Show HN: Smart Photo Finder – Semantic photo search, not filename(100% local AI)

2 points

6 months ago

I got frustrated scrolling through thousands of photos to find "person holding camera" or "sunset on beach." Filenames are useless, manual tagging never happens, and cloud solutions compromise privacy.

Built this using NexaAI's local VLM (Qwen3-VL) + semantic embeddings. How it works:

1. VLM generates natural language descriptions of each image (one-time processing) 2. Descriptions converted to 384D embeddings 3. Search by meaning using cosine similarity (<100ms for 1000 images)

Everything runs locally. Your photos never leave your device. Zero API costs.

Technical reality: CPU processing is slow (~20-30s per image initially), but searching is instant afterward. JSON database works fine for personal collections; you'd want FAISS for 10k+ images.

It's a prototype, but it works and solves a real problem I had. Open to feedback.

Built for NexaAI's Builder Bounty Program (on-device AI, privacy-first).

Demo video: https://youtu.be/YVkPa-aJpEo Medium writeup: https://medium.com/@pankajgoyal4152/building-smart-photo-fin...