Classifier-Free Guidance

Classifier-Free Guidance linearly extrapolates between a conditional and an unconditional prediction so that sampling amplifies the condition $c$. During training the condition is dropped with probability $p$ (CFG dropout, usually 10–15%) and replaced by the empty token $\emptyset$, so one network learns both $\epsilon_\theta(x_t, c)$ and $\epsilon_\theta(x_t, \emptyset)$. At sampling the two noise predictions are combined with strength $w$:

$w=1$ reduces to plain conditional sampling; larger $w$ pushes the trajectory toward the conditional distribution (stronger prompt adherence) at the cost of diversity and oversaturated colors; typical values are $w = 7\text{–}8$ for Stable Diffusion and $w = 3\text{–}5$ for video/cascaded models. It is a score reweighting: it matches the behavior of classifier guidance $\nabla_{x_t}\log p(x_t) + \gamma\nabla_{x_t}\log p(c \mid x_t)$, pushing the score toward conditional likelihood, but needs no extra classifier.

prompt adherence in text-to-image/text-to-video (enabled by default in SD/Flux); a classic interview question — “how is CFG implemented, what w is typical, and why does large w oversaturate and reduce diversity?”.

plain conditional diffusion gives weak text control and off-prompt samples; classifier guidance needs a separately trained discriminator that is sensitive to adversarial noise; CFG needs one training run (dropout) and two forward passes (conditional + unconditional) at inference — no extra model — while substantially improving condition fidelity.

Write the CFG sampling formula and explain every symbol.

What is the typical CFG dropout probability, and how does one network learn both branches?

How does increasing w affect adherence, diversity and oversaturation, and why?

How does classifier-free guidance relate to classifier guidance; pros and cons of each?

What do w=1 and w=0 correspond to, and what is CFG's extra inference cost?