Latent Space in plain English.

Latent space is the internal, abstract mathematical space inside a trained AI model where it represents the meaning of whatever it has processed — words, pixels, sounds — as points or directions rather than as the original text or image. It's closely related to embeddings: an embedding is a specific point plotted in latent space. What makes latent space powerful is that meaningful relationships become geometric ones — in a well-trained model's latent space, moving in a certain direction might correspond to "make this more formal" or "make this image brighter," and similar concepts cluster near each other. Image generators like Stable Diffusion literally generate in a compressed latent space before converting the result back into pixels, which is faster than working with full-resolution images directly. Latent space is largely invisible to end users but is the conceptual territory researchers actually work in when they talk about what a model "understands."