What it does

TIDY indexes every photo on your Android device using a quantized CLIP model, then lets you search your own library with natural language or by showing it another image. The indexing happens once at first launch; new photos get picked up automatically afterward. Everything stays on-device — no cloud, no accounts, no connectivity required.

The interesting bit

The trick is making a ~2B-parameter vision-language model run comfortably on a phone. The author uses ONNX Runtime with quantization to squeeze CLIP (specifically an OpenCLIP variant trained on LAION-2B) into something that can batch-process your camera roll without melting the battery. The result is genuine semantic search — “dog looking guilty” works, not just filename matching.

Key highlights

• Pure offline operation: no network calls, no data exfiltration
• Text-to-image and image-to-image retrieval in one app
• Automatic incremental indexing after initial scan
• Distributed via GitHub releases and F-Droid
• Built in Kotlin around ONNX Runtime inference

Caveats

• Initial indexing time is unspecified but implied to be nontrivial for large libraries
• Only Android; no iOS or desktop builds mentioned
• Model specifics (size, speed, memory footprint) aren’t quantified in the README

Verdict

Grab it if you’ve got 10,000 unlabeled photos and a patience for first-run indexing. Skip if you’re hoping for cloud-backed cross-device sync or fine-grained EXIF filtering — this is search by semantic embedding, not by metadata.