CFG scale

in English: classifier-free guidance (CFG) scale

advancedwidespread

The CFG scale is a configurable number that determines how closely an image-generating AI follows your text prompt and how much freedom it is granted.

In plain terms

When you describe what you want to see to an image generation program, the machine does not execute the request literally: it has some room for maneuver. The CFG scale is the slider with which you control this room for maneuver. At low values, the model interprets the description loosely; at high values, it adheres to it.

It is like giving an order to a painter. If you say, 'I want something beach-related, the rest is up to you,' it is natural, but you get an unpredictable image. If you demand every detail, you get exactly what you asked for, but the result will be rigid and overdrawn.

The good setting lies between the two. At too low a value, the image is blurred and barely resembles your request; at too high a value, the colors are garish, the contrast is harsh, and details are distorted. Therefore, most programs have a proven default value to start with.

Not every service displays this number. Image generators built into chatbots mostly hide it, while in open, locally runnable tools it is one of the most important adjustable values.

How it works

A generative image model starts from noise and refines the image through many small steps.
At each step, it makes two estimates: one taking your text description into account, and one without it, as if you had requested nothing.
The difference between the two estimates shows which direction your description 'pulls' the image.
The CFG scale multiplies this difference: the higher the number, the more strongly the model pushes the image toward the description.
Since two calculations are needed per step, this method roughly doubles the workload; its effect is often pre-built into accelerated models.

An everyday example

In an image generation program, you ask: 'a red fox in a snowy forest, watercolor style.' With a low CFG value, you get an atmospheric but inaccurate image where the fox could even be brown; with a very high value, the fox is bright red and the snow is blinding, but the whole thing looks artificial and oversaturated.

Who is affected

marketerWhen generating images for a campaign, this setting lets you choose between exact matching or a more natural, freer visual style. → Marketing

developerIn an application incorporating image generation, this is one parameter you must choose well, because its value affects image quality, and enabling the method roughly doubles computing costs. → Code

parent, student, teacherFor school projects or images made as a hobby, this explains why the same description sometimes produces a blurred image and other times a vivid one. → Education & Family

A common misconception

MythThe higher you set the CFG scale, the better and more accurate the image becomes.

In realityBeyond a certain point, image quality degrades: colors become oversaturated, details distort, and variety decreases. The best results usually come from the mid-range.

Limits and risks

There is no universally good value: what works for one model may produce unusable images in another, so always start from the specific model's recommendations.
High value does not replace good description: if your request is inaccurate or contradictory, the model will only follow the poorly phrased instruction more strongly.
The scale does not fix the model's fundamental weaknesses, such as malformed hands or nonsensical text appearing in the image.
In many convenience services it cannot be adjusted at all, so the provider decides for you.

Similar, but not the same

Temperature – Temperature controls the randomness of text model responses, while the CFG scale determines how strongly the image should adhere to the description.
negative prompt – The negative prompt specifies what the image should not contain, while the CFG scale determines how strongly your instruction is enforced. The two are related: the negative prompt relies on the same mechanism, so the scale also amplifies the effect of the negative prompt.
sampling step count – The step count specifies how many rounds the image is refined, while the CFG scale determines how much the description matters in each round.

Related concepts

Broader

Diffusion model, Image generation, Generative AI

Narrower

–

On the edge

negative prompt, Prompt, Temperature, sampling step count, Seed, Stable Diffusion

A brief history

The method was described by Google researchers Jonathan Ho and Tim Salimans in their paper titled Classifier-Free Diffusion Guidance. It was first presented at a NeurIPS workshop conference at the end of 2021 and uploaded to arXiv in July 2022. The name indicates that, unlike previous solutions, no separate classifier model is needed to guide the image.

Where things stand

By the end of 2026, it is a basic tool of diffusion image and video generators and can be freely adjusted in open tools. Accelerated models that work in few steps, however, increasingly have it built in, so users less often encounter it directly.

Status: 2026 November

Also known as

CFG · CFG scale · CFG value · guidance scale · classifier-free guidance · classifier-free guidance · control scale · prompt adherence strength

Sources

Ho, J. – Salimans, T.: Classifier-Free Diffusion Guidance (arXiv:2207.12598)

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.