Classifier-Free Guidance (CFG)
CFG makes diffusion models obey prompts by a training trick and a sampling trick: randomly drop the condition so ONE network learns both "denoise with the prompt" and "denoise without it", then amplify the difference between the two predictions at every step. The guidance scale w is the fidelity-vs-diversity dial behind every guidance_scale and negative_prompt in production - and 2024-era distillation folds its 2x cost back into one pass.
01.The Problem: The Model Hears the Prompt, But Barely Listens
You trained a conditional diffusion model (topics 187-189): it takes a text embedding c and denoises. You type "a red fox in snow" and... you get a blurry white nothing with maybe a reddish smear. Technically on-distribution. Practically useless.
Why? Training maximizes the average fit. Most of the loss is paid by learning what images look like in general; the prompt only nudges the answer slightly. The model is a people-pleaser with a shrug.
So the question becomes
Can we take the tiny "the prompt pushed me this way" component of the prediction — and deliberately overshoot in that direction?
Classifier-Free Guidance (Ho & Salimans, 2022) does exactly that, and it is arguably the most-used sampling-time technique in all of diffusion. Every guidance_scale and negative_prompt box you have ever seen in Stable Diffusion is CFG.
The analogy to carry: two tour guides in a strange city.
- Guide U wanders wherever — the average tourist route through the city.
- Guide C heads toward the museum (the condition).
- Neither guide is fast. But the difference between their two directions tells you exactly which way the museum is.
- CFG is: walk in that difference-direction at a run, not a stroll.
Interpolating Conditional and Unconditional Predictions
Interpolating Conditional and Unconditional Predictions
The same network is run twice per step - once with the condition and once with a null token. Their difference is amplified by the guidance scale w and added back to the unconditional prediction, pushing the sample further toward the condition than plain sampling would.
Unlock Topic #191: Classifier-Free Guidance (CFG)
You are viewing a preview. The full in-depth technical walkthrough, worked derivations, and code notebooks for this concept, along with self-assessment quizzes, are available with Pro or Lifetime Access.
Failure modes, high-throughput bottlenecks, and real FAANG implementation decisions.
Interactive system topology diagrams, live parameter simulators, and downloadable SVG charts.
Staff-level multiple-choice quiz questions with instant feedback and answer explanations.
Firebase Google authentication automatically syncs your completed topics and quiz scores.
How clear and actionable was this distributed systems breakdown?