CFG Scale Stable Diffusion Meaning: Guide + Examples

Cole B.
March 18, 2024

Today, I am going to explain the CFG scale stable diffusion meaning and how to use it effectively.

Learning to master this setting is crucial to producing high-quality images.

So if you want to learn how to use this setting adequately, this guide will provide you with everything you need.

What Is The CFG Scale Stable Diffusion Meaning?

The CFG scale (classifier-free guidance scale) determines how closely generated images follow prompts in stable diffusion.

If you increase the guidance scale value, then the generated images should more closely resemble the prompt. Low CFG values give generated images more creativity and diversity but don’t closely follow the prompt.

Deciding what value to use will be important for every image you produce because it will greatly affect the generated images.

How Does the Guidance Scale Effect Images?

The guidance scale will effect your images differently depending on the value used and other factors like sampling method, and sampling steps.

These are the common effects and use cases at different values:

1-5: At very low values, lets say one or two, images will come out monotone and desaturated. While three to five is usually a sweet spot for realistic images that somewhat follow the prompt.
6-10: This range can be decent for certain styles and prompts. On the lower end, realistic prompts are viable, although they will be somewhat saturated looking. On the higher end, prompts and styles that try to achieve cartoons or unrealistic images are viable.
11-15: This range is uncommonly used. Images will be saturated and usually bright-looking. These values might be useful for unrealistic images that really need to follow the prompt.
16<: Values this high usually affect the quality and make the image too saturated and exaggerated. Not recommended.

What CFG Scale Values Should You Use?

My recommendation:

2-5 values for realistic images
2-10 values for unrealistic images (cartoon, anime, graphics)

The reason why this works:

Realistic images get hurt by the saturated, brightness, and exaggerated nature of high values diminishing their authentic aesthetic.

Unrealistic images don’t need an authentic look, so they can have higher values, especially if you are having problems with images following the prompt.

Other Places The Guidance Scale is Used

Negative Prompts

Because the CFG scale affects positive prompts, it might be natural to think it has the same effect on negative prompts.

Although there is some confusion on this issue, increasing the CFG scale guides images further from negative prompts.

Meaning if you increase the guidance scale value, then stable diffusion will generate images that deviates from the negative prompt more.

Img2Img

The guidance scale will have an effect on images generated by img2img.

Although the inputed or original image will still have a huge factor in the generated image, increasing or decreasing the guidance value will still influence how much the output image follows the prompt.

Let’s look at an example:

This demonstrates the difference between a high and low guidance value in img2img. A high value ends up in an animated-looking image with brighter colors.

Hires.Fix

Typically hires.fix doesn’t have an option for the guidance scale. But with the extension, sd webui hires fix tweaks. You can have the options available to you:

Then you can have control over images generated by hires.fix more closely following the prompt or not.

What is the Difference Between the CFG Scale and Denoising Strength?

The difference between the CFG scale and denoising strength is:

The CFG scale influences how much the generated image follows the prompt.

Denoising strength controls how closely the generated image will resemble the original image.

When using img2img and hires.fix finding the right combination of both settings will be important to producing the image that you want.

CFG Values Comparisons and Examples

Now, I am going to demonstrate some comparisons and examples using the CFG scale and Web UI (Automatic1111).

This will help you understand the different effects of adjusting the guidance scale or with other settings in tandem so you can determine what is the best setting for you:

Low vs High Guidance Scale Comparison

Prompt: a wizard holding a potion, in a castle, realistic, 8k

Very clearly, increasing the setting beyond 15 creates a cartoony and saturated image, as I had previously warned. Although I still wouldn’t recommend having the setting that high.

Most importantly, the comparison demonstrates that the best values for realistic images are low (1-3), while animated images can have high and low values (1-15).

Comparison Of Sampling Steps And CFG Values

Prompt: a woman sitting down, with long black hair, in a cafe, photograph, realistic, high quality, detailed, 8k

Increasing the guidance scale creates a necessity to choose a higher value for sampling steps as well. On the other hand, it’s possible to reduce sampling steps with a low guidance scale value.

Higher CFG Scale Value = More Sampling Steps

Comparison Of Samplers And CFG Values

Prompt: a garden, flowers, sharp focus, dreamy, atmospheric, highly detailed

Some samplers have worse output with a higher CFG scale value (DPM++ SDE and Euler a), but overall, there isn’t much difference in how the scale interacts with different samplers.

Using Different Values With Low And High-Detailed Prompts

The next two tests are going to compare different CFG values for high-detailed prompts and low-detailed prompts to show how the prompt can influence your settings.

Example Of Using High-Detailed Prompts

Prompt: double exposure portrait of a beautiful woman with brown hair and a snowy tree under the bright moonlight by Dave White, Conrad Roset, Brandon Kidwell, Andreas Lie, Dan Mountford, Agnes Cecile, splash art, winter colors, gouache, triadic colors, thick opaque strokes, brocade, depth of field, hyperdetailed, whimsical, amazing depth, dynamic, dreamy masterwork

As you can see, the image looks much more artistic and conceptual as the cfg increases. This gives us the desired output because of the increased cfg value because the prompt was not originally meant to have a realistic image of a woman.

Although rare, this is an example of how a higher value would give us a better output than a lower value. Usually, this is only found with detailed prompts and a very specific output image in mind.

Example Of Using Low-Detailed Prompts

Prompt: sports car, poster, concept art

Using a low value added some creativity to the prompt with adding a building, clouds, and a marble looking floor.

But as the value is increased to 9 and above, everything is removed but the car itself. The prompt is being taken more literally and removes anything that isn’t being specifically mentioned in the prompt.

What Guidance Scale Value Works For You?

Now, I’d like to hear about your experience with the stable diffusion CFG scale:

Which guidance values work best when you generate images, and for what style?

Let me know by leaving a comment below right now.