Earngenix Logo
Skip to main content

Glossary · ComfyUI Basics

What Is CFG Scale in ComfyUI?

By Earngenix Team ·

⚡ Quick Answer

CFG scale (Classifier-Free Guidance) controls how strictly ComfyUI follows your prompt versus how much creative freedom the model takes. Higher CFG means stricter prompt-following but can burn or oversaturate the image; lower CFG means more creative freedom but a looser match to what you typed.

Think of it as a volume knob for "listen to my prompt." Turn it up and the model pays closer attention to every word — turn it up too far and details start looking harsh, overexposed, or exaggerated.

Where You'll See It

The cfg field sits inside the KSampler node, alongside steps, sampler_name, and scheduler. Most checkpoints default to a value around 7–8, but the ideal range depends heavily on the model.

Quick Example

For most SD1.5 and SDXL checkpoints, a CFG between 4 and 12 is typical, with 6–8 being a solid starting point. Newer distilled or "turbo" models — like some FLUX and SD3 variants — bake guidance behavior directly into training, so they often expect CFG values near 1 and can look worse at the higher values older models want.

If your image looks crunchy, overexposed, or unnaturally sharp, lower the CFG value before touching anything else. That's the most common cause.

Frequently Asked Questions

For most SD1.5 and SDXL checkpoints, 6–8 is a safe starting range. Check the model card for anything newer or distilled — recommended values vary a lot by architecture.

CFG is likely set too high. Pushing the model to follow your prompt too aggressively often produces harsh contrast, blown-out colors, and exaggerated detail.

No. Some newer distilled or "turbo" models bake guidance behavior directly into training and expect CFG near 1, or ignore it almost entirely. Check the model’s documentation before assuming 7–12 is safe.

See It In Action

Ready to dial in your first generation?

Our text-to-image guide shows exactly how CFG interacts with sampler and steps, with real before/after examples.

Published: 2026-08-25 · Last updated: 2026-08-25

Discussion

Join the discussion

Sign in to leave a comment or reply

💬

No comments yet

Be the first to share your thoughts!