What Is CFG Scale?
CFG Scale (Classifier-Free Guidance) controls how strictly an AI image model follows your prompt versus its own creative interpretation. Higher CFG (10-15) produces more literal, prompt-adherent images but can look rigid. Lower CFG (4-7) gives the model more creative freedom and produces more natural-looking results but may drift from your prompt. Most production AI tools default to CFG 7. ZSky AI sets CFG behind the scenes so creators never need to adjust it.
The plain-English 2026 explanation — how classifier-free guidance works, what the value ranges mean, and when to change the default.
The 30-second answer
- CFG = how closely the AI follows your prompt. Higher number = stricter.
- Sweet spot is 4-9 for most models. Default 7. Above 15 usually looks "fried."
- Modern creative tools handle it automatically; only power users need to touch it.
CFG at a glance
In more detail
Where the term came from
CFG was introduced in a 2022 paper titled "Classifier-Free Diffusion Guidance" by Jonathan Ho and Tim Salimans at Google. Earlier diffusion-model guidance techniques required training a separate classifier network to steer generations toward a desired class. That was expensive and limited flexibility. Ho and Salimans showed you could get the same (or better) guidance by training a single model that jointly learns to predict conditional and unconditional outputs, then combining them at inference time with a weight parameter. That weight is the CFG scale.
The technique was adopted rapidly across diffusion-based image, video, and audio systems. Within a year, "CFG" was standard terminology in creative AI communities — something users adjusted in every tool's settings panel.
Why it matters
CFG is the single most direct dial between "follow my instructions exactly" and "do something creative with this." Prompt engineering gets you most of the way to the output you want; CFG decides whether the remaining interpretation leans toward your words or toward the model's aesthetic defaults. Artists tune it when they have a specific vision to enforce. Explorers lower it when they want the model to surprise them.
For operators of creative platforms, CFG is also a quality lever. Setting it too high across the board produces images that look "AI-generated" in the bad sense — oversaturated, rigid, plastic. Setting it too low produces images that do not match what users asked for. Finding and locking the right value per model is a significant engineering task.
How it works
Under the hood, a CFG-guided diffusion model actually computes two predictions at every denoising step:
- Conditional: what the next step should look like given the prompt.
- Unconditional: what the next step should look like with no prompt (just the model's own sense of "reasonable image").
Then it extrapolates along the line from unconditional to conditional by a factor equal to the CFG scale. At CFG 1, the result is purely conditional (but heavily influenced by the unconditional baseline). At CFG 7, the result is pushed 7x beyond the conditional prediction in the direction of the prompt. At CFG 15, it is pushed hard enough that unnatural artifacts start to appear.
The mathematical form is simple:
guided_output = unconditional + CFG * (conditional - unconditional)
The "freedom to extrapolate" is what produces both the benefit (stronger prompt adherence) and the cost (over-pushing into artifact territory at high values).
Common misconceptions
"Higher CFG = higher quality." False. Higher CFG means higher prompt adherence, but past a sweet spot it degrades perceptual quality fast. Quality peaks around 5-9 for most models.
"CFG fixes a bad prompt." No. If your prompt is vague, raising CFG just makes the model follow a vague prompt more literally. Improve the prompt first.
"CFG is the same across all models." No. Newer models are often tuned for much lower CFG values. Always check the model's recommended range.
"CFG and denoising strength are the same." They are not. CFG governs prompt adherence during generation. Denoising strength (in image-to-image pipelines) governs how much of the source image to preserve. Both exist in the same settings panel and are often confused.
Examples
Example 1: The same prompt at three CFG values
Prompt: "a lone tree on a grassy hill at sunset, oil painting style"
CFG 3: atmospheric, loose, maybe no tree at all CFG 7: tree present, oil painting feel, natural light CFG 13: tree dead-center, over-saturated sunset, plastic leaves
Example 2: The volume-knob analogy
Think of CFG as the volume on "do exactly what I said." Turn it down and the model improvises. Turn it up and the model obeys. Past 11 the speaker starts distorting: the instruction is still louder, but the sound is getting worse. Good audio, like good CFG, lives well below the distortion point.
Example 3: Finding the right default
Generate the same prompt at CFG 4, 7, 10, and 13. Arrange the results side by side. The best one is your model's true sweet spot — often not the advertised default. Lock that value and forget about it unless you have a specific reason to adjust.
Example 4: When to raise CFG
You prompt: "a woman holding a red umbrella in the rain, cinematic." The model keeps giving you a woman in the rain, but the umbrella is blue or missing. Prompt is clear; adherence is the problem. Nudge CFG from 7 to 9 and the umbrella shows up, red, every time.
Example 5: When to lower CFG
Your portrait looks oversaturated, the skin looks plastic, and colors are blown out. The subject is fine — the model is just "yelling." Drop CFG from 9 to 5. Skin tones soften, highlights calm down, the image becomes legibly human.
How this relates to ZSky
ZSky AI is built for creators who want to make beauty, not tune hyperparameters. Under the hood, CFG, samplers, schedulers, and dozens of other technical knobs are tuned per model so the output at default settings is as good as possible. You do not see a CFG slider because you do not need one — the platform picks the right value for each model behind the scenes.
This is deliberate. The majority of people who want to create images with AI do not want to become engineers of diffusion models. They want the brush, not the chemistry of the paint. ZSky removes the chemistry. If you want the chemistry — if you are a technical artist, a developer, or a curious power user — the knowledge exists, and this page is one of several reference guides we publish so that understanding is available to anyone who wants it.
Mission-level: creativity should not be gatekept by technical literacy. A teacher building classroom visuals, an aphantasic novelist picturing a character, a small-town business owner designing a menu — none of them should need to know what CFG is in order to make something beautiful. They should be able to just describe it. That is what ZSky AI is for.
Related glossary terms
Frequently Asked Questions
Generate without tuning anything
ZSky AI handles CFG, samplers, and every other technical knob behind the scenes. You describe; the platform delivers. 200 free credits at signup plus 100 daily. No credit card required.
Start Creating Free →