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

The single slider that changes more AI artwork than any prompt tweak is the CFG scale — and most creators still use it blindly. If your images look burnt, oversaturated, or weirdly rigid, your CFG is probably too high. If they drift off-prompt and ignore half of what you asked for, it’s probably too low. The CFG scale and its modern cousin, the guidance scale, are the primary dials for AI image control: they decide how strictly a model listens to your prompt versus how freely it improvises. This guide explains what these values actually do in 2026’s model landscape, what numbers to use for which tools, and how to combine them with your prompts for consistently better results.

What Is CFG Scale in AI Art Generation?

CFG stands for Classifier-Free Guidance. It’s a technique introduced in 2022 that lets diffusion models evaluate your prompt twice per step: once with the text conditioning applied, and once without it (the “unconditional” pass). The model then pushes the final result further in the direction of the difference between those two predictions.

That difference is scaled by a number — the CFG scale. A higher value amplifies the conditioned direction, so the output adheres more tightly to your prompt. A lower value keeps the result closer to the model’s learned average, which tends to be more natural, more varied, and softer in contrast.

How CFG Works Under the Hood

The math is simpler than it sounds. At each denoising step, the model blends two predictions: the unconditional one and the prompt-guided one, using a formula like final = uncond + scale × (cond − uncond). At a scale of 1, you get almost pure model instinct. At 7, you get a strong prompt signal. At 20+, the formula extrapolates far beyond anything the model was trained to produce, which is why extreme values cause artifacts.

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

The terms are often used interchangeably, but they aren’t quite the same thing anymore.

Classifier-Free Guidance (CFG)

The original technique: two passes, one prompt, one blank. It’s accurate and predictable, but it roughly doubles compute per step. Stable Diffusion 1.5, SDXL, and many community checkpoints still work this way.

Distilled & Reference Guidance

Modern architectures — Flux, SD3.5, and most 2026 diffusion transformers — bake guidance into training. The model learns to predict the guided output directly, so the “guidance scale” is a learned control rather than a runtime multiplier. Flux in particular has a distilled guidance value embedded in the model, and developers can override it with a runtime guidance parameter. The practical effect is the same, but the sweet spot is far lower: what looked right at CFG 7 in SDXL often looks fried at guidance 7 in Flux.

What the Numbers Actually Mean

Think of the scale as a dial between creativity and obedience:

• 1.0–3.0: Maximum interpretation. Dreamy, painterly, unpredictable. Great for texture and mood; risky for specific compositions.
• 4.0–7.0: The balanced zone for most modern models. Prompt fidelity is strong, images stay natural.
• 7.0–12.0: The classic SD 1.5/SDXL range. Tighter adherence, slightly higher contrast and saturation.
• 12.0–20.0: Overdriven. Expect halo artifacts, clipped highlights, and repetitive anatomy.
• 20+: Almost always destructive. Occasionally useful for extreme stylization, rarely for anything else.

How to Tune CFG for Specific Models

Stable Diffusion 1.5 and SDXL

Start at 7. Move in increments of 0.5, not 2. SDXL tends to tolerate 5–8 well; going above 9 usually produces waxy skin and blown-out whites. Older anime checkpoints sometimes genuinely need 9–12 to lock in line art.

Flux, SD3.5, and DiT Models

Start at 3.5. These models were trained with low guidance, so 3–4 is the sweet spot and anything above 5 starts to look crunchy. If your Flux outputs feel flat, fix the prompt before touching the guidance.

Midjourney, Firefly, and Closed Platforms

Most hosted tools hide CFG behind names like stylize, weirdness, or prompt strength. Midjourney’s --stylize operates inversely in feel: higher values mean more model interpretation, not less. Read each platform’s documentation — the labels rarely map one-to-one.

Practical Tips for Better AI Image Control

• Change one variable at a time. Fix the seed, then adjust CFG. If you change both, you learn nothing.
• Lower CFG when adding detail words. Long prompts already increase adherence; high CFG on top of that causes color banding.
• Raise CFG for text and typography. Precise letterforms usually need more guidance, sometimes 8–10 even on modern models.
• Lower CFG for photographic realism. Skin, foliage, and fabric look more believable at the low end of the range.
• Pair low CFG with a strong negative prompt. This recovers prompt control without the artifacts high guidance creates.
• Use dynamic CFG when available. Schedulers like Rescale or dynamic thresholding let you start high for composition and ease off for refinement.
• Batch three values side by side. Generate at 4, 6, and 8 with identical seeds. It takes seconds and tells you more than any tutorial.

Common Mistakes to Avoid

The biggest error is treating CFG as a “quality” slider. It isn’t. Higher values don’t make images better — they make them more literal, and past a point, more broken. Newer users also carry settings across models: SDXL’s perfect 7 will destroy a Flux render. Finally, don’t chase artifacts by cranking guidance when the real problem is a vague prompt. “Beautiful woman, masterpiece, 8k” gives the model almost nothing to guide toward, so it over-interprets whatever it has.

Advanced: Dynamic CFG and Multi-Stage Workflows

In tools like ComfyUI and diffusers pipelines, you can vary guidance across the denoising schedule. A common 2026 workflow runs higher guidance for the first 30–40% of steps to establish composition, then drops to a low value to keep textures organic. Others generate a base at low CFG for realism, then run a second pass with img2img at moderate CFG to sharpen prompt adherence. Both approaches beat any single static number.

Conclusion

The CFG scale and guidance scale are the most direct levers you have over how literally an AI interprets your prompt. The rules for 2026 are short: start around 7 for SDXL and SD1.5, around 3.5 for Flux and other DiT models, and adjust in halves rather than jumps. High values buy prompt fidelity at the cost of naturalism; low values buy realism at the cost of control. Master that trade-off, combine it with well-written prompts and a sensible negative prompt, and you’ll spend far less time rerolling — and far more time making images that look intentional.