Posted by jamesxv7 16 hours ago
Ask HN: What's the 2025 stack for a self-hosted photo library with local AI?
My goal is to create a system with smart search capabilities, and one of the most important requirements is that it must run entirely on my local hardware. Privacy is key, but the main driver is the challenge and joy of building it myself (an obviously learn).
The key features I'm aiming for are:
Automatic identification and tagging of family members (local face recognition).
Generation of descriptive captions for each photo.
Natural language search (e.g., "Show me photos of us at the beach in Luquillo from last summer").
I've already prompted AI tools for a high-level project plan, and they provided a solid blueprint (eg, Ollama with LLaVA, a vector DB like ChromaDB, you know it). Now, I'm highly interested in the real-world human experience. I'm looking for advice, learning stories, and the little details that only come from building something similar.
What tools, models, and best practices would you recommend for a project like this in 2025? Specifically, I'm curious about combining structured metadata (EXIF), face recognition data, and semantic vector search into a single, cohesive application.
Any and all advice would be deeply appreciated. Thanks!
We no longer are auto uploading to Google or Apple.
So far, I really like it. I haven't quite gone 100%, as we're still uploading with Synology's photo app, but Immich provides a much more refined, featured interface.
I think you can use Immich to just look at a folder and not use the backup from phone bits.
I’m running a DS1813+. It’s stopped getting new feature updates. This approach lets me keep the storage running while migrating away the server components.
May I ask why? Just curious as the main reason I use Immich is for the auto upload
Edit: Ugh. Can’t read. I somehow read don’t auto upload to Immich.
No hate here, I'm really grateful for what they've achieved so far, but I think there's a lot of room for improvement (e.g: proper R/W query split, native S3 integration, faster endpoints, ...). I already mentioned it in their channel (they're a really welcoming community!) and I'm working on an alternative drop-in replacement backend (written in Go) [1] that will hopefully bring all the needed improvements.
TL;DR: It's definitely good, especially for an open-source project, and the team is very dedicated - but it's definitely not Postgres-good
A lot of existing tooling supports the s3 protocol, so it would simplify the storage picture (no pun intended).
Near zero maintenance stack, incredibly easy to update, the client mobile apps even notify you (unobtrusively) when your server has an update available. The UI is just so polished & features so stable it's hard to believe it's open source.
Is it really that stable and flawless in terms of updates?
Because I'm sat here with ZFS, snapshotting and replication configured and wondering why people scare others off of it when the tools to mitigate issues are all free and should be used anyway as part of a bog-standard self-hosted stack.
Given how good the new multimodal models are, I've been thinking it would be much better to just have a multimodal model describe the image, and let the searching be done by the already included melleisearch.
That said, due to reasons I haven't had time to mess with it past couple of months, so perhaps something drastic has changed.
The dev is really reluctant of accepting external contributions, which has driven away a lot of curious folks willing to contribute.
Immich seems to be the other extreme. Moving really fast with a lot of contributors, but stuff occasionally breaks, the setup is fiddly, but the Ai features are 100x more powerful. I just don't like the ui as much as photoprism. I with there was some kind of blend of the two, on a middle ground of their dev philosophies.
While I really like it — snappy and encrypted — I was surprised by how much the missing Ultra HDR implementation affects me. Photos are currently uploaded with brightness information but not displayed with it. Therefore, my photos look great in Google Photos but far less vivid in Ente.
For what it's worth, I found a discussion about Ultra HDR. It doesn't seem to be a priority right now, though: https://github.com/ente-io/ente/discussions/779
No. (I self-host Ente and use their published ios app.)
2. The software is provided without modification; I think it would be stranger to remove the encryption.
This is exactly how I self-host Ente and it has been great.
Machine leaning for image detection has worked really well for me, especially facial recognition for family members (easy to find that photo to share).
I have the client on my Android mobile, Fire tablet (via F-Droid), and my Windows laptop.
My initial motivation was to replace "cloud" storage for getting photos copied off the phone as soon as possible.
Also, my house is less secure than commercial data centers, so e2ee gives me greater peace of mind about data safety.
I must be wasting so much storage on the 4 photos I took in a row of the family pose, or derivatives that got shared on whatsapp and then stored back to my gallery, and so on, and I know I'm not the only one.
As of now, I use SentenceTransformer model to chunk files, blip for captioning (“Family vacation in Banff, February 2025”)) and mtcnn with InsightFace for face detection. My index stores captions, face embeddings, and EXIF metadata (date, GPS) for queries like “show photos of us in Banff last winter.” I’m working on integrating ChromaDB for faster searches.
Eventually, I aim to store indexes as:
{
"filename": "/Vacation/Banff/Wife.jpg",
"chunk_id": 0,
"text": "Family at Banff, February 2025",
"caption_embedding": [0.1, 0.2, ...],
"face_embeddings": [{"name": "NT", "embedding": [0.3, 0.4, ...]}, ...],
"exif": {
"DateTimeOriginal": "2025:02:15",
"GPSCoordinates": "18.387, -65.992"
}
I also built an UI (like Spotlight Search) to search through these indexes.
Code (in progress): https://github.com/neberej/smart-search
I've used gemma to process pictures and get descriptions and also to respond questions about the pictures (eg. is there a bicycle in the picture?). Haven't tried it for face recognition, but if you already have identified someone in one photo, it can probably tell you if the person in that photo is also in another photo
Just one caveat, if you are processing thousands of pictures, it will take a while to process them all (depending on your hardware and picture size). You could also try creating a processing pipeline, first extracting faces or bounding boxes of the faces with something like opencv, and then passing those to gemma3
Please post repo link if you ever decide to open source
And for sure, if I get this to a point where it's open-source, I'll post the link here!
I focused more on fast rendering in [photofield] (quick [explainer] if you're interested), but even the hacked up basic semantic search with CLIP works better than it has any right to. Vector DBs are cool, but what is cooler is writing float arrays to sqlite :)
[deepface]: https://github.com/serengil/deepface
[photofield]: https://github.com/SmilyOrg/photofield
[explainer]: https://lnar.dev/blog/photofield-origins/
I'm using docker compose to include some supporting containers like go-vod (for hardware transcoding), another nextcloud instance to handle push notifications to the clients, and redis (for caching). I can share some more details, foibles and pitfalls if you'd like.
I initiated a rescan last week, which stacks background jobs in a queue that gets called by cron 2 or 3 times a day. Recognize has been cranking through 10k-20k photos per day, with good results.
I've installed a desktop client on my dad's laptop so he can dump all of the family hard drives we've accumulated over the years. The client does a good job of clearing up disk space after uploading, which is a huge advantage in my setup. My dad has used the OneDrive client before, so he was able to pick up this process very quickly.
Nextcloud also has a decent mobile client that can auto-upload photos and videos, which I recently used to help my mother-in-law upload media from her 7-year-old iPhone.
Gonna check the apps that you mentioned. Feel free to share more details of your set up. Why are you running 2 instances? Edit: I see, probably for the memories app.
I have an OpenMediaVault VM with a 10tb volume in the network that runs the S3 plugin (Minio-based) which is connected through Nextcloud's external storage feature (I want to migrate to Garage soon). I believe notify_push helps desktop clients cut down on the chatter when querying the external storage folder. Limiting the users that can access this also helps.
I was having issues getting the notify_push app [1] to work in the container with my reverse-proxy. I found some similar setups that did this [2], so I added another nextcloud container to the docker-compose yaml like so:
notify_push:
image: nextcloud
restart: unless-stopped
ports:
- 7867:7867
depends_on:
- app
environment:
- PORT=7867
- NEXTCLOUD_URL=http://<local ip address of docker server>:8081
entrypoint: /var/www/html/custom_apps/notify_push/bin/x86_64/notify_push /var/www/html/config/config.php
volumes:
- /path/to/nextcloud/customapps:/var/www/html/custom_apps
- /path/to/nextcloud/config:/var/www/html/config
[2] - https://help.nextcloud.com/t/docker-caddy-fpm-notify-push-ca...
For Features. I dont know why there's isn't a tag for Screen Caps. I made lots of them and I want to group them together.