CFG Scale & Guidance in AI Image Generation (2026)

Ask ten AI artists what setting they tweak most often, and most will say the CFG scale. Classifier-free guidance — usually shortened to CFG or simply “guidance scale” — is the dial that decides how obediently an image model follows your prompt. Set it too low and your carefully written prompt drifts into vague, washed-out noise. Set it too high and you get oversaturated, crunchy, over-baked renders that look like a JPEG pushed past its limits. In 2026, with distilled models, embedded guidance, and dynamic schedulers now standard, understanding what CFG actually does is the difference between fighting your tool and directing it.

This guide breaks down the CFG scale from first principles, explains how modern models interpret guidance differently, and gives you practical numbers and workflows you can apply today.

What Is CFG Scale?

CFG scale is a number — typically between 1 and 20 — that controls how strongly the model’s output is pushed toward your text prompt and away from an unguided, “generic” result. It is one of the three core levers of AI image control, alongside the seed and the sampling step count.

The Two-Pass Trick Behind Guidance

In a classic classifier-free guidance setup, the model runs two predictions at every denoising step:

• Conditional prediction: what the image should look like given your prompt.
• Unconditional prediction: what the image should look like with no prompt at all.

The sampler then extrapolates away from the unconditional result and toward the conditional one, multiplied by the CFG value. A CFG of 1.0 means the unconditional pass carries no weight — the prompt barely steers the image. A CFG of 12 means the prompt’s direction is amplified heavily, which is why high values produce that signature contrast-and-saturation blowout.

Why Guidance Matters More Than Almost Any Other Setting

Steps affect refinement. Samplers affect texture and speed. But CFG changes the entire relationship between your words and your pixels. It influences:

• Prompt adherence: whether that specific jacket color, camera angle, or background element actually appears.
• Color and contrast: higher values darken shadows and intensify saturation.
• Compositional freedom: low values let the model improvise; high values lock the composition in place.
• Detail density: moderate increases sharpen structure, but extreme values create artifacts and repeating patterns.

Reading the Scale: Low vs. High CFG

Here is the practical behavior you will observe across most diffusion architectures:

• CFG 1–2: Minimal prompt influence. Useful for distilled and turbo models where guidance is baked in, and for very loose, painterly exploration. On a standard model, expect muddy, undefined output.
• CFG 3–6: The creative sweet spot on modern large models. Natural color, soft lighting, believable textures, and enough freedom for the model to add pleasing detail.
• CFG 7–10: The classic range for older Stable Diffusion checkpoints. Strong prompt fidelity with a manageable risk of oversaturation.
• CFG 11–16: Hard prompt enforcement. Expect deep shadows, electric colors, halos around edges, and occasional anatomy breakdowns.
• CFG 17+: Almost always destructive on non-distilled models. Images become flat, contrasty, and repetitive.

Guidance in 2026: Every Model Treats It Differently

The single biggest misconception in 2026 is assuming CFG behaves the same way everywhere. It does not. Guidance-distilled models bake the prompt-following behavior into training, so their sliders operate on a completely different scale.

Model-Specific Starting Points

• Stable Diffusion 1.5 and its derivatives: start at 7.0; range 5–9.
• SDXL and SDXL-based finetunes: start at 5.0; many artists hover between 4 and 7.
• Flux-style guidance-distilled models: start around 3.0–3.5; values above 5 tend to burn detail. Many users simply leave it at 1 and control strength through prompt wording.
• SD 3.5 and embedded-guidance architectures: treat the slider as a gentle modifier rather than a primary control.
• Turbo, Lightning, LCM, and other few-step models: keep CFG between 1 and 2. Higher values wreck these samplers.

The Negative Prompt Connection

Negative prompts only function because of the guidance mechanism — they widen the gap between what you want and what you don’t. This means high CFG amplifies negatives too. If your image suddenly develops strange artifacts after you add a long negative prompt, drop the CFG rather than trimming the negative.

Practical Tips for Dialing In Your CFG Scale

1. Change One Variable at a Time

Lock your seed, then run a comparison grid at CFG 3, 5, 7, 9, and 11. You will instantly see where the model’s “personality” shifts. This single exercise teaches more than any tutorial.

2. Pair CFG With Step Count

Guidance and steps interact. A very low step count with high CFG produces harsh, half-formed images. If you are rendering at 15–20 steps, stay at the lower end of the CFG range. If you are rendering at 40+ steps, you have more headroom to raise it.

3. Rescue Blown-Out Highlights With CFG Rescale

CFG rescale is a post-guidance correction that renormalizes the conditional output, pulling back the over-contrast that high values introduce. If you need CFG 9 for prompt fidelity but hate the color burn, enable rescale around 0.5–0.7 instead of lowering the guidance value.

4. Use Dynamic or Scheduled CFG

Many 2026 interfaces let you ramp guidance across the denoising timeline. A common recipe is high guidance early to lock in composition, then a gentle taper toward 3–4 in the final steps for natural color and soft lighting.

5. Watch for the Baked-In Guidance Trap

If your images look harsh no matter what you do, check whether your model already has embedded or distilled guidance. On these architectures, a “normal” CFG of 7 is effectively a 2x overdrive.

6. Keep Prompt Weighting Modest

Token weighting and CFG compound. If you are already at CFG 8, adding strong weights multiplies the effect and produces artifacts fast. Prefer one control or the other.

Common CFG Mistakes to Avoid

• Assuming higher means better. Past a model-specific threshold, adherence stops improving and quality collapses.
• Copying settings between models. A CFG of 7 that sings on SD 1.5 may ruin a distilled model.
• Ignoring the sampler. Some samplers are far more sensitive to guidance than others; re-test when you switch.
• Never lowering it. Many “bad renders” are simply over-guided — try CFG 3 before rewriting your prompt.

Conclusion

The CFG scale remains the most powerful and most misunderstood control in AI image generation. In 2026, the rule of thumb has shifted: modern models want lower guidance than the defaults of two years ago, distilled models want almost none, and pairing CFG with rescale or scheduling delivers better results than brute-forcing a single number. Start at your model’s recommended default, run a seed-locked comparison grid, and let the images tell you where the sweet spot lives. Master that dial, and you move from hoping the model listens to knowing exactly how hard to push.