← All demos
Computer Vision
Zero-Shot Image Classification (CLIP)
Type your own labels and CLIP scores the photo against them, no training, in your browser.
Production proof: Open-vocabulary vision on-device. From my edge-vision work: flexible perception without retraining a model for every new class. Read the case study →
How it’s engineered
- CLIP (~120 MB) embeds the image and your text labels into the same space and scores their match, entirely in your browser. No training, no upload, no key.
- Zero-shot means you classify into any labels you invent at runtime, not a fixed list baked in at training time. That flexibility is what makes CLIP useful in production.
- The same technique powers visual search and content moderation where the categories change faster than you could retrain a classifier.