What Is CFG Scale? Guidance Scale Explained, With Values for Every Model (2026)

CFG scale (classifier-free guidance scale, also called guidance scale) is the number that controls how strongly a diffusion model follows your prompt. Low values (1–4) give looser, more varied images that may ignore parts of the prompt; high values (12+) follow the words closely but often look burnt, oversaturated and samey. For classic Stable Diffusion and SDXL models, 5–8 is the usual sweet spot, while distilled and turbo models use 1–4 or ignore it entirely.

This guide explains what CFG actually does inside the sampler, what each value range looks like, which numbers to start with for each model family (SD 1.5, SDXL, Pony, Flux, turbo models) and how CFG interacts with steps and samplers. Updated as of 23 October 2026.

What does CFG stand for?

CFG stands for classifier-free guidance, not "configuration", which is a common mix-up in tutorials. The technique comes from the paper Classifier-Free Diffusion Guidance by Jonathan Ho and Tim Salimans (Google Research, first posted July 2022). Before it, diffusion models needed a separate image classifier to steer them toward a label. Ho and Salimans showed you could train one model to work both with a prompt (conditional) and without one (unconditional), then blend the two predictions at sampling time to get the same quality-versus-diversity trade-off without any classifier.

Every tool names the setting a little differently:

The defaults for Diffusers, FLUX and SDXL Turbo come from their official docs and model cards (linked in the sources). The A1111 and ComfyUI defaults are the values those interfaces ship with.

How does CFG scale work?

At every denoising step, a model using CFG runs two predictions of the noise in the image:

1. Unconditional: what the model would draw with an empty prompt (or with your negative prompt).
2. Conditional: what it would draw with your prompt.

It then takes the difference between the two (the direction "towards your prompt") and multiplies it by the CFG scale:

final = unconditional + CFG × (conditional − unconditional)

That is the exact line in the Hugging Face Diffusers tutorial for writing your own Stable Diffusion pipeline. A few things follow from that formula:

• CFG = 1 gives you just the conditional prediction, with no extra push. Diffusers treats guidance_scale ≤ 1 as "guidance off" and skips the unconditional pass.
• CFG > 1 exaggerates whatever separates your prompt from "no prompt". That is why higher values follow the prompt more literally, and also why colours and contrast get pushed too far.
• Two predictions per step means classic CFG roughly doubles the compute per step. This is one reason newer distilled models bake guidance in: they get CFG-like results from a single prediction.
• Negative prompts only work through CFG. The negative prompt replaces the empty "unconditional" side, so the model is pushed away from it. With guidance off (CFG 1 or a distilled model) a negative prompt has nothing to act on. We cover this in more detail in our negative prompts guide.

What happens at low vs high CFG values?

The exact thresholds shift between models, but the pattern is the same on any CFG-trained model:

Ho and Salimans describe this trade-off in terms of sample quality versus diversity: more guidance gives images that match the condition more strongly but look more alike. In practice, "quality" stops improving well before the top of the slider, which is why most guides land on 5–9.

Signs your CFG is too high

• Colours look neon or posterised, skin looks plastic or orange
• Hard black outlines and blown-out highlights
• Repeated, "fried" textures in hair, grass or fabric
• Every seed produces almost the same composition

Signs your CFG is too low

• Objects or attributes from the prompt are missing
• The image looks muddy, grey or low-contrast
• Results drift toward generic subjects that weren't asked for
• Text and fine details come out vaguer than usual

What CFG scale should I use for each model?

There is no single best value. Use the model card's recommendation first, then adjust by one or two points at a time on the same seed. These are reasonable starting points as of October 2026:

The Pony range comes from several community guides rather than an official spec, so treat it as a starting point. We cover Pony's other quirks on the Pony XL model page.

Why do Flux models use such low guidance?

FLUX.1 [dev] was trained with guidance distillation, which its model card lists as a key feature. A distilled model learns to produce the guided result from a single pass. The "guidance" number you pass is a conditioning input that tells it how strongly to behave as if it were guided, so the sensible range is around 2–5, not 7. Pushing a distilled Flux model to SDXL-style values of 7–10 usually gives the same over-cooked look as high CFG elsewhere.

The FLUX.2 [klein] Base models are different. Black Forest Labs describes them as trained "without step or guidance distillation", which keeps them flexible for fine-tuning and gives more output diversity. They run real two-pass guidance, and the card's example uses 4.0. You can compare the Flux 2 variants on our Flux 2 model page.

Do turbo and lightning models use CFG?

Mostly no. Step-distilled models such as SDXL Turbo were trained to work in 1–4 steps without guidance. Stability's card says SDXL Turbo "does not make use of guidance_scale or negative_prompt". Some LCM and Lightning setups tolerate CFG 1–2, but anything higher quickly burns the image.

How do CFG scale, steps and samplers interact?

CFG never acts alone. Three settings work together:

• Steps: higher CFG pushes harder on each step, so very low step counts with high CFG often look rough. If you raise CFG, give the sampler enough steps (25–40 for SDXL-class models) to converge. Going far past that rarely helps.
• Sampler: some samplers handle high CFG better than others. Ancestral samplers (Euler a, DPM++ SDE) add fresh noise each step, which can hide some burn but also changes the image as steps change. Our samplers and schedulers explainer covers Euler, DPM++ 2M, Karras and Beta in detail.
• CFG rescale / guidance rescale: a fix proposed by Lin et al. in Common Diffusion Noise Schedules and Sample Steps are Flawed (2023). It rescales the guided prediction so it stops overshooting brightness and contrast, which the paper calls over-exposure. Diffusers exposes it as guidance_rescale (0 = off by default), and values between 0 and 1 blend in the fix. It is especially useful for v-prediction models that look overexposed at normal CFG values.

In img2img and inpainting, CFG interacts with denoising strength too. At low strength the model can only change a little, so high CFG mostly adds saturation rather than following the prompt better.

How to find the right CFG scale in 5 steps

1. Fix the seed. Use one seed for every test so the only thing changing is CFG.
2. Start at the model card value (or the table above): 7 for SD 1.5, 5–6 for SDXL, around 3.5 for FLUX.1 dev, 0 for turbo models.
3. Run a sweep. Generate the same prompt at three or four values, for example 4, 6, 8 and 10. Most UIs have an "X/Y plot" or batch option for this.
4. Read the failures. Missing prompt elements mean it's too low. Burnt colours and same-looking seeds mean it's too high. Pick the lowest value that still includes everything you asked for.
5. Fix the prompt before cranking CFG. If an element only appears at CFG 12+, reword the prompt, move the key subject earlier, or add weight to it. That almost always beats forcing it with guidance.

Do I need to set CFG scale on Kenerate?

No. Kenerate does not expose a CFG or guidance slider in its image studio. Each model runs with fixed default settings, so you control results with the prompt, the size preset and, on models that support it, a seed for repeatable results. That suits most people who want good images without tuning samplers. If you need full control over CFG, steps and samplers for a local checkpoint, a local tool like ComfyUI or Forge is the better choice.

You can try the model families covered in this guide in the AI image generator. Credits per image as of 23 October 2026, from Kenerate's model config:

For stylised characters, the AI anime generator page shows which models handle anime best, and the Chroma model page covers the open Flux-based Chroma model.

Common CFG scale mistakes

• Using SD 1.5 values on every model. CFG 7 is fine on SD 1.5, too high for many SDXL fine-tunes and much too high for distilled Flux or turbo models.
• Raising CFG to fix anatomy. Hands and faces depend far more on the model, resolution and steps than on guidance. High CFG usually makes them crisper, not more correct.
• Expecting negative prompts to work at CFG 1. With guidance off, the negative prompt is skipped entirely.
• Changing several settings at once. Change CFG, steps and sampler one at a time on a fixed seed, or you won't know which one helped.

Sources

• Ho, J. and Salimans, T. — Classifier-Free Diffusion Guidance (arXiv, 2022)
• Hugging Face Diffusers — Understanding pipelines, models and schedulers
• Black Forest Labs — FLUX.1 [dev] model card and FLUX.2 [klein] Base 9B model card
• Stability AI — SDXL Turbo model card
• Lin, S. et al. — Common Diffusion Noise Schedules and Sample Steps are Flawed (arXiv, 2023)