← All models · CLIP

use case · basics

Basics: one image, your labels

This is the whole idea in its simplest form. Give CLIP a picture and a list of words. It tells you how well each word matches — even words it has never been "trained on," because to CLIP a label is just more text to compare against the image.

Drop an image, or click to choose
Two cats on a couch A cozy bedroom A riverside at golden hour
Selected image

Try a preset:

The bars are a softmax over the whole label set, so they always add up to 100%. Change the words and the same picture can "become" a different thing — that is zero-shot classification.

← Back to CLIP · Next: Practical →