An Android gallery that understands 'sunset by the sea'

What it does

PicQuery is an Android app that indexes your local photos into CLIP embeddings and lets you retrieve them with natural language like “a laptop on the desk” or “kitty in the grass.” It also supports querying by selecting a reference image from your gallery, and keeps the entire index on the device.

The interesting bit

Instead of calling a cloud API, the app runs quantized CLIP-family models—OpenAI CLIP, Apple MobileCLIP, or TensorFlow Lite variants—directly on the phone, scoring text-to-image similarity locally. That trades a hefty initial indexing pass for sub-second search across thousands of photos with no network traffic.

Key highlights

• Fully offline operation: no cloud uploads, no in-app purchases, and no network required.
• Bilingual support: accepts natural-language queries in both English and Chinese.
• Swappable backends: switch between CLIP, MobileCLIP, or TFLite by changing the module and bundling the matching quantized model files.
• Fast local retrieval: the README claims sub-second results when searching across 8,000+ photos after the first-launch index is built.
• Built with Jetpack Compose and Material Design 3.

Caveats

• Building from source is a hands-on affair: you must source quantized model files, drop them into app\\src\\main\\assets, and select the correct backend module or the app will crash on launch.
• The README dedicates an FAQ to common build misconfigurations, suggesting the setup is finicky enough to trip people up regularly.
• It ships as a standalone app rather than a reusable library, so adopting the logic means extracting it from the project structure.

Verdict

Worth a look if you need a working example of on-device, CLIP-powered image retrieval on Android, or if you simply want a gallery that respects your privacy. Look elsewhere if you need a polished, drop-in SDK with no assembly required.