Michele LaurelliBlog

Classifier-free guidance

Classifier-free guidance: Guidance that combines conditional and unconditional predictions to increase prompt adherence. It requires no separate classifier: the model learns both modes during training. Excessive scale can reduce diversity and introduce artifacts. Examples include Prompt and empty prompt, Guidance scale 7.5, Visual fidelity control.

In italiano: Classifier-Free GuidanceLegge II · Probabilità, softmax e modelli generativi

It requires no separate classifier: the model learns both modes during training. Excessive scale can reduce diversity and introduce artifacts.

Formula

ε_guided = ε_uncond + s(ε_cond − ε_uncond)

Examples

  • Prompt and empty prompt
  • Guidance scale 7.5
  • Visual fidelity control