What is Classifier-Free Guidance?

How it works

Classifier-free guidance works by generating two outputs at each step of the diffusion process: one conditioned on your prompt and one unconditional (without any prompt). The final output is computed by amplifying the difference between these two outputs by the CFG scale factor. Higher CFG values amplify the prompt's influence more, producing output that more strongly reflects your text description. This is computationally expensive (essentially doubling the work per step) but produces significantly better prompt adherence than conditioned generation alone. Most tools handle this automatically, but understanding the mechanism helps explain why very high CFG values can produce over-saturated or artifact-heavy results.