Self-Hosting

Immich vs PhotoPrism vs Nextcloud: self-host photos

Immich vs PhotoPrism vs Nextcloud Memories in 2026: mobile auto-backup, AI/face search, GPU acceleration, RAW/video, backups, and which to self-host.

Problem summary

There's no single winner among the self-hosted Google Photos alternatives — they're built around different priorities. Immich is the phone-replacement: the best native mobile auto-backup (especially reliable iOS background upload), CLIP natural-language + OCR + face search, and the only one that GPU/iGPU-accelerates its core ML (Intel OpenVINO, NVIDIA CUDA, AMD ROCm, Rockchip NPUs) — at the cost of a heavier stack (Postgres + Redis + ML container, 6–8 GB RAM) and the occasional breaking release. PhotoPrism is the stable, format-broad library: CPU-only built-in AI that runs anywhere, the widest RAW/HEIC/JPEG-XL and Live/Motion-photo support, and the lowest fuss — but no first-class mobile backup and label-based (not semantic) search. Nextcloud + the Memories app is the right layer only if you already run Nextcloud: a clean photo UI over your existing files, with AI via the Recognize app (CPU-only, and officially maintenance-only — the biggest long-term risk). One rule applies to all three: a file copy is NOT a backup — you must dump the database too.

Operator snapshotEvidence first

First proof

Name your top priority.

Screen to open

Check the immich Postgres volume is a local disk, not an NFS/SMB share

Expected signal

You know whether phone auto-backup, AI/semantic search, format breadth/stability, or reusing Nextcloud matters most.

Stop boundary

Don't adopt this stack just for photos if you don't already run Nextcloud — Recognize is feature-frozen.

Layer path

1There is no single 'best' self-hosted photo app — Immich, PhotoPrism, and Nextcloud Memories are built around different priorities, so the decision is a use-case map across mobile auto-backup, AI search, format support, stability, and footprint.

2Immich: the phone-replacement — best native mobile auto-backup (reliable iOS background upload), CLIP natural-language + OCR + face search, and the only one that GPU/iGPU-accelerates its core ML (Intel OpenVINO, NVIDIA CUDA, AMD ROCm, Rockchip NPUs; no Coral/Hailo). Heavier stack: Postgres + Redis + ML container, 6–8 GB RAM, Postgres on local SSD only.

3PhotoPrism: the stable, format-broad library — CPU-only built-in TensorFlow AI that runs anywhere, the widest RAW/HEIC/JPEG-XL and Live/Motion-photo support, native Google-Takeout JSON, Go single binary + SQLite/MariaDB. Weakness: no first-class mobile backup, and label-based (not semantic) search.

4Nextcloud + Memories: a photo UI that runs inside Nextcloud (not standalone), with AI via the Recognize app (CPU-only, ~4 GB RAM, officially maintenance-only — the biggest long-term risk). Right only if you already run Nextcloud. Safety boundary for all three: a file copy is NOT a backup — the database (paths/metadata/faces) must be dumped too, or photos restore orphaned with an empty timeline.

Runbook

Step-by-step runbook

Start here. Do each check in order, compare it to the expected result, and stop when the evidence explains the failure or the safe stop point applies.

Pin your constraints

Check: Write down your top priority, whether the phone is the source, RAM/GPU budget, formats, and stability tolerance.

Expected result: You have the inputs that decide the fit.

If not: Skipping this picks a tool weak on the axis you care about.

Map constraints to a primary app

Check: Phone backup / semantic search → Immich; stability / broad formats / CPU-only → PhotoPrism; already-Nextcloud → Memories.

Expected result: You have a primary candidate (a fit, not a ranking).

If not: If you have a spare iGPU/GPU for AI, that points to Immich.

Provision the stack correctly

Check: Immich: Postgres on local SSD + Redis + ML container, 6–8 GB RAM, enable the GPU backend. PhotoPrism: SQLite or MariaDB. Memories: install on existing Nextcloud + Recognize, matched versions.

Expected result: The app runs with its dependencies met.

If not: Don't put Immich's Postgres on a network share; don't mix incompatible Nextcloud/Memories/Recognize versions.

Import from Google Photos

Check: Immich: use immich-go against Takeout (preserves albums/metadata). PhotoPrism: drop extracted Takeout files in originals and index (reads JSON). Nextcloud: upload then index with Memories/Recognize.

Expected result: Your existing library is imported with metadata intact.

If not: immich-go is community-maintained and early — verify a sample album after import.

Set up and verify mobile backup

Check: Immich: configure per-album background backup in the app. PhotoPrism/Memories: configure the Nextcloud app or a sync tool.

Expected result: New phone photos land on the server automatically.

If not: On iOS, confirm Background App Refresh; see /self-hosting/immich-photos-not-backing-up if uploads stall.

Prove the backup with a restore drill

Check: Dump the database, back up the media, then restore both to a scratch instance and confirm the timeline rebuilds.

Expected result: You've proven a file-copy-plus-DB-dump restores correctly.

If not: If the timeline is empty on restore, you backed up files without the DB — fix the order.

Decision tree

If: Replacing Google Photos on the phone; reliable auto-backup is the point.

Then: Immich fits best.

Action: Run Immich (Postgres on local SSD, 6–8 GB RAM); set per-album background backup. If uploads stall, see /self-hosting/immich-photos-not-backing-up.

If: Want most stable, lowest-fuss library with the broadest RAW/format support; CPU-only is fine.

Then: PhotoPrism fits best.

Action: Run PhotoPrism (SQLite for small libraries, MariaDB at scale); index Google-Takeout files directly (it reads the JSON sidecars).

If: Already run Nextcloud and want photos over your existing files.

Then: Nextcloud + Memories fits best.

Action: Install Memories + Recognize, keeping all three (Nextcloud core/Memories/Recognize) version-compatible; use the Nextcloud app for auto-upload.

Safe stop: Don't adopt this stack just for photos if you don't already run Nextcloud — Recognize is feature-frozen.

If: Have an Intel iGPU / NVIDIA / Rockchip NPU you want to use for photo AI.

Then: Immich is the only one that benefits.

Action: Enable the matching ML backend (OpenVINO/CUDA/ROCm/RKNN); plan it with /nas/local-ai-on-nas-immich-frigate-ollama.

Evidence

Evidence table

Symptom	Evidence to collect	Likely layer	Next action
Choosing a self-hosted photo app.	Your top priority: backup vs search vs formats vs reuse-Nextcloud.	Core fit	Immich / PhotoPrism / Memories respectively.
Need reliable phone auto-backup.	Whether the tool has a first-class native backup app.	Mobile backup	Immich (native apps); others rely on third-party / Nextcloud app.
Want to use a GPU/iGPU for AI.	Presence of Intel iGPU / NVIDIA / Rockchip NPU.	ML acceleration	Immich (OpenVINO/CUDA/ROCm/RKNN); PhotoPrism/Recognize are CPU-only.
Restoring after a failure.	Whether you have a database dump, not just files.	Backup correctness	Dump Postgres/MariaDB/Nextcloud DB first, then media — or restore orphaned.

Reference

Commands and settings paths

Confirm Postgres is on local SSD (Immich)

Check the immich Postgres volume is a local disk, not an NFS/SMB share

Where: On the Immich host / compose file.

Expected: The DB lives on a local (ideally SSD) filesystem.

Failure means: Immich's Postgres on a network share corrupts/underperforms — unsupported.

Safe next step: Move the DB volume to local storage before importing a large library.

Check available ML acceleration

ls /dev/dri (Intel iGPU) or nvidia-smi (NVIDIA); confirm the model line for Rockchip

Where: On the photo-app host.

Expected: A supported accelerator is present for Immich's OpenVINO/CUDA/RKNN backend.

Failure means: No accelerator means Immich ML runs on CPU (slower); PhotoPrism/Recognize are CPU-only regardless.

Safe next step: Pick the ML backend per /nas/local-ai-on-nas-immich-frigate-ollama, or accept CPU indexing.

Verify a database dump exists before trusting a backup

Immich: check UPLOAD_LOCATION/backups for daily dumps; PhotoPrism: `photoprism backup`; Nextcloud: DB dump + data dir

Where: On the server, before relying on the backup.

Expected: A recent DB dump exists alongside the media backup.

Failure means: Files without a matching DB dump restore as orphaned assets with an empty timeline.

Safe next step: Schedule DB dumps first; verify with a real restore (see /self-hosting/home-server-backup-restore-tested).

Check RAM headroom for the chosen app

free -h (and the container memory limit)

Where: On the host.

Expected: >=6–8 GB free for Immich with ML; ~3–4 GB is workable for PhotoPrism/Recognize.

Failure means: Immich ML on 4 GB must be disabled; over-tight limits cause OOM during indexing.

Safe next step: Right-size RAM, or disable ML / pick PhotoPrism on constrained boxes.

Hardware boundary

Hardware and platform boundary

Change only when

Choose Immich when phone auto-backup and natural-language/face search are the point and you can give it 6–8 GB RAM (ideally a GPU/iGPU).
Choose PhotoPrism when you want the most stable, format-broad library and CPU-only AI is fine.
Choose Nextcloud + Memories only when you already run Nextcloud and want photos over your existing files.

Evidence that matters

RAM headroom (6–8 GB for Immich ML; 3–4 GB for PhotoPrism/Recognize) and SSD for the database/thumbnails.
A supported accelerator if you want fast AI (Intel iGPU/NVIDIA/Rockchip — Immich only).
A database-dump backup plan and a tested restore.

Evidence that does not matter

Coral/Hailo for photo AI — none of these use them (those are camera-NVR detectors).
Chasing the 'most AI' — match the search model (semantic vs label) to how you actually search.

Avoid

Treating a file copy as a backup — always dump the database too.
Putting Immich's Postgres on a network share (unsupported, corrupts).
Adopting the Nextcloud+Memories+Recognize stack just for photos when you don't already run Nextcloud (Recognize is feature-frozen).
Upgrading Immich across a major (e.g. v3.0) without reading release notes and backing up first.

Related tool

Use the linked tool to turn this runbook into a guided check for your exact setup.

NAS setup planner

Last reviewed

2026-06-23 · Reviewed by HomeTechOps. Built from June-2026 research verified against first-party Immich docs/GitHub (v2.0 stable Oct 2025, v3.0 June 2026; OpenVINO/CUDA/ROCm/RKNN ML accel; Postgres-on-local-SSD; automatic DB dumps), PhotoPrism docs (CPU-only TensorFlow, broad RAW/HEIC/JPEG-XL, native Takeout JSON), and Nextcloud Memories/Recognize repos (Recognize CPU-only and maintenance-only). Framed as a use-case decision map across mobile-backup/AI-search/formats/stability/footprint, never a 'winner'; exact versions and RAW matrices flagged to re-verify before relying on them.

Sources/assumptions

Facts verified 2026-06-23 against first-party Immich docs/GitHub (v2.0 stable Oct 2025, v3.0 June 2026), PhotoPrism docs (rolling calendar builds), and Nextcloud Memories/Recognize repos; version-specific behavior changes fast — re-verify exact versions before relying on them.
Immich requires Postgres on local SSD (not a network share) and 6–8 GB RAM with ML enabled; PhotoPrism's built-in ML is CPU-only; Nextcloud Recognize is CPU-only and officially 'maintained with limited effort' (critical fixes only).
This is a use-case decision map across mobile backup, AI search, format support, stability, and footprint — not a ranking; the right pick depends on your priority.

Source-backed checks

HomeTechOps turns official docs and conservative safety rules into a shorter runbook. These links are the source trail for the page direction.

Immich Docs: Hardware-Accelerated Machine LearningUsed for the Immich machine-learning acceleration paths (CUDA/OpenVINO/ARM-NN) and why the model store and PGData belong on a fast pool for smart-search and face-recognition performance.Immich docs: Backup and RestoreUsed for the pg_dump database backup, which asset folders to back up (library/upload/profile, not thumbnails), the restore command, and the earth/reverse-geocoding restore fix.PhotoPrism: Features (official)Used for PhotoPrism's CPU-only built-in TensorFlow AI (label/face, not CLIP-semantic), broad RAW/HEIC/JPEG-XL and Live/Motion-photo support, native Google-Takeout JSON metadata, and Go single-binary + SQLite/MariaDB architecture.Nextcloud Memories (official app / GitHub)Used for Memories being a photo UI that runs inside Nextcloud (not standalone), relying on the Nextcloud mobile app for auto-upload and on Recognize/facerecognition for AI tagging.Nextcloud Recognize (official app / GitHub)Used for Recognize's CPU-only TensorFlow.js AI tagging (faces/objects/landmarks/video), ~4 GB RAM need, AVX-or-WASM fallback, and its official 'maintained with limited effort' (critical-fixes-only) status — the long-term risk in the Nextcloud-photos stack.

Diagnostic rule

1. Check the simplest signal first.
2. Change one thing at a time.
3. Confirm the layer (app, container, network) before deep config.
4. Prove the restore; don't assume it.

FAQ

Which self-hosted photo app is best — Immich, PhotoPrism, or Nextcloud Memories?

It depends on your priority. Pick Immich if replacing Google Photos on your phone (reliable auto-backup) and natural-language/face search matter most, and you can give it 6–8 GB RAM (ideally an iGPU/GPU). Pick PhotoPrism if you want the most stable, lowest-fuss library with the broadest RAW/HEIC/JPEG-XL support and CPU-only AI. Pick Nextcloud Memories only if you already run Nextcloud and want a photo layer over your existing files. Many people self-host Immich for the phone backup and keep Nextcloud for documents.

Which has the most reliable mobile auto-backup?

Immich, clearly. It has first-class native iOS and Android apps with per-album background backup, device→server sync, and 'Free Up Space' to offload backed-up local files — and its iOS background-upload reliability is the most-cited reason people choose it. PhotoPrism has no first-party auto-backup app (it's browser/PWA-first; you'd use a third-party sync tool). Nextcloud Memories relies on the general Nextcloud mobile app's auto-upload, which is widely reported as the least reliable for iOS background uploads. If your phone is the camera you're protecting, that points to Immich. If your Immich phone won't upload, see Immich photos won't back up in the background.

Which can use my GPU or Intel iGPU for AI search and faces?

Only Immich accelerates its core ML on a GPU/iGPU — it supports Intel OpenVINO (Iris Xe/Arc), NVIDIA CUDA, AMD ROCm, and Rockchip NPUs (no Coral/Hailo). PhotoPrism's built-in TensorFlow AI and Nextcloud's Recognize are both CPU-only by design (a GPU only helps PhotoPrism's optional Ollama caption path). So if you have an Intel iGPU or an NVIDIA card you want to put to work on photo indexing, Immich is the one that benefits. Plan accelerators with local AI on a NAS.

What's the #1 backup mistake with these?

Copying only the photo files. All three keep paths/metadata/albums/faces in a database (Immich: Postgres + Redis; PhotoPrism: SQLite or MariaDB; Memories: the Nextcloud DB), and they do NOT rebuild that DB by re-scanning the library. Restore the files without the matching database dump and you get orphaned photos with an empty timeline. Always dump the database first, then back up the media — Immich now ships automatic daily DB dumps. Verify the whole thing with a real restore drill: test your home-server backup.

Which is the most stable / least likely to break on upgrade?

PhotoPrism — a Go single binary on rolling stable builds, no pre-1.0 churn. Immich is far more mature since its v2.0 stable (Oct 2025) but still ships occasional breaking releases (v3.0 in June 2026), so read release notes and back up the DB before upgrading. The riskiest is the Nextcloud + Memories + Recognize chain: three independently-versioned parts that must stay compatible, and Recognize is officially feature-frozen (critical fixes only) — fine if Nextcloud is already central to you, but plan for version-matching friction.

Planning a purchase?

We keep a source-backed, price-free comparison so you can buy once and right. No star ratings, every spec cited.

Synology vs UGREEN vs DIY NAS in 2026 →