'Text-to-image' systems like Stable Diffusion really are Large Language Models: an encoder such as BERT creates a mapping from text tokens to the latent space o

'Text-to-image' systems like Stable Diffusion really are Large Language Models: an encoder such as BERT creates a mapping from text tokens to the latent space of the image generation model. As part of this training step, the system is learning the concepts of certain words and grammatical constructs. There are quite a few in-depth explanations of the whole system; here's one for instance: https://jalammar.github.io/illustrated-stable-diffusion/ https://jalammar.github.io/illustrated-stable-diffusion/