Note · 2026-06-12
How we catch images downloaded from stock libraries
Five layers. (1) EXIF strip detection — most downloaded images lack any metadata. (2) Qwen-VL authenticity prompt asks the model to name watermarks verbatim if it sees any. (3) Perceptual hashing against a corpus of known stock-image samples. (4) Reverse image search via a provider abstraction (Tineye, Bing, Google Vision). (5) Live-capture-only flow on supported channels.
We deliberately don't trust any single layer. Each one feeds a 0-1 confidence score; the aggregate goes to the adjuster as a flag, not a block. Even authoritative-looking 'this image was downloaded from Shutterstock' verdicts get a human review — because a partner NGO might legitimately upload a sample photo as a reference.