What is classifier-free guidance (CFG) in diffusion?

hard

Answer

$Train the model to accept both conditional (c) and unconditional (\emptyset, e.g., empty prompt) inputs, by dropping the condition 10-20% of the time.$
$\mathrm{eps}_{\mathrm{guided}}\; = \;\mathrm{eps}$
$Standard trick in Stable Diffusion / Imagen — biggest single quality lever after model size.$

Check yourself — multiple choice

CFG requires a classifier
Train with occasional empty conditioning; at sampling, extrapolate: $\mathrm{eps}_{\mathrm{uncond}}\; + \;w$ · $(\mathrm{eps}_{\mathrm{cond}}\; - \;\mathrm{eps}_{\mathrm{uncond}})$ → stronger prompt alignment
Removes conditioning
Same as DDIM

CFG: extrapolation between cond/uncond predictions — huge quality boost for text-to-image.

#diffusion#generative

214 interview questions in this topic.