How Does Clip Work?

Clip was trained on a very large data set, which consisted not just of images, but of also textual descriptions. And so clip is trained in order to basically grade the extent to which an arbitrary image and an arbitrary text input are the same. If you have a text cat running through a yard and an image of the same thing, it should give you a high score. But if you have a completely irrelevant text input, it'll give you a low score. So basically what you could do is use clip as a guide, basically in combination with another model that produces images.