If you have ever nudged a single slider and watched an AI image shift from a soft, coherent illustration into a harsh, over-saturated mess, you have already met the CFG scale. Guidance scale — as it is often called — is one of the few settings that touches every stage of the generation process, and in 2026 it remains the most misunderstood control in the AI image toolbox. This guide explains what the CFG scale actually does, how it behaves differently across modern models, and how to tune it for reliable, repeatable results.
What Is the CFG Scale, Exactly?
CFG stands for classifier-free guidance. Despite the name, it no longer has anything to do with classifiers. It is a mathematical trick that lets a diffusion model compare two predictions at every denoising step:
- Conditional prediction — what the model thinks the image should look like, given your prompt.
- Unconditional prediction — what the model thinks the image should look like with no prompt at all.
The gap between those two predictions is then amplified by the CFG value:
output = unconditional + scale × (conditional − unconditional)
At a CFG of 1, the model ignores that difference entirely and simply follows the unconditional path — meaning your prompt has almost no influence. As the value climbs, the model is pushed harder toward your text. Push too far and the math starts to overshoot, producing the classic high-CFG artifacts: crunchy textures, blown-out contrast, and anatomy that looks like it was melted and re-frozen.
Why the Default Is Usually Around 7
During the Stable Diffusion 1.5 era, a value between 7 and 8 tested best across the widest range of prompts, and that number became the default in most interfaces. It is a reasonable starting point, but it is not universal — and in 2026 it is frequently wrong for newer architectures.
Guidance Scale vs. CFG Scale: Same Thing, Different Names
Terminology across tools is fragmented, which causes a lot of confusion:
- CFG scale — used by Automatic1111, ComfyUI, and Forge.
- Guidance scale — used by the diffusers library and most hosted APIs.
- Distilled guidance — used by models like Flux, where guidance is baked into training rather than computed by running the model twice.
The important distinction is between true CFG and distilled guidance. True CFG requires two model passes per step and behaves predictably across a wide range. Distilled models learn a guidance pathway during training, which means the useful range is much narrower — often 2.5 to 4 — and values above 5 tend to break the image rather than sharpen it.
What Different CFG Values Actually Do
Low Values (1–3)
The model has more creative latitude. You get softer lighting, more natural color grading, and greater variety between seeds — but also looser prompt adherence. Useful for artistic, painterly, or photographic looks where you want the model to interpret rather than obey.
Mid Values (4–8)
The practical sweet spot for most community checkpoints. Prompt adherence is strong, colors stay balanced, and composition follows your description without becoming rigid.
High Values (9–15)
Text rendering and rare tokens get pushed harder, so specific details like a logo, an unusual object, or a precise color may appear more reliably. The trade-off is flatter lighting, reduced realism, and a tendency toward oversaturation.
Extreme Values (16+)
Beyond this point, you are usually fighting the model rather than guiding it. Expect posterized colors, fried detail, and duplicated limbs. If you need something at this level, a better prompt or a LoRA will almost always serve you better.
Recommended CFG Ranges by Model (2026)
Treat these as starting points, not laws — every checkpoint has its own personality:
- SD 1.5 and older SDXL checkpoints: 6–9
- Modern SDXL fine-tunes: 4–7
- SD 3.5 Medium / Large: 4–7
- Flux.1 dev and distilled derivatives: 2.5–4.5
- Turbo, Lightning, and LCM models: 1–2
- Rectified-flow models with true CFG support: 5–8
Practical Tips for Better AI Image Control
1. Change one variable at a time
If you adjust CFG, sampler, and steps simultaneously, you will never know which change produced the result. Move CFG in increments of 1 and compare.
2. Lock your seed while testing
A fixed seed isolates the effect of guidance. Once you find a CFG value you like, unlock the seed and check whether it holds up across a dozen random generations.
3. Rebalance your negative prompt
Negative prompts scale with CFG. A negative prompt that is mild at 5 becomes aggressive at 12, sometimes erasing legitimate details. If you raise CFG, consider trimming your negatives.
4. Pair CFG with step count
Higher guidance amplifies errors at every step, so it helps to add steps alongside it. A rough rule: 20–30 steps in the 5–8 range, 30–40 steps if you are pushing past 10.
5. Use guidance rescale when you go high
Most 2026 interfaces include a guidance rescale or “CFG rescale” option. A factor around 0.7 lets you keep the prompt adherence of high CFG while restoring natural contrast and color. It is the single most useful companion setting to the CFG slider.
Advanced Guidance Techniques Worth Knowing
Dynamic and scheduled CFG
Instead of one fixed value, schedules vary guidance across the denoising process — typically starting lower to establish composition, then rising to lock in detail. Several ComfyUI nodes and hosted APIs now expose start/end CFG percentages as standard controls.
Smoother guidance variants
Newer research variants reduce the overshoot inherent in classic classifier-free guidance, delivering high-CFG prompt adherence with far fewer artifacts. Support is spreading through popular samplers, and it is worth checking whether your pipeline offers one before you accept burned highlights.
Common CFG Mistakes to Avoid
- Cranking CFG to fix a weak prompt. Detail comes from specificity, not from the slider.
- Using SDXL-era values on distilled models. A CFG of 7 on a Flux-based model will produce mush.
- Ignoring the sampler. Some samplers behave very differently at high guidance — test both together.
- Assuming a value transfers between checkpoints. A setting that sings on one model can fail on the next.
Conclusion
The CFG scale is not a quality dial — it is a control that measures how hard the model leans into your prompt versus its own learned instincts. Low values invite creativity, mid values deliver balance, and high values demand compliance at the cost of realism. The most reliable workflow in 2026 is simple: start at your model’s recommended range, lock a seed, adjust in single increments, and pair any high-CFG experiment with a rescale setting. Master that loop, and you gain genuine AI image control instead of hoping the next roll of the dice lands better than the last.