CFG Scale & Guidance: AI Image Generation (2026)

If you’ve ever generated an image that ignored half your prompt — or one that looked like a deep-fried poster of itself — the CFG scale is probably to blame. Classifier-free guidance remains the single most consequential dial in diffusion-based image generation, and in 2026 it’s more misunderstood than ever. Newer models have renamed it, hidden it, distilled it, or removed it entirely, and each of those changes alters how you should approach AI image control. This guide breaks down what the CFG scale actually does, how the guidance scale in modern tools differs from the classic implementation, and the practical settings that produce reliable results across Stable Diffusion, FLUX, and today’s video generators.

What Is CFG Scale?

CFG stands for classifier-free guidance. It’s a numerical value — typically between 1 and 20 — that determines how aggressively the model pushes its output toward your text prompt and away from generic, unremarkable imagery.

The mechanism is straightforward. During generation, the model evaluates your image twice at each step: once conditioned on your prompt, and once with an empty or null prompt. It then extrapolates in the direction of the difference. A low CFG scale keeps the two predictions close together, producing softer, more varied, more “creative” results. A high CFG scale amplifies the gap, forcing the image to conform tightly to your text.

The Short Version of the Math

Without getting lost in the algebra, the formula is roughly: final prediction = uncond + scale × (cond − uncond). When the scale equals 1, you get the raw conditioned model with no guidance at all. When it climbs to 12, the difference is multiplied twelvefold — which is exactly why images start to burn out at high values.

Why It’s Called “Classifier-Free”

Early guidance methods required a separate classifier network to judge how well an image matched a prompt. The 2022 breakthrough was showing you could get the same steer without a second model, by simply training with random prompt dropout. That’s the technique essentially every modern diffusion model still uses.

CFG Scale vs. Guidance Scale: Same Idea, Different Names

Here’s where newcomers get tripped up. Different platforms label the same concept differently, and in a few cases the underlying implementation genuinely differs.

• CFG scale / CFG Scale: The classic term, used in A1111, Forge, ComfyUI, and most Stable Diffusion interfaces. Typical range 1–20, with 7 as the default.
• Guidance scale: The term used by Stability AI’s APIs, Diffusers, and many hosted services. Functionally identical to CFG.
• Guidance (FLUX): In FLUX.1 [dev], guidance is a distilled parameter rather than true CFG. FLUX is trained to follow an embedded guidance value, so the default of 3.5 behaves very differently from a CFG of 3.5 in SDXL.
• CFG 1.0 models: Turbo, Lightning, and schnell variants are distilled to work without true classifier-free guidance. Setting CFG above 1 typically destroys the image rather than improving it.

The practical takeaway: never assume a guidance number from one model transfers to another. Read the model card.

What Different CFG Values Actually Do

Assuming a standard, non-distilled diffusion model such as SDXL or SD 3.5:

• 1.0–2.0: Almost no prompt adherence. Useful only for img2img refinement or when you want the model to riff freely on a composition.
• 3.0–5.0: Loose, painterly, and often inventive. Great for artistic styles where you want the model to contribute its own interpretation.
• 6.0–8.0: The sweet spot for most work. Strong prompt adherence, natural contrast, and minimal artifact risk.
• 9.0–12.0: Tighter conformity, but colors begin to oversaturate and shadows crush. Useful for precise text rendering or strict composition matching.
• 13+: Visible burning, halos, and distorted anatomy. Rarely worth it outside of deliberate stylistic effects.

Higher CFG also reduces variation between seeds. If you’re generating a batch and every image looks nearly identical, your guidance is too high.

CFG in the 2026 Toolbox

Distilled and Turbo Models

Four-step and eight-step models now dominate fast inference. They trade away classifier-free guidance because it requires two forward passes per step. Leave CFG at 1.0 on these models and control the output through prompt engineering, LoRAs, and negative embeddings instead.

FLUX and Guidance-Distilled Architectures

With FLUX.1 [dev], guidance is a learned conditioning signal. The 3.5 default sits in a different conceptual space than SDXL’s 3.5, and pushing it to 7 does not produce the same effect. Treat published model defaults as your real starting point, not folklore from a 2023 forum thread.

Image-to-Image and Inpainting

In img2img workflows, CFG interacts with the denoising strength. High CFG plus high denoising produces drastically reinterpreted images; moderate CFG around 5–6 with denoising at 0.4 preserves structure while allowing stylistic shifts. For inpainting, drop CFG slightly — masked regions tend to blow out faster than full-frame generations.

Practical Tips for Dialing In Guidance

• Start at the model default, then adjust in increments of 0.5. Large jumps make it impossible to isolate what changed.
• Hold the seed constant while tuning. Change one variable at a time or you’re not testing anything.
• Pair CFG with a negative prompt. On classic models, a CFG of 6 with a well-crafted negative prompt may out-perform a CFG of 10 with none.
• Lower CFG for realism, raise it for typography. Photographic skin texture degrades quickly at high guidance, while legible text generally benefits from tighter adherence.
• Watch for crushed blacks. If shadows lose detail before your prompt is fully satisfied, you’ve hit the ceiling for that model.
• Check the scheduler. Some samplers amplify guidance effects more sharply than others; a value that works with DPM++ 2M may misbehave with Euler.

Common Mistakes to Avoid

The most frequent error is treating CFG as a quality slider — it isn’t. Cranking it to 15 doesn’t make an image “better,” it makes it more literal and more damaged. The second most frequent error is ignoring model-specific guidance implementations, which leads to wasted hours on turbo models set to CFG 7. And the third is assuming that a prompt which fails at CFG 7 will succeed at CFG 12; usually it’s the prompt that needs rewriting, not the dial.

Conclusion

The CFG scale and its guidance-scale cousins remain the most direct lever you have over AI image control — but they reward precision, not force. Understand which architecture you’re running, start from the model’s documented default, and tune in half-point increments with a fixed seed so you can actually see what’s changing. On classic diffusion models, the reliable zone sits between 5 and 8. On distilled and guidance-embedded models in 2026, the correct answer is often “leave it alone and fix the prompt.” Master that distinction and you’ll spend less time fighting artifacts and more time generating images that match what you actually pictured.