What is classifier-free guidance (CFG) in image generation?
Computer VisionLevel: IntermediateGenerative Models
What is classifier-free guidance (CFG) in image generation?
Hints
- Identify the prediction unit: image, object, pixel or sequence.
- How do occlusion, lighting or domain changes affect results and evaluation?
Answer
CFG controls how strongly the generated image adheres to the text prompt vs unconstrained quality:
Training: randomly drop the conditioning (text prompt) with probability p. Model learns both conditional and unconditional generation.
Inference:
noise_pred = unconditional_pred + guidance_scale × (conditional_pred - unconditional_pred)
guidance_scale = 1: no guidance, creative but may ignore promptguidance_scale = 7-12: good balance (typical default)guidance_scale > 15: strictly follows prompt but may produce artifacts
Intuition: the difference between conditional and unconditional predictions represents what's "prompt-specific." Amplifying that difference makes the image more prompt-adherent.
Source and licence: JustInCache & contributors — Awesome AI Interviews · Apache-2.0