How to Use CFG Scale & Guidance in AI Image Generation 2026

If you have ever typed a perfect prompt only to get an over-saturated, melted mess — or a beautiful image that ignored half your instructions — the culprit is usually one slider: the CFG scale. Classifier-free guidance is the single most powerful control knob in diffusion and flow-matching image models, and in 2026 it is also the most misunderstood. Modern models such as Flux, SD 3.5, and the latest video generators handle guidance very differently than the Stable Diffusion 1.5 workflows most tutorials still describe. This guide explains what the CFG scale and guidance scale actually do, how they differ across current model families, and the practical settings that give you precise AI image control without frying your output.

What Is CFG Scale (Classifier-Free Guidance)?

CFG scale — sometimes called the guidance scale, CFG weight, or simply “guidance” — is a number that controls how strongly your text prompt steers the generation process. At its core, the model predicts two things at every denoising step:

• The conditioned prediction: what the image should look like given your prompt.
• The unconditioned prediction: what the image should look like with no prompt at all (the empty or negative prompt).

Classifier-free guidance amplifies the difference between those two predictions. A higher value pushes the result further away from the “generic” image and further toward your prompt. A value of 1.0 means no amplification — the prompt has minimal influence.

How CFG Scale Actually Works in Practice

Think of it as a contrast dial. At low settings, the model wanders freely and produces soft, varied, sometimes dreamlike results that only loosely match your prompt. As you raise the value, adherence improves — but so does contrast, saturation, and edge hardness. Push it too far and you get the classic “deep-fried” look: crushed shadows, neon colors, duplicated limbs, and repetitive compositions.

There is no universal “correct” CFG value. The right number depends on the model architecture, the sampler, the scheduler, and how specific your prompt is.

CFG Scale vs. Guidance Scale in 2026: What Changed

Historically, “CFG scale” and “guidance scale” were interchangeable terms. In 2026 that is no longer strictly true, because the underlying math has evolved.

• True CFG (dual-pass): The model runs both a positive and negative prompt pass per step. This is what SDXL and SD 1.5 use. Cost: roughly double the compute.
• Distilled / embedded guidance: Flow-matching models like Flux train guidance directly into the network. The “guidance” slider is a conditioning input, not a true two-pass calculation — so it is cheaper but behaves differently.
• Guidance rescale: A post-processing trick that re-normalizes the guided prediction to prevent over-saturation at high values. Now standard in most 2026 interfaces.
• CFG++ and related methods: These reformulate guidance so that higher values no longer degrade quality the way they once did, allowing stronger prompt adherence at lower effective strength.

Practical takeaway: if your UI offers both “CFG scale” and “guidance,” they may refer to genuinely different mechanisms. Always check the model card.

Recommended CFG Values by Model Family

These ranges reflect common 2026 defaults. Treat them as starting points, not laws.

• SD 1.5 / SD 2.x: 6–8. Lower for realism, higher for illustration.
• SDXL and SDXL Turbo variants: 4–7 for base SDXL; turbo/LCM models often want 1.5–2.5.
• Stable Diffusion 3.5: 3.5–6. MMDiT architecture penalizes high values more than SDXL.
• Flux (dev): 2.5–4.0. Above 4.5, artifacts and stiff compositions appear quickly.
• Flux (schnell) and other distilled models: 1.0. These models are trained without guidance; raising it breaks them.
• Distilled video models: 1.0–2.0, often paired with a separate motion strength control.

How to Tune Guidance for Specific Workflows

1. Photorealistic Portraits

Start at the lower end of the model’s range and raise slowly. Realism rewards softness; a CFG of 3–5 on Flux dev or 5–6 on SDXL typically yields natural skin and lighting. Above that, skin turns plastic and highlights blow out.

2. Illustration, Logos, and Text

These tasks need literal prompt adherence, so push higher — usually the top of the recommended range. Pair it with a strong negative prompt that lists the artifacts you want to avoid.

3. ControlNet and Structural Workflows

When ControlNet, IP-Adapter, or depth maps already constrain the composition, drop your CFG by 1–2 points. The structural conditioning is doing part of the work that guidance used to do, and doubling up causes rigidity and washed-out textures.

4. Inpainting and Regional Prompting

Higher guidance helps small regions follow instructions, but it also amplifies seams. Use moderate values plus generous feathering, and consider scheduling the CFG downward in the final steps.

5. LoRA and Fine-Tuned Models

Every LoRA shifts the effective sweet spot. A LoRA trained at CFG 7 may produce muddy results at 4. Test each new LoRA at two or three guidance values before committing.

Practical Tips for Better AI Image Control

• Change one variable at a time. Fix your seed, then adjust CFG alone. Otherwise you cannot tell what caused the change.
• Use CFG scheduling. Many 2026 samplers let you start high for structure and decay to low for texture. This gives crisp composition without burnt colors.
• Enable guidance rescale when available. A value around 0.5–0.8 lets you run higher CFG without saturation collapse.
• Write negatives for the value you’re using. A negative prompt tuned for CFG 4 will behave very differently at CFG 9.
• Keep a personal preset log. Record model, sampler, steps, and CFG for every good result. Reproducibility beats guessing.
• Lower CFG when using prompt-enhancers. LLM-expanded prompts are already detailed; high guidance then over-constrains the image.

Common Mistakes to Avoid

Chasing prompt adherence with maximum CFG. If the image ignores your prompt at CFG 5, the problem is usually prompt phrasing, model capability, or step count — not guidance. Adding more guidance often just adds artifacts.

Using SD 1.5 habits on modern models. Values of 7–12 are destructive on Flux and SD 3.5. The most common 2026 support question is still “why are my images neon orange?” — and the answer is almost always a leftover high CFG.

Ignoring the sampler interaction. A deterministic sampler and a stochastic sampler feel different at the same CFG. Compare samplers at a fixed scale before blaming guidance.

Frequently Asked Questions

What is a good CFG scale for AI image generation?

For most modern models, 3–6 is the practical sweet spot. Distilled models want 1.0. Legacy SD 1.5 workflows still use 6–8. Always start at the model’s documented default.

Does higher CFG scale mean better images?

No. It means stronger prompt adherence, which comes with a quality penalty past a model-specific threshold. The goal is the lowest value that reliably follows your prompt.

Why do some models only allow a guidance scale of 1?

Those are distilled or guidance-embedded models trained without classifier-free guidance. The slider remains for interface compatibility but has no meaningful effect above the minimum.

Conclusion

The CFG scale and guidance scale remain the most direct lever you have over AI image control in 2026 — but the era of one universal number is over. Flow-matching architectures, distilled models, guidance rescaling, and CFG++ have fragmented the rules, and the operators who get consistently great results are the ones who match their guidance settings to the specific model in front of them.

Start with the model’s recommended default. Lower it by one point and compare. Raise it by one and compare. Lock your seed, log what worked, and build a personal preset library. That simple discipline will improve your output far more than any new prompt trick — because knowing exactly how much guidance to apply is what separates a lucky render from a repeatable workflow.