# Photo Search

> Search a folder of your own photos by describing what is in them ("a dog on the grass", "sunset over the water") or by a look-alike photo, and find duplicate and near-duplicate shots. Image models run in the browser; the index and small previews stay in the browser's own storage. Nothing is uploaded. Free, no account, no API key.

URL: https://photo-search.skillsafe.ai/

## What it does

- Input: a folder (or a selection) of photos: JPEG, PNG, WebP, GIF, BMP, AVIF. HEIC and camera RAW files are skipped and listed.
- Search by description: type what you remember; every photo gets a 0-100 match score; a minimum-match slider hides weak matches.
- Search by photo: pick a reference photo from your device, or press "Similar" on any result.
- Duplicates: a 64-bit layout fingerprint per photo groups copies and resized or re-saved versions; optionally also look-alikes (different shots of the same scene). The largest copy in each group is marked Keep. Nothing is deleted.
- Export: results and duplicate groups as CSV.
- Library: kept in this browser (IndexedDB), so later visits search instantly; pick the folder again to add new photos or open full-size originals.
- Free example: 18 CC0 photos from Wikimedia Commons with their index precomputed, so trying it downloads no model.

## How it works

- Meaning search: CLIP ViT-B/32 (OpenAI, MIT licence; ONNX int8 export), image and text towers, 512-number vectors, cosine similarity. It switches on automatically when its weights are served by SkillSafe's model host.
- Quick search (used until then): ResNet-50 ImageNet-1k (Apache-2.0) names 1,000 everyday things in each photo; all-MiniLM-L6-v2 (Apache-2.0) matches the words of your query to those names; colour words use each photo's colour shares. The page says which words it understood and which it did not.
- Runtime: onnxruntime-web (MIT), WebAssembly. Weights are downloaded once from models.skillsafe.ai, checked against their SHA-256 and kept in the browser cache.
- Privacy: no server step. Photos, previews and the index never leave the device.

## Limits

- Text search understands English best.
- Quick search only knows the 1,000 ImageNet categories plus colours; scenes like "sunset" or "birthday" need meaning search.
- Indexing speed depends on the device: roughly 5-15 photos a second on a laptop.