Sort into categories
Classify photos, screenshots and videos into bucket folders you configure (people, food, nature, receipts, chat screenshots, memesβ¦), across a whole tree of subfolders into one flat set of buckets.
Privacy-first AI for your photo and video library.
Sort photos, screenshots and videos into folders, gather the ones from "the birthday party last Sunday", find anything with a plain-language search, and recognize the people in them, all with AI models running on your own machines. Nothing is sent to a cloud AI.
Classify photos, screenshots and videos into bucket folders you configure (people, food, nature, receipts, chat screenshots, memesβ¦), across a whole tree of subfolders into one flat set of buckets.
Ask in plain language, for example "birthday party on last Sunday". Matching photos and videos go into one folder; everything else stays where it is. Dates come from the files' capture time, not a guess.
Each file gets a local caption and embedding, so "receipt from the hardware store" or "kids on the boat" finds it. Browse results as thumbnails, and carry the index to another PC with export and import.
Match people in photos and videos against a few reference pictures per person, with no training.
Don't have references? faces-discover groups the faces in a folder into one folder per person
for you to name.
Run jobs with live progress, browse a media gallery, review every past run, watch the model cluster and privacy checks, and edit settings in a checked editor.
Preview first, then apply exactly what the preview found. Every move is journaled and can be undone, even after Ctrl+C. Files are moved byte-for-byte: never re-encoded, resized or stripped.
127.0.0.1.
127.0.0.1:11434 spreads requests across
your GPU machines.
gemma4:12b sorts and captions; qwen3-embedding:8b powers
search. Face detection (InsightFace) runs on LAN workers over SSH.
Videos are read as 8 evenly spread keyframes per request (faces use every keyframe). Audio is never used.
β¦-cloud names, or ones PAIR reports as remote) are
refused before any photo is sent.group in a public test setYour numbers depend on your GPUs and network. Details and methods are in the benchmark.
utils)gemma4:12b) and an embedding model (qwen3-embedding:8b) from your
machinesgit clone https://github.com/canvito/PhotoSort.git
cd PhotoSort
conda create -n utils python=3.12
conda run -n utils pip install -r requirements.txt
.\photosort.cmd check
.\photosort.cmd sort "<folder>" --action report # preview
.\photosort.cmd sort "<folder>" # move into buckets
.\photosort.cmd undo "<folder>\_photosort_journal.jsonl"
.\photosort.cmd group "<folder>" "birthday party on last Sunday"
.\photosort.cmd index "<folder>"
.\photosort.cmd search "kids on the boat"
.\dashboard.ps1 # desktop dashboard
Full setup, every option and troubleshooting: the operator guide.