← AI Terminology
Classifier-Free Guidance
Classifier-free guidance (CFG) steers diffusion sampling by extrapolating between conditional and unconditional score estimates, amplifying adherence to the condition (e.g. text) without a separate classifier.
It is the standard control knob for prompt strength in modern diffusion.
It is the standard control knob for prompt strength in modern diffusion.
Why It Matters in AI
CFG transformed text-to-image quality by trading diversity for prompt fidelity with one scale parameter. Every T2I user and paper tunes guidance strength.
Key Points
| Aspect | Description |
|---|---|
| Use | Stable Diffusion, Imagen-style systems |
| Origin | Ho & Salimans, classifier-free guidance |
| Related | CLIP guidance (older), ControlNet |
| Scale s | Higher → stronger prompt match, less diversity/artifacts risk |
| Training | Randomly drop condition so model learns both |
| Mechanism | ε̃ = ε_uncond + s(ε_cond − ε_uncond) |
Simple Analogy
Mixing a pure daydream with a tightly briefed illustration, then pushing past the briefed mix to overshoot toward the brief — volume knob on obedience.
Common Usage Examples
- Set CFG 3–12 in SD UIs and watch fidelity
- Too high CFG → oversaturated artifacts
- Papers report CFG scales with samples
- Combine CFG with ControlNet carefully
Summary
In short: Classifier-free guidance amplifies conditional diffusion signals — the standard slider for how strongly images follow prompts.