compose configuration above exposes every setting that commercial equivalents often hide behind enterprise plans: custom networking, environment variables, storage backends, and authentication integrations. You decide when to update, how to configure backups, and what access controls to apply.

The honest tradeoff: you're responsible for updates, backups, and availability. For teams running any production workloads, this is familiar territory. For individuals, the learning curve is real but the tooling (Docker, Caddy, automated backups) is well-documented and widely supported.

Server Requirements and Sizing

Before deploying PhotoPrism, assess your server capacity against expected workload.

Minimum viable setup: A 1 vCPU, 1GB RAM VPS with 20GB SSD is sufficient for personal use or small teams. Most consumer VPS providers — Hetzner, DigitalOcean, Linode, Vultr — offer machines in this range for $5-10/month. Hetzner offers excellent price-to-performance for European and US regions.

Recommended production setup: 2 vCPUs with 4GB RAM and 40GB SSD handles most medium deployments without resource contention. This gives PhotoPrism headroom for background tasks, caching, and concurrent users while leaving capacity for other services on the same host.

Storage planning: The Docker volumes in this docker-compose.yml store all persistent PhotoPrism data. Estimate your storage growth rate early — for data-intensive tools, budget for 3-5x your initial estimate. Hetzner Cloud and Vultr both support online volume resizing without stopping your instance.

Operating system: Any modern 64-bit Linux distribution works. Ubuntu 22.04 LTS and Debian 12 are the most commonly tested configurations. Ensure Docker Engine 24.0+ and Docker Compose v2 are installed — verify with docker --version and docker compose version. Avoid Docker Desktop on production Linux servers; it adds virtualization overhead and behaves differently from Docker Engine in ways that cause subtle networking issues.

Network: Only ports 80 and 443 need to be publicly accessible when running behind a reverse proxy. Internal service ports should be bound to localhost only. A minimal UFW firewall that blocks all inbound traffic except SSH, HTTP, and HTTPS is the single most effective security measure for a self-hosted server.

Backup and Disaster Recovery

Running PhotoPrism without a tested backup strategy is an unacceptable availability risk. Docker volumes are not automatically backed up — if you delete a volume or the host fails, data is gone with no recovery path.

What to back up: The named Docker volumes containing PhotoPrism's data (database files, user uploads, application state), your docker-compose.yml and any customized configuration files, and .env files containing secrets.

Backup approach: For simple setups, stop the container, archive the volume contents, then restart. For production environments where stopping causes disruption, use filesystem snapshots or database dump commands (PostgreSQL pg_dump, SQLite .backup, MySQL mysqldump) that produce consistent backups without downtime.

For a complete automated backup workflow that ships snapshots to S3-compatible object storage, see the Restic + Rclone backup guide. Restic handles deduplication and encryption; Rclone handles multi-destination uploads. The same setup works for any Docker volume.

Backup cadence: Daily backups to remote storage are a reasonable baseline for actively used tools. Use a 30-day retention window minimum — long enough to recover from mistakes discovered weeks later. For critical data, extend to 90 days and use a secondary destination.

Restore testing: A backup that has never been restored is a backup you cannot trust. Once a month, restore your PhotoPrism backup to a separate Docker Compose stack on different ports and verify the data is intact. This catches silent backup failures, script errors, and volume permission issues before they matter in a real recovery.

Security Hardening

Self-hosting means you are responsible for PhotoPrism's security posture. The Docker Compose setup provides a functional base; production deployments need additional hardening.

Always use a reverse proxy: Never expose PhotoPrism's internal port directly to the internet. The docker-compose.yml binds to localhost; Caddy or Nginx provides HTTPS termination. Direct HTTP access transmits credentials in plaintext. A reverse proxy also centralizes TLS management, rate limiting, and access logging.

Strong credentials: Change default passwords immediately after first login. For secrets in docker-compose environment variables, generate random values with openssl rand -base64 32 rather than reusing existing passwords.

Firewall configuration:

ufw default deny incoming
ufw allow 22/tcp
ufw allow 80/tcp
ufw allow 443/tcp
ufw enable

Internal service ports (databases, admin panels, internal APIs) should only be reachable from localhost or the Docker network, never directly from the internet.

Network isolation: Docker Compose named networks keep PhotoPrism's services isolated from other containers on the same host. Database containers should not share networks with containers that don't need direct database access.

VPN access for sensitive services: For internal-only tools, restricting access to a VPN adds a strong second layer. Headscale is an open source Tailscale control server that puts your self-hosted stack behind a WireGuard mesh, eliminating public internet exposure for internal tools.

Update discipline: Subscribe to PhotoPrism's GitHub releases page to receive security advisory notifications. Schedule a monthly maintenance window to pull updated images. Running outdated container images is the most common cause of self-hosted service compromises.

Troubleshooting Common Issues

Container exits immediately or won't start

Check logs first — they almost always explain the failure:

docker compose logs -f photoprism

Common causes: a missing required environment variable, a port already in use, or a volume permission error. Port conflicts appear as bind: address already in use. Find the conflicting process with ss -tlpn | grep PORT and either stop it or change PhotoPrism's port mapping in docker-compose.yml.

Cannot reach the web interface

Work through this checklist:

Confirm the container is running: docker compose ps
Test locally on the server: curl -I http://localhost:PORT
If local access works but external doesn't, check your firewall: ufw status
If using a reverse proxy, verify it's running and the config is valid: caddy validate --config /etc/caddy/Caddyfile

Permission errors on volume mounts

Some containers run as a non-root user. If the Docker volume is owned by root, the container process cannot write to it. Find the volume's host path with docker volume inspect VOLUME_NAME, check the tool's documentation for its expected UID, and apply correct ownership:

chown -R 1000:1000 /var/lib/docker/volumes/your_volume/_data

High resource usage over time

Memory or CPU growing continuously usually indicates unconfigured log rotation, an unbound cache, or accumulated data needing pruning. Check current usage with docker stats photoprism. Add resource limits in docker-compose.yml to prevent one container from starving others. For ongoing visibility into resource trends, deploy Prometheus + Grafana or Netdata.

Data disappears after container restart

Data stored in the container's writable layer — rather than a named volume — is lost when the container is removed or recreated. This happens when the volume mount path in docker-compose.yml doesn't match where the application writes data. Verify mount paths against the tool's documentation and correct the mapping. Named volumes persist across container removal; only docker compose down -v deletes them.

Keeping PhotoPrism Updated

PhotoPrism follows a regular release cadence. Staying current matters for security patches and compatibility. The update process with Docker Compose is straightforward:

docker compose pull          # Download updated images
docker compose up -d         # Restart with new images
docker image prune -f        # Remove old image layers (optional)

Read the changelog before major version updates. Some releases include database migrations or breaking configuration changes. For major version bumps, test in a staging environment first — run a copy of the service on different ports with the same volume data to validate the migration before touching production.

Version pinning: For stability, pin to a specific image tag in docker-compose.yml instead of latest. Update deliberately after reviewing the changelog. This trades automatic patch delivery for predictable behavior — the right call for business-critical services.

Post-update verification: After updating, confirm PhotoPrism is functioning correctly. Most services expose a /health endpoint that returns HTTP 200 — curl it from the server or monitor it with your uptime tool.

See all open source photo management tools at OSSAlt.com/categories/photos.

The SaaS-to-Self-Hosted Migration Guide (Free PDF)