潜空间

The low-dimensional semantic space into which data is compressed and encoded.

Models map raw data (pixels, text) into a latent space where similar semantics sit closer together. Latent diffusion generates images in this space, and embeddings are essentially text mapped into it.