Guidance Scale
A hyperparameter controlling how strongly the model follows a conditioning signal (like a text prompt) during generation.
Think of it like
Leash length: 0 = completely free dog (ignores you), 7 = dog paying attention, 15 = dog tugging hard (over-correcting).
Example
In Stable Diffusion, guidance scale 7 balances creativity and prompt adherence. Scale 15 follows the prompt rigidly, producing bland or distorted outputs. Scale 0 ignores it entirely.
How it actually works
Classifier-free guidance blends unconditional and conditional denoising predictions: pred = pred_uncond + scale x (pred_cond - pred_uncond). Scale = 0 -> unconditional. Scale = 1 -> conditional. Scale > 1 -> over-emphasis of conditioning. High scales sharpen the prompt but can create artifacts (oversaturation, impossible geometry). Optimal scale depends on model, prompt complexity, and desired creativity-faithfulness tradeoff.
For product teams
Lets users trade off between following their prompt precisely and creative variation.
For engineers
Typical range: 7-15 for text-to-image. Experiment per model. Trade-off: scale vs. sample quality.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome