The CFG scale is one of the few settings in AI image generation that can make or break a render. Set it too low and your prompt gets ignored; push it too high and you get burned colors, crunchy textures, and over-saturated artifacts. In 2026, with models like Flux, SDXL derivatives, and the latest proprietary generators all handling guidance differently, knowing how to tune this single number is the difference between a usable image and twenty wasted generations.
This guide breaks down what CFG scale and guidance scale actually do, how they differ across model families, and the practical ranges and workarounds you can apply today for tighter AI image control.
What Is CFG Scale?
CFG stands for classifier-free guidance. It’s a numeric value that tells the model how strongly it should follow your text prompt versus how freely it can improvise. Every diffusion and flow-matching model compares two predictions during generation: one with your prompt attached (the conditional prediction) and one without it (the unconditional prediction). The CFG scale determines how much weight the difference between those two gets.
In practical terms:
- Low CFG scale — the model leans on its own learned image distribution. Results feel creative, soft, and often off-prompt.
- High CFG scale — the model chases your prompt aggressively. Results match the text more literally but can look harsh, plastic, or distorted.
That trade-off is the entire game. There is no universally “correct” value — only the right value for a given model, prompt, and desired aesthetic.
How CFG Scale Works Under the Hood
The Core Formula
Most implementations follow a simplified version of this:
final_output = uncond + CFG × (cond − uncond)
When CFG equals 1, the guidance term becomes the conditional prediction alone — effectively no guidance. As the value climbs, the model amplifies the direction your prompt pushes it in, and that amplification is what creates both sharper prompt adherence and the artifacts that come with over-steering.
What Happens at Each End of the Range
- CFG 1–2: Dreamlike, loose, frequently ignores specific prompt details. Useful for abstract or painterly work.
- CFG 4–7: The classic sweet spot for most base diffusion models. Balanced adherence with natural texture.
- CFG 8–12: Strong prompt fidelity. Watch for waxy skin, crushed shadows, and color banding.
- CFG 15+: Almost always degraded. High contrast, neon saturation, and structural breakdown.
CFG Scale vs. Guidance Scale: What’s the Difference?
Functionally, very little — and this trips up a lot of people. “Guidance scale” is simply the newer, more generalised name for the same parameter, adopted because transformer-based and flow-matching architectures don’t always implement guidance through the traditional classifier-free path.
- Stable Diffusion 1.5 / SDXL: Typically labelled “CFG Scale.” Comfortable range: 5–8.
- Flux family: Labelled “Guidance.” Distilled variants respond well around 2.5–4.
- Midjourney: No direct CFG slider. The –stylize and –chaos parameters act as rough analogues.
- Hosted API models: Often expose “guidance” or “prompt strength” with narrow recommended windows in their docs.
The lesson: never carry a CFG value from one model family to another. Always start from the model’s documented default and adjust from there.
How to Choose the Right CFG Value
Start at the Default
Every model ships with a tuned default because its training pipeline was built around that number. Generate one image at the default before changing anything. This gives you a baseline for what “correct” looks like for that model.
Adjust in Small Increments
Move in steps of 0.5 to 1.0 — not jumps of five. Diffusion behaviour is non-linear, and a change from 6 to 7 can matter far more than 3 to 4. Generate the same seed at each step so you’re comparing guidance, not randomness.
Match CFG to Your Prompt’s Specificity
- Vague prompts (“a moody forest”) — lower CFG, let the model interpret.
- Precise prompts (“a red 1967 Ford Mustang on wet asphalt, neon reflections, 35mm”) — higher CFG to lock in details.
- Multi-subject prompts — moderate CFG plus regional prompting or inpainting beats cranking the slider.
Practical Tips for Stronger AI Image Control
- Use negative prompts before raising CFG. Removing unwanted elements via the negative prompt is gentler than forcing adherence upward.
- Enable CFG rescale or dynamic thresholding. These techniques counteract the over-saturation that high guidance produces, letting you run 8–10 without blowing out highlights.
- Lower CFG for stylised LoRAs. Adapter layers already carry strong conditioning; stacking high CFG on top creates harsh, over-fitted results.
- Raise CFG when using ControlNet. Structural guidance competes with text guidance, and a small bump often helps the prompt stay present.
- Test at a fixed seed and fixed sampler. Change one variable at a time or you’ll never learn what the CFG change actually did.
- Try guidance scheduling. Several 2026 pipelines let you start with high guidance for composition and decay it in later steps for smoother detail.
Common Mistakes to Avoid
- Believing higher CFG means better quality. It means stronger adherence — a very different thing.
- Copying settings across models. A value of 7 that’s perfect for SDXL may destroy a Flux render.
- Using CFG to fix a bad prompt. If your prompt is contradictory, no guidance value will resolve it. Rewrite instead.
- Ignoring the interaction with steps. Fewer sampling steps often need slightly higher CFG; more steps can tolerate lower values.
What’s Changed in 2026
Three shifts are worth knowing. First, guidance distillation has become standard — many modern models bake guidance into training, meaning you get strong prompt adherence at very low or even fixed guidance values. Second, guidance scheduling is now exposed in mainstream interfaces rather than being a research curiosity. Third, hosted models increasingly hide the parameter entirely, replacing it with a single “prompt strength” slider that maps to an internal range.
The practical consequence: CFG scale is no longer a universal dial. It’s a model-specific one, and reading each model’s documentation matters more than memorising a magic number.
Conclusion
CFG scale and guidance scale remain the most direct levers you have over how literally an AI image model interprets your prompt. The fundamentals haven’t changed — low values favour creativity, high values favour fidelity, and every model has a narrow band where both coexist. What has changed in 2026 is that the parameter is increasingly model-specific, distilled into training, and often scheduled across the sampling process.
Start at the default, move in small increments, hold your seed steady, and combine guidance adjustments with negative prompts and rescaling techniques rather than relying on the slider alone. Do that, and you’ll spend far less time regenerating and far more time shipping images that actually match the picture in your head.