What Is CFG Scale in NSFW AI? Explained Simply (2026)

CFG scale, short for classifier-free guidance scale, is the dial that controls how strictly an AI image generator follows your prompt. A low value gives the model creative freedom and softer results, while a high value makes it stick tightly to your words and can look harsh. For most models, around 5 to 8 is a safe start.

If you have opened a generator and seen a slider labeled CFG scale or guidance, and had no idea what it does, this explainer is for you. CFG scale is one of the most misunderstood settings, and getting a feel for it separates flat, disappointing images from ones that actually match what you pictured. This page is definitional: it explains what CFG scale is and the visible effect it has on your image. When you want exact per-model numbers and a full tuning workflow, it hands you to the settings guide.

This article is for adults only and covers concepts intended for people 18 and over.

What CFG scale actually is

Every time you generate, the model is doing a quiet balancing act. On one side is your prompt, telling it what to make. On the other side is its own learned sense of what a good image looks like, regardless of your words. CFG scale is the number that decides how much weight to give your prompt in that balance. It is, in plain terms, the prompt-obedience dial.

Turn it up and you are telling the model to obey your prompt above all else. Turn it down and you are giving it permission to wander, leaning more on its own instincts and less on your exact wording. That is the whole concept. The name classifier-free guidance is just the technical term for the math that makes this steering possible, and you can safely ignore the jargon. What matters is the behavior: higher means stricter adherence, lower means looser and more creative.

It sits right alongside the negative prompt, which is the flip side, telling the model what to steer away from. In fact CFG controls the strength of both the positive pull and the negative push at once, which is why turning it up intensifies everything. Both fit into the broader process described in how AI image generation works, where guidance is the stage that sets how hard the denoising loop leans toward your words.

What it does to your image

The effect of CFG scale is easiest to understand at the extremes. At a very low value, the model barely listens to your prompt. Images come out soft, dreamy, and often miss details you asked for, but they can look natural and organic, with pleasant lighting and no harshness. At a very high value, the model clamps hard onto every word. It follows your prompt closely, but the picture often turns harsh: colors oversaturate, contrast spikes, edges get crunchy, and the whole thing takes on a fried, artificial look that experienced users call deep-fried.

The sweet spot is in the middle, where the model follows your prompt well while keeping natural color and detail. For most Stable Diffusion style models that middle sits somewhere around 5 to 8, though this depends heavily on the model you are using. Some newer distilled and Flux-style models are designed to run at much lower guidance, sometimes as low as 1 to 2, and pushing them higher wrecks the output entirely. This is why there is no single universal number, and why anyone who gives you one fixed value is oversimplifying. The settings overview for beginners shows where the slider lives, and the CFG and sampler tuning guide covers exact values model by model.

Low versus medium versus high

Here is the tradeoff at a glance. Read it as a spectrum, not three fixed boxes, and remember the numbers shift with the model, especially for distilled or turbo variants.

CFG level	Prompt adherence	Creativity	Artifact risk
Low (roughly 1 to 4)	Loose, may ignore details	High, soft and organic	Low, but can look vague or underbaked
Medium (roughly 5 to 8)	Follows the prompt well	Balanced	Low, the usual sweet spot
High (roughly 9 and up)	Sticks tightly to every word	Rigid	High, oversaturated and harsh

For most NSFW work on standard models, staying in the medium band gives you reliable, natural-looking results. Nudge upward when the model keeps ignoring an important detail you clearly described, and nudge downward when the image looks harsh or overcooked. CFG works hand in hand with the sampler and step count, so if you change one, expect to revisit the other, because the same CFG value can behave differently under a different sampler.

Why it matters most for NSFW output

CFG scale carries extra weight in adult generation for one practical reason: anatomy and pose accuracy depend on the model actually following your description. If guidance is too low, the model drifts and you get vague or incorrect anatomy because it is ignoring the specifics you wrote, so poses come out generic or wrong. If guidance is too high, the harsh oversaturation ruins skin tones and lighting, which matters enormously for realistic work where believable skin is the whole point.

The medium band is where you get accurate posing and description without frying the image, which is why most NSFW model cards recommend values in that range. If you find yourself constantly raising CFG to force the model to render a detail correctly, that is usually a sign the prompt itself needs work rather than more guidance. Strengthening and clarifying the description with the help of the prompt formula guide fixes the root cause, where cranking the dial only masks it and adds harshness.

How to adjust it in practice

A sensible workflow is to leave CFG at a moderate value while you get the prompt right, then adjust it last. Start in the middle of the recommended range for your model. Generate, look at the result, and make one small move. If the image is soft and ignoring detail, raise CFG a point or two. If it looks harsh, oversaturated, or crunchy, lower it. Change nothing else while you do this, so you can clearly see what the dial is doing. Once it looks right, lock it in and leave it, because a good CFG value tends to hold across similar prompts on the same model.

Keep the seed fixed while you test CFG. With the seed locked, the only variable changing is the guidance, so you can compare two renders directly and see exactly how the dial affects that specific image rather than chasing a moving target.

Where to find CFG in your tool

In almost every local interface, CFG scale is a labeled slider or number box near the sampler and steps, usually called CFG scale or simply CFG, and sometimes just Guidance on hosted tools. It is one of the main generation settings, so it sits in plain view rather than buried in an advanced menu. In node-based tools the value is a field on the sampler node, so the ComfyUI guide shows exactly where it plugs in if that is your setup. On a hosted generator it may be simplified to a small set of presets like more creative versus more precise, which are just friendly labels for low versus high CFG.

Wherever it lives, the way you use it is the same. Set it once to the middle of your model’s recommended range, generate, and only move it when the image tells you to, harsh means lower it, soft and vague means raise it. Because it is a top-level control, changing it is instant and needs no reload, which makes it easy to run a quick comparison at two or three values on a fixed seed.

Realistic versus anime and CFG

The ideal CFG range shifts a little with the kind of model. Realistic photo-style checkpoints tend to look best in the lower part of the medium band, because high guidance quickly turns believable skin harsh and plasticky, and realism is exactly where oversaturation is most obvious. If your realistic renders look sunburned or waxy, lowering CFG is one of the first things to try. Anime and illustration models, including the popular Illustrious checkpoints, often tolerate slightly higher guidance because their flatter shading and bold linework hide oversaturation better, and they sometimes want the extra adherence to render clean, specific poses.

This is a tendency, not a law, and the model card still wins. But it is a useful instinct: when you move from a realistic model to an anime one, do not assume the CFG value that looked perfect carries over. Re-test it, because the same number can read as gentle on one model and harsh on another. This interacts with clip skip and the sampler too, which is another reason anime and realistic workflows end up with different default settings even on the same machine.

When CFG matters most

CFG earns its keep in two situations. The first is a detailed, specific prompt where you need the model to actually honor every element, a particular pose, setting, and lighting all at once. There, a medium-to-slightly-higher value keeps the model from dropping details, which is common on complex scenes. The second is realism, where the wrong value is immediately visible as either mushy anatomy at too-low CFG or fried skin at too-high CFG. In both cases CFG is the difference between a render that matches your intent and one that misses.

For simple prompts on a well-behaved model, CFG matters less, and the default from the model card will usually just work. That is worth knowing, because it tells you where to spend attention. Do not agonize over one CFG point on an easy prompt. Do dial it in carefully when you are pushing a complex NSFW scene or chasing believable skin, because that is exactly where the dial repays the effort.

It also matters when you combine several LoRAs or a heavily weighted prompt, because those already push the model hard, and stacking high CFG on top can tip a balanced image into harshness. In that situation, easing CFG down a little often restores natural color while keeping the adherence you need. The broader point is that CFG is not a fixed setting you learn once, but a dial you rebalance whenever the rest of your setup changes, and the more demanding your prompt, the more that rebalancing pays off.

What to watch for

The most common mistake is cranking CFG scale to fix a weak prompt. Guidance controls how hard the model follows your words, but it cannot invent detail that your prompt never described. If an element is missing, add it to the prompt clearly; do not just raise the dial. High CFG on a vague prompt gives you a harsh image that is still missing the thing you wanted, which is the worst of both worlds.

The second trap is ignoring the model’s own recommendation. Distilled and turbo models often want very low guidance, and treating them like a standard model by running high CFG will produce a burnt mess. Always check the model card, because it usually states the intended CFG range for that model. The third is changing CFG in isolation from the sampler and steps. Because these settings interact, a value that looked great on one setup can look off on another, so treat them as a group rather than solo dials. And keep every subject adult and fictional.

Before and after: what a CFG change looks like

It helps to picture the visible shift as you move the dial, so you can recognize it in your own images. Start with a moderate value on a realistic portrait of an original adult character and the result looks natural: even skin tones, believable lighting, soft but present detail. Now push CFG well past the sweet spot. The same character comes back with skin that looks sunburned or waxy, colors that glow unnaturally, and hard, crunchy edges around hair and clothing. The prompt is being followed more literally, but the picture has lost its realism. That is the deep-fried look, and once you have seen it a couple of times you will spot it instantly.

Drop the dial far below the sweet spot instead and the swing goes the other way. The character softens, the lighting turns dreamy, and specific things you asked for, a particular pose or a detail of the setting, start to fade or vanish because the model is barely listening. The image can be pretty, but it is no longer your image. The lesson from watching both extremes is that the middle band is not a compromise, it is the target, and the whole skill is nudging toward it from whichever side you landed on.

How this connects to the rest of your settings

CFG scale does not decide image quality on its own; it is one voice in a small chorus. Your prompt supplies the actual content, the sampler and steps decide how cleanly the image is rendered, the checkpoint sets the overall look, and CFG governs how tightly all of that hews to your words. When people describe an image as overcooked, the cause is often high CFG combined with too many steps or an aggressive sampler, several settings pushing the same direction at once. The fix is rarely one dial; it is easing the group back toward balance.

This is why experienced users change one setting at a time and keep the seed fixed while they test. It is the only way to know whether the improvement came from CFG or from something else. Treat CFG as the obedience dial in that chorus, get the prompt and model right first, then use CFG to tune adherence, and reach for the full settings walkthrough when you want the exact numbers for your specific model.

The short version

CFG scale is the prompt-obedience dial: low lets the model be creative and soft, high forces it to follow your words tightly at the risk of a harsh, oversaturated look. There is no universal number because it depends on the model, but a common sweet spot for standard models is roughly 5 to 8, while distilled and Flux-style models want much lower. Stay in the middle band, fix missing detail in the prompt rather than by cranking guidance, keep the seed fixed while you test, and when you want exact per-model values and a full walkthrough, go to the CFG and sampler tuning guide.

Frequently asked questions

What is CFG scale in AI image generation?

CFG scale, or classifier-free guidance scale, is the setting that controls how strictly the model follows your prompt. A low value gives the model creative freedom and softer results, while a high value forces it to stick tightly to your words. Think of it as a prompt-obedience dial that balances your description against the model’s own instincts about a good image.

What CFG scale value should I use?

It depends on the model, so treat any number as a starting point. For most standard Stable Diffusion style models, a value around 5 to 8 is a reliable sweet spot. Newer distilled and Flux-style models are built for much lower guidance, sometimes 1 to 2. Always check the model card, which usually lists the intended range for that model.

What happens if CFG scale is too high?

The model clamps hard onto every word of your prompt, which sounds good but often looks bad. Colors oversaturate, contrast spikes, edges get crunchy, and the image takes on a fried, artificial appearance that users call deep-fried. For realistic NSFW work this ruins skin tones and lighting, so back the value down toward the medium band to recover.

What happens if CFG scale is too low?

The model barely listens to your prompt, so it leans on its own instincts and often ignores details you asked for. Images can look soft, dreamy, and natural, but they may miss specifics like pose or anatomy. If your results feel vague or keep dropping requested elements, raising CFG toward the middle usually tightens them up noticeably.

Does CFG scale work the same on every model?

No, and this trips up many beginners. Standard models like SD 1.5 and SDXL want a medium value around 5 to 8, but distilled and turbo models are designed for very low guidance, and running them high produces a burnt mess. The right range is model-specific, so the model card’s recommendation always beats any universal number.

Can raising CFG scale fix a bad prompt?

No. Guidance only controls how hard the model follows the words you wrote; it cannot invent detail your prompt never described. If an element is missing, the fix is to add it clearly to the prompt, not to crank the dial. High CFG on a vague prompt gives you a harsh image that still lacks the thing you actually wanted.

How is CFG scale different from a negative prompt?

They are two sides of the same steering. CFG scale controls how strongly the model pulls toward your positive prompt, while the negative prompt tells it what to pull away from. High CFG intensifies both directions at once. Used together, a clear positive prompt at a medium CFG plus a focused negative prompt gives the cleanest control over the result.

Why does my image change when I adjust sampler and CFG together?

Because they interact. CFG scale, the sampler, and the number of steps all shape the same denoising process, so changing one shifts how the others behave. A CFG value that looked great with one sampler can look off with another. Treat these settings as a group and expect to revisit CFG when you change the sampler or step count.