guidance scale

The guidance scale is the single user-facing knob — usually labeled 'CFG scale' in tools — that controls how literally the model obeys your prompt versus how freely it invents. Turn it low and the model treats the prompt as a loose suggestion, producing diverse but sometimes off-topic images. Turn it high and the model clamps hard onto every word of the prompt, producing on-topic but increasingly stylized, contrasty, and less varied images. It is the practical dial on the abstract weight w in classifier-free guidance.

In the CFG formula epsilon_guided = epsilon_u + w * (epsilon_c - epsilon_u), the guidance scale is exactly w. A value of 1 means no amplification (plain conditional sampling); values above 1 extrapolate beyond the conditional prediction in the prompt-aligned direction. There is a genuine tradeoff captured by quality metrics: increasing w typically improves precision/prompt-fidelity (often measured by CLIP score, how well image matches text) while reducing recall/diversity (often measured by FID, how well the set of generations matches the real distribution). The best setting is a balance, not a maximum.

Typical defaults reflect this balance: Stable Diffusion commonly uses around 7.5, Imagen and similar systems report sweet spots in a comparable range, and very high values (15+) are reserved for cases where strict prompt adherence outweighs naturalness. Because the ideal scale depends on the model, the schedule, the sampler, and even the specific prompt, it is one of the first parameters practitioners sweep when results look either too generic (raise it) or too garish and warped (lower it).