How to Use CFG Scale & Guidance in AI Images (2026)

If your AI-generated images keep ignoring half of your prompt — or come out looking like over-saturated, crispy-fried plastic — the culprit is often a single number: the CFG scale. Classifier-free guidance is one of the most powerful levers you have for AI image control, and it is also one of the most misunderstood. Push it up and the model obeys you more literally; push it down and you get creative, painterly, sometimes dreamlike results that ignore your instructions.

This guide explains what CFG scale and guidance scale actually do, how the setting differs across the major 2026 models, and exactly how to tune it for portraits, products, landscapes, and text rendering. By the end, you’ll know how to diagnose any image problem and fix it with a single slider.

What Is CFG Scale?

CFG stands for classifier-free guidance. During sampling, the model makes two predictions at every step: one with your prompt (the conditional prediction) and one with a blank or negative prompt (the unconditional prediction). The CFG scale determines how aggressively it pushes the final result away from the unconditional version and toward your prompt.

The math is simple: a CFG of 1 means “ignore the prompt almost entirely,” while a CFG of 15 means “obey the prompt at nearly any cost.” Most models live comfortably between 3 and 9.

CFG scale vs. guidance scale: are they the same thing?

In practice, yes — but the terminology matters when you switch tools.

• CFG scale is the classic term used by Stable Diffusion, SDXL, ComfyUI, and Automatic1111-style interfaces.
• Guidance scale is the term used by FLUX, many API providers, and newer diffusion transformers. It means the same thing but often runs on a lower numeric range.
• Distilled guidance appears in models like FLUX.1 schnell, where guidance is baked into training. Setting CFG to 1 is correct here — raising it will usually break the image.

When you see “guidance” instead of “CFG,” check the model card before assuming the scale is comparable. A guidance of 4 in FLUX is not the same as a CFG of 4 in SDXL.

Recommended CFG Ranges by Model

Defaults exist for a reason, and they shift as architectures evolve. Here’s a practical 2026 reference:

How to Tune CFG Step by Step

1. Start at the model default

Generate one baseline image. Never tune from a cold start — you need a reference to compare against. Save the seed so every change is the only variable.

2. Move in increments of 0.5 to 1.0

CFG is not linear in feel. A jump from 7 to 12 tells you almost nothing useful. Small steps reveal the exact point where composition locks in and texture quality starts to degrade.

3. Watch three signals

• Prompt adherence: Are all the elements present — the red jacket, the two birds, the correct sign text?
• Contrast and saturation: Higher CFG increases both, often unnaturally.
• Artifacts: Blown highlights, halos, duplicated limbs, and waxy skin are classic high-CFG symptoms.

4. Lock in and refine elsewhere

Once you find the range where prompt adherence is high and artifacts are absent, stop adjusting CFG. Improve the image with your sampler, step count, or prompt wording instead.

Low vs. High CFG: A Troubleshooting Guide

Use this as a quick diagnostic when an image isn’t working.

• Image ignores key details → raise CFG by 1–2 points.
• Colors look neon or burnt → lower CFG by 1–2 points.
• Skin looks plastic or waxy → lower CFG and increase steps.
• Composition is chaotic or surreal → raise CFG slightly, or add negative prompts.
• Text or logos are garbled → raise CFG modestly and use a model with strong typography training.
• Results feel generic or “stock” → lower CFG to let the model’s priors add flavor.
• Everything is a blurry mess → CFG is probably fine; check your step count and resolution instead.

How CFG Interacts With Other Settings

Sampler and steps

Higher CFG amplifies error as well as instruction. If you like a CFG of 10 for prompt accuracy, pair it with 30–40 steps to smooth out the noise it introduces. Fast samplers and high CFG are a poor match.

Negative prompts

Negative prompts scale with CFG. At CFG 4, a negative prompt has mild influence; at CFG 12, it can dominate and introduce artifacts of its own. If you’re using aggressive negatives, keep CFG moderate.

CFG rescale and dynamic guidance

Modern pipelines include tools to get the accuracy of high CFG without the contrast blowout:

• CFG rescale normalizes the guidance effect and restores natural contrast.
• Dynamic or scheduled CFG starts high for composition, then tapers low for texture. This is available in most advanced node-based interfaces.
• Distilled guidance in newer models removes the need for manual tuning on many prompts.

Practical Tips for Better AI Image Control

• Fix the prompt before the slider. If two concepts keep blending, clarify the wording before cranking CFG.
• Keep a personal settings log. Note model, CFG, sampler, and steps for images you love — your own data beats generic advice.
• Match CFG to intent. Product shots and infographics want higher guidance; concept art and painterly portraits want lower.
• Never copy CFG values across model families. The numbers are not portable.
• Test with a fixed seed. Random seeds hide the effect of your changes.
• Try mid-range first. If you’re unsure, CFG 5–6 with a strong prompt outperforms CFG 12 with a vague one almost every time.

Common Mistakes to Avoid

• Treating CFG as a quality slider — it is a compliance slider.
• Raising CFG to fix a prompt that is genuinely ambiguous.
• Leaving CFG at 1 on a non-distilled model and blaming the model for ignoring the prompt.
• Ignoring the fact that some 2026 models have no CFG control at all, and require language-based steering instead.

Conclusion

CFG scale and guidance scale remain the most direct form of AI image control available in 2026, even as newer architectures push toward distilled, guidance-free generation. The skill is not memorizing a magic number — it’s learning to read an image and know whether it needs more instruction or less. Start at the model default, move in small increments, watch for contrast and artifact signals, and pair your setting with an appropriate sampler and step count.

Do that consistently, and you’ll spend far less time rerolling and far more time getting exactly the image you described.