links: Immich | Immich repo
I spent a few days using AI to sort out my photo library in Immich. Years of phone backups, scanned photos, screenshots, receipts, and random images had piled up. Having everything backed up was one thing. Actually finding things was another.
I used Python scripts and a vision model to classify roughly 25,000 images and build 12 albums. Immich’s search, text recognition, and face search narrowed down the candidates first, then the model looked at the images to decide whether they belonged. That worked better than keyword matching alone, which was too loose for separating screenshots and other categories.
The albums cover things like artwork, cars, gym photos, documents, office equipment, and different kinds of screenshots. They are also useful as review queues. The classifier puts images somewhere I can inspect them; it does not automatically trash anything.
Dates were another big part of the cleanup. I corrected 4,570 asset dates, using the old year and month folders as evidence first. One scan batch had given around 1,100 older photos the same date. Where reliable dates were missing, visual clues helped estimate them. Those are estimates, not exact dates recovered by AI.
I also dropped a custom organizer that maintained a separate tree of category folders and links. It added complexity I did not need. Immich now handles the albums, and its storage template keeps the files arranged by year and month.
A couple of details mattered more than I expected: midnight UTC placeholder dates could land in the previous month locally, and the metadata sidecar files needed to stay so later metadata extraction would not undo the date fixes.
There is still duplicate review to do, but the library is much easier to work with now. This was a useful change of pace from using AI to write code. The scripts were just the means to get years of photos into a more usable state.