Term · Advanced
Latent Space
The compressed abstract space where a model represents concepts before rendering output.
Latent space is the internal, compressed map of concepts a generative model works in, points in this space correspond to images or sounds, and nearby points are similar. Diffusion models generate by navigating and denoising within it rather than working on raw pixels. Understanding it helps explain why seeds, interpolation, and prompt tweaks behave the way they do.
Also called latent representation, latent
Related
- Diffusion ModelThe AI architecture behind most image and video generators, it denoises randomness into a picture.
- SeedThe starting number that determines an AI image's random noise, reuse it to reproduce results.
- Img2ImgGenerating a new image guided by an input image plus a prompt.
- Prompt EngineeringThe craft of writing inputs that steer an AI model toward the result you want.
- Fine-TuningFurther training a base model on your own data to specialize its output.