How to Use CFG Scale and Guidance in AI Art (2026)
The CFG scale is the most misunderstood slider in AI image generation. Most creators set it once, never touch it again, and then wonder why their prompts either get ignored or turn into a saturated, over-baked mess. Yet among the handful of parameters that genuinely shape your output — seed, sampler, steps, resolution — the CFG scale and its modern cousin, the guidance scale, consistently deliver the biggest visible change per unit of effort. In 2026, with guidance-distilled models like FLUX and SD3.5 becoming the default for many workflows, understanding how guidance actually works is no longer optional. This guide breaks down what the CFG scale does, how it differs across model families, and how to dial it in for precise AI image control.
What Is the CFG Scale?
CFG stands for Classifier-Free Guidance. It’s a technique that runs your prompt through the model in two directions at once: a “positive” pass that leans toward what you asked for, and a “negative” or unconditioned pass that represents everything your prompt is not. The CFG scale is the multiplier that determines how aggressively the model pushes away from the unconditioned result and toward your positive prompt.
Think of it as a confidence dial:
- Low CFG — the model trusts its own training distribution and improvises freely.
- High CFG — the model obeys your prompt literally, but with diminishing returns and rising artifacts.
Crucially, the CFG scale does not add detail. It amplifies the contrast between prompt and non-prompt. That distinction explains almost every weird thing you’ve seen at CFG 15.
CFG vs. Guidance Scale: Same Idea, Different Names
You’ll see this parameter labelled differently depending on the tool. Functionally, these all control the same concept:
- CFG scale / CFG Scale — Automatic1111, Forge, ComfyUI, most Stable Diffusion frontends.
- Guidance scale — Hugging Face Diffusers, SD3, SD3.5, FLUX.1.
- Guidance / embedded guidance — distillation-trained models where guidance is baked into the network.
- Stylize, prompt strength, or adherence — proprietary tools that expose a simplified, rebranded version.
The numbers don’t translate cleanly across model families, which trips up a lot of people moving from Stable Diffusion 1.5 to SDXL to FLUX. A CFG of 7 is “normal” on SD 1.5, slightly high on SDXL, and outright destructive on FLUX.1 dev, where 3.5 is the standard.
How the CFG Scale Changes Your Images
Low CFG (1–4)
At low values, the model behaves loosely. Outputs feel painterly, dreamlike, and unpredictable. Prompt fidelity drops — extra elements creep in, colours drift, and composition is largely left to chance. Negative prompts have almost no effect here. This range is useful for:
- Abstract, textured, or atmospheric work
- Blending multiple concepts without hard edges
- Guidance-distilled models (FLUX, some SD3 variants) where 1–4 is the intended operating range
Balanced CFG (5–8)
This is the sweet spot for most traditional diffusion models. You get reliable prompt adherence, natural colour, clean anatomy, and negative prompts that actually do something. If you’re unsure where to start, begin here and adjust in increments of 0.5 rather than whole numbers.
High CFG (9–15+)
Past roughly 9–10, the model starts over-committing. Typical symptoms include:
- Crushed contrast and oversaturated colours
- Burned edges, halos, and “deep-fried” textures
- Repeated or fused limbs and objects
- Loss of fine detail in shadows and skin tones
There are niche uses — forcing a stubborn token, hardening line art, pushing a strong graphic style — but high CFG is a corrective tool, not a default.
Typical Guidance Ranges by Model Family (2026)
These are starting points, not laws. Samplers, schedulers, and LoRAs all shift the optimum.
| Model family | Typical guidance | Notes |
|---|---|---|
| SD 1.5 | 6–8 | Classic range; 7 is the community default |
| SDXL / SDXL Turbo | 4–7 | Lower than 1.5; Turbo variants near 1–2 |
| SD3 / SD3.5 | 3.5–6 | Responds sharply to changes |
| FLUX.1 dev | 2.5–4 | Guidance-distilled; high values break the image |
| FLUX.1 schnell | n/a | No true CFG; distilled for speed |
| Anime-tuned checkpoints | 5–7 | Often paired with specific negative prompt sets |
Practical Tips for Better AI Image Control
- Change one variable at a time. Lock your seed, then adjust only the CFG scale. Otherwise you can’t tell what caused the change.
- Move in 0.5 steps. The difference between 6.5 and 7.5 is often more meaningful than between 7 and 10.
- Fix prompt problems with prompts first. CFG cannot add something the model doesn’t understand. If a concept is missing, rewrite the prompt instead of cranking guidance.
- Use negative prompts as a lever at mid CFG. Their influence scales with CFG, so 6–8 is where they earn their keep.
- Lower CFG for photoreal skin. High guidance over-sharpens and plasticizes faces faster than almost any other artifact.
- Raise CFG for flat graphic styles. Logos, vector art, and illustration often benefit from 8–10, where colours stay punchy and shapes stay crisp.
- Pair low CFG with more steps. If you drop guidance, give the sampler a little more room to resolve detail.
- Watch the sampler. Some schedulers (DPM++ 2M Karras, Euler a) tolerate high CFG better than others; ancestral samplers amplify noise at high values.
A Quick Sweet-Spot Workflow
- Generate the same prompt at CFG 4, 6, 8, and 10 with a fixed seed.
- Identify the highest value before artifacts appear or colours shift unnaturally.
- Step back by 0.5–1.0. That’s your working value for this model and style.
- Save it as a preset — guidance optima are model-specific and worth tracking.
Common CFG Mistakes to Avoid
- Assuming one number fits all models. It doesn’t. Recalibrate whenever you switch checkpoints.
- Using CFG to fix a vague prompt. Vague prompts plus high CFG equals confident nonsense.
- Ignoring distilled architectures. On FLUX and similar models, pushing guidance to 7+ actively degrades output.
- Treating settings as permanent. Guidance interacts with LoRA strength, resolution, and step count — recheck it after any major workflow change.
Conclusion
The CFG scale and guidance scale are not magic quality sliders — they’re adherence controls. They decide how literally the model reads you, and how much it’s allowed to improvise. Once you internalise that, the parameter stops being mysterious: low values give you freedom, high values give you obedience, and the useful range almost always sits somewhere in the middle. In 2026, the biggest practical adjustment is simply remembering that distilled models like FLUX expect far lower guidance than the Stable Diffusion checkpoints most tutorials were written for. Lock a seed, sweep a few values, and note what your specific model likes. That single habit will improve your AI image control more than any prompt trick you’ll read this year.