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

If your AI images keep coming out either washed-out and generic or crunchy and over-saturated, the culprit is often a single number: the CFG scale. Classifier-free guidance remains the most powerful lever you have over how literally a model follows your prompt — and in 2026, it’s also the setting most people get wrong, because newer architectures like Flux and SD3.5 handle guidance very differently from the Stable Diffusion models of 2022. This guide explains what CFG scale and guidance scale actually do, how they differ, and exactly which values to use for AI image control across today’s most popular models.

What Is CFG Scale?

CFG stands for classifier-free guidance. It’s a technique that runs your prompt through the model twice at each denoising step:

• Once with your text prompt (the conditional prediction)
• Once without it, or with your negative prompt (the unconditional prediction)

The model then amplifies the difference between the two. A higher CFG scale means a stronger push away from the “average” image and toward your prompt. A lower value means the model leans on its own learned aesthetic and statistical priors.

A Quick Definition

CFG scale (guidance scale) is a multiplier that controls how strongly the model steers toward your prompt versus its own internal tendencies. In most interfaces it appears as a slider or number input roughly in the range of 1 to 20, with a default around 7 for older Stable Diffusion checkpoints.

How CFG Scale Works in Practice

At CFG 1, guidance is effectively switched off — the model ignores the contrast between conditional and unconditional predictions. Below that, negative prompts stop working entirely, because there’s no unconditional branch left to subtract.

CFG Scale vs. Guidance Scale: What’s the Difference?

The two terms are used almost interchangeably, but there’s a nuance worth knowing in 2026:

• CFG scale traditionally refers to the classic two-pass classifier-free guidance multiplier.
• Guidance scale is the broader, vendor-neutral term. Flux.1, for example, exposes a “guidance” value (default 3.5) that is not the same mechanism as classic CFG.
• Some modern models bake guidance into the architecture during training, so the exposed slider is really a distillation parameter rather than a live multiplication.

Practically: if your UI says “CFG,” you’re on classic guidance. If it says “guidance,” check the model’s documentation before assuming the numbers translate.

How CFG Scale Changes Your Image

Low CFG (1–4)

• Softer, more painterly, and more “dreamlike”
• Greater variety between seeds
• Weaker prompt adherence; details may drift
• Negative prompts barely register

Balanced CFG (5–7)

This is the sweet spot for most SDXL and SD 1.5 workflows. You get solid prompt adherence without the telltale over-contrast of high guidance.

High CFG (8–12)

• Very literal prompt following
• Increased saturation, sharpness, and contrast
• Higher risk of artifacts: extra fingers, halos, burnt highlights
• Reduced creative variety across seeds

Extreme CFG (13+)

Beyond around 13, most models start producing what’s commonly called “deep-fried” output: crushed shadows, neon saturation, and visible structural errors. It’s occasionally useful for stylized graphic work, but rarely for photorealism.

Recommended Guidance Settings by Model (2026)

• SD 1.5 checkpoints: CFG 7–8
• SDXL base and most fine-tunes: CFG 5–7
• SD3.5 Medium / Large: CFG 4.5–5.5
• Flux.1 [dev]: guidance 3.5 (CFG 1 with embedded guidance enabled)
• Flux.1 [schnell]: guidance 0 or CFG 1 — it’s a distilled, guidance-free model
• Turbo / Lightning / LCM / Hyper checkpoints: CFG 1–2
• Midjourney: no CFG slider; use --stylize, --weird, and --chaos instead
• API image models (GPT-image, Gemini image): guidance is not user-exposed

These numbers are starting points. Every fine-tune has its own personality, and LoRA stacks can shift the ideal value by a point or two in either direction.

Why Many 2026 Models Ignore CFG

The biggest shift in the last two years is the rise of distilled and guidance-embedded architectures. Because classic CFG requires two forward passes per step, it roughly doubles inference cost. Distilled models like Flux schnell, SDXL Turbo, and Lightning variants were trained to hit high quality in one to eight steps with no guidance branch at all.

The result: if you crank CFG to 9 on a Turbo model, you don’t get sharper images — you get noise. The correct move is almost always CFG 1, and let the model’s training do the steering.

Practical Tips for Better AI Image Control

• Change one variable at a time. Lock your seed before you touch the CFG slider so you’re comparing like for like.
• Fix the prompt before the CFG. If the composition is wrong, more guidance won’t help — it’ll just make the wrong composition more contrasty.
• Pair CFG with step count. Low CFG plus 30–40 steps often beats high CFG plus 15 steps for realism.
• Use negative prompts intentionally. They only work above CFG 1, and they get stronger as you push guidance higher.
• Lower CFG to reduce artifacts. If you’re seeing extra limbs or melted faces, drop guidance by 1–2 points before changing anything else.
• Batch-test three values. Generate the same seed at CFG 4, 6, and 8, then compare side by side. It takes seconds and teaches you the model’s range fast.
• Read the model card. Recommended guidance values are usually listed, and they’re almost always more reliable than default sliders in third-party UIs.

Common CFG Mistakes to Avoid

• Assuming high CFG equals higher quality. It equals higher contrast, not better fidelity.
• Using CFG 7 on every model. That habit dates back to SD 1.5 and actively hurts Flux and SD3.5 results.
• Ignoring the sampler. DPM++ 2M Karras and Euler a respond differently to the same CFG value.
• Fighting a distilled model. If the checkpoint was trained for CFG 1, let it run at CFG 1.

A Simple Workflow to Dial It In

1. Read the model card and set guidance to the recommended value.
2. Generate one image with a fixed seed and a detailed prompt.
3. If prompt adherence is weak, raise guidance by 0.5. If artifacts appear, lower it by 0.5.
4. Once the guidance is dialed in, start tuning steps, then the sampler, then LoRA weights.
5. Save your working values in a preset so you’re not rediscovering them every session.

Conclusion

CFG scale and guidance scale are still the fastest way to gain real AI image control in 2026 — but the playbook has changed. Classic CFG values around 5 to 8 remain the standard for SDXL and SD 1.5, while newer models like SD3.5 and Flux expect noticeably lower guidance, and distilled checkpoints want none at all. The reliable approach is the same as it’s always been: start with the model’s recommended value, fix your prompt first, change one variable at a time, and use the slider as a fine adjustment rather than a quality dial. Master that habit, and your outputs will go from inconsistent to intentional.