How to Use CFG Scale & Guidance in AI Art (2026)

If your AI-generated images look washed out, oversaturated, or stubbornly ignore your prompt, the culprit is often a single slider: the CFG scale. Classifier-free guidance (CFG) and its modern cousin, the guidance scale, are among the most powerful levers you have for AI image control — and among the most misunderstood. Push them too low and your model wanders off-prompt. Push them too high and you get the infamous “deep-fried” look: crushed contrast, neon edges, and plastic skin. This guide breaks down exactly what CFG scale and guidance scale do in 2026’s model landscape, what values to use for which styles, and how to tune them like a pro.

What Is CFG Scale? Classifier-Free Guidance Explained

CFG scale (Classifier-Free Guidance scale) determines how strongly your text prompt steers the diffusion process. Internally, the model makes two predictions at every step: one conditioned on your prompt, and one unconditioned (essentially “no prompt”). The CFG scale is the multiplier that amplifies the difference between them.

In plain terms:

• Low CFG (1–3): The model follows its own artistic instincts more than your words. Outputs are soft, diverse, and dreamlike — but frequently off-brief.
• Mid CFG (4–9): The balanced sweet spot where prompt adherence and natural rendering coexist.
• High CFG (10–20): Aggressive prompt compliance at the cost of contrast, color accuracy, and image diversity.

CFG vs. Guidance Scale: Same Thing, Different Labels

You’ll see this parameter named differently depending on your interface — CFG scale in Automatic1111 and Forge, cfg in ComfyUI’s KSampler node, and “guidance” in FLUX and other guidance-distilled models. They’re conceptually identical, but the numbers are not interchangeable. A guidance value of 3.5 on FLUX.1 [dev] is not comparable to a CFG of 3.5 on SD 1.5. Always calibrate per model family.

Why CFG Scale Matters for AI Image Control

CFG is the primary dial for the trade-off between fidelity and naturalness. Adjusting it changes several things at once:

• Prompt adherence: Higher values force the model to honor specific details, colors, and compositions.
• Contrast and saturation: Elevated CFG over-sharpens and over-saturates, often clipping highlights.
• Diversity: Low CFG produces more variety across seeds; high CFG collapses toward similar outputs.
• Negative prompt strength: Negative prompts only work meaningfully when CFG is above 1.0, because the guidance term is what activates the contrast between positive and negative conditioning.
• Anatomy and text: Complex subjects like hands, typography, and multi-subject scenes usually need more guidance to stay coherent.

CFG Scale Cheat Sheet by Style

Photorealism and Portraits (CFG 3–6)

Realism punishes high guidance. Skin becomes waxy, shadows crush, and lighting turns harsh. Stay low and let the model’s training data do the heavy lifting. Pair with a realistic checkpoint and 25–35 steps.

Illustration, Anime, and Graphic Art (CFG 6–9)

Stylized work tolerates — and often benefits from — stronger guidance. It locks in line weight, color palettes, and distinctive character features. Anime checkpoints typically look best around 7–8.

Text, Logos, and Precise Layouts (CFG 9–13)

When the prompt must be obeyed literally, raise CFG. Modern transformers-based models such as SD 3.5 and FLUX handle text rendering well, and higher guidance helps characters stay legible.

Distilled and Turbo Models (Guidance 0–2)

This is the biggest 2026 caveat. Distilled models — FLUX schnell, LCM, Lightning, Hyper, and Turbo variants — were trained to generate in 1–8 steps with guidance disabled or near zero. Setting CFG to 7 on a 4-step turbo model produces noise. Use guidance 0–1.5 and let the scheduler do the work.

Practical Tips for Tuning CFG and Guidance

• Change one variable at a time. Fix your seed, leave the prompt untouched, and sweep CFG in increments of 1.5 so you can see the real effect.
• Never compensate with steps. Adding sampler steps won’t fix a bad CFG value; it just costs time and compute.
• Use CFG rescale or dynamic thresholding. These features tame the oversaturation that strikes above CFG 10, letting you keep strong prompt adherence with natural colors.
• Match CFG to your sampler. Some schedulers (DPM++ 2M Karras, Euler a) behave very differently at identical CFG values. Build a personal preset library.
• Lower CFG for img2img and inpainting. Because the source image already supplies structure, high guidance tends to destroy it. Try 3–5 for refinements.
• Raise CFG when the prompt is crowded. Multi-subject scenes and complex spatial instructions need more steering.
• Check the model card. Most 2026 checkpoints list a recommended guidance range. Trust it before you start experimenting.

Common Mistakes to Avoid

• Assuming all models use the same scale. FLUX guidance, SDXL CFG, and turbo guidance are different numeric worlds.
• Setting CFG to 0 by accident. That disables conditioning entirely — the model ignores your prompt and generates from noise alone.
• Relying on the negative prompt at low CFG. Below roughly 3, negatives barely register.
• Chasing “maximum detail” with maximum CFG. Detail comes from steps, resolution, and model choice — not from cranking guidance.

Conclusion

CFG scale and guidance scale remain the most direct form of AI image control you have — but they reward restraint. The winning formula in 2026 is simple: start at your model’s recommended value, adjust in small increments, keep guidance low for realism and distilled models, and let tools like CFG rescale handle the extremes. Treat the slider as a precision instrument rather than a power dial, and your generations will look sharper, more natural, and far closer to the image you actually had in mind.