Continual CLIP

A CLIP-like system that adds visual concepts directly in latent space from a handful of images without updating model weights.