Definition
Plain language
A trick used by image-generation models to control how strongly the output follows the prompt.
As stated in the literature
Classifier-Free Guidance, a diffusion-model inference technique that interpolates between conditional and unconditional score predictions to tune prompt adherence.
Also called: classifier-free guidance
Why it matters: It's the main knob that lets users trade off prompt adherence against image diversity and naturalness in diffusion models.
For example, turning the CFG scale up makes a text-to-image model stick closer to 'a red bicycle on a beach' instead of drifting toward a generic outdoor scene.
Heard on the show
“… the DiT — you train it with something called flow matching, and you steer generation with classifier-free guidance. …”Episode 127 — What Diffusion Language Models Were Missing: A Map, Not an Algorithm