It is a dial that you have most likely looked at, and perhaps never, really, understood. It is a silent slider in nearly all ai image generators interfaces, called CFG Scale or guidance Scale. The majority drag it between 7 and 12 and wish the best. But what does this number do? That alters all that you create images.
Without the jargon coma, let’s break it down.
The means of CFG Scale
CFG is an abbreviation of Classifier-Free Guidance. Sounding like a doctoral thesis, right? Yet the gist of it is very easy. The CFG scale informs the model of the degree to which it should adhere to your text prompt. High number = it is more attentive. Low number = it becomes rogue.
Imagine it is as if you were employing a painter. You provide them with a reference image and description. That CFG of 1 implies that they peep at what you are reading, and paint whatever the mood they feel like, perhaps a fever dream, perhaps a masterpiece. A CFG of 20? They’re using a ruler to gauge pixel distances, and they’re attempting to sweat over every word you typed and they hardly breathe without your authorization.
Both extremes are not right. They are simply other tools.
The scale usually goes between 1 and 30, but most of the models have a sweet spot. On the low end, images are loose, painterly, abstract. At the extreme they become stiff, even unnaturally sharp, occasionally blown out in contrast and colour. You can easily get it very wrong on both extremes.
Low CFG: Accidents Made Art
CFG of 1-4 is essentially pouring a cocktail in your model and switching the lights off. It is not so blind to your prompt as to worship it. The result? Surprising pictures.
This scale is indeed practical in the study of textures, abstract art and experimental visual art. When you say “ocean at dusk,” a low CFG may give you what hardly appears to be water, but which has a mood so powerful that it halts your scroll. That is where sometimes it ends.
Unpredictability is the snare. You have to get many refusals, before you strike a big one. It’s a high-variance game.
Artists that create to explore creativity are likely to camp here. This model is more of a partner and not an implementer. You present the seed of an idea; it presents weird interpretations.
The Middle Ground (CFG 6 -12): The location of the majority of people
This is the workhorse range. The majority of the default settings are placed here, and not without reason. The model is a real-time attentive one yet stylistically interpretative. Faces are like faces. Skies are skies. But there’s still breathing room in the output.
In portrait work – particularly when you are using a free AI headshot generator – this range deals with likeness and composition without entering the uncanny. The products are professional, but not sterile. It is not as easy to strike that balance as it sounds.
What is interesting about this range is the way it responds differently in model architectures. SDXL at CFG 7 doesn’t behave the same as SD 1.5 at CFG 7. Flux models react to guidance yet again differently. You can’t just memorize one number and call it a day. Every model has its personality and the CFG scale is the way you negotiate with it.
High CFG (15-30): Accuracy at a Cost
Turn it up to 20 and the model takes your prompt like a life raft. Every adjective matters. Each noun is questioned. CFG 20 In Red apple on a marble table will provide you with the reddest apple, the most obviously marble table, and very little room to deviate artistically.
This would be perfect, right? Greater control, greater accuracy?
Not always. Very large CFG values bring about artifacts. Colors begin to be oversaturated. Edges get crunchy. Faces can distort. The model, in its attempt to do justice to each word, almost disembowels itself in the process. It is the pictorial version of a person echoing your words to you over and over again until they are screaming and have lost their sense.
High CFG performs optimally using short and accurate prompts. The less moving the parts you provide the model with, the more it is responsive to extreme guidance. Assuming that your prompt is three lines and you have a dozen adjectives, a CFG of 22 will give you chaos that nevertheless perfectly fits all the words you typed.
Prompt Length and CFG Are a Team
This is the section most guides do not cover. CFG scale does not act alone, but it is in combination with prompt complexity. A low CFG, minimalist prompt provides you with crisp, focused outputs. High CFG with a complex prompt provides you with a visual argument that has all the elements competing.
On the other hand, really strange things are the results of long prompts at low CFG. The model absorbs all that, and decides that it does not need to stick to it, and selects perhaps two or three ideas to pursue. Could be interesting. Usually isn’t what you wanted.
A practical rule: the longer your prompt, the smaller your CFG. Allow the model a leeway to make decisions. The narrower your prompt, the closer to the point, the higher you can push guidance.
Steps, Sampler, and CFG: The Holy Trinity.
CFG scale doesn’t exist in a vacuum. It operates with your number of steps and your sampler. Further steps tend to enable higher CFG values to converge correctly – the model has additional iterations to balance its requirement to obey instructions with the real image it is creating.
High CFG is more easily capable of artifacting at low step counts (10-15) when compared to low step counts. The model doesn’t have enough runway to sort itself out. At 30 or more steps, the same CFG values could yield clean, sharp results.
CFG is also treated differently by different samplers. DPM++ 2M Karras is significantly more permissive of greater guidance values than Euler a. DDIM has a propensity to act consistently along the scale. When you are regularly getting blown-out or oversaturated at your favorite CFG, then changing samplers is a good place to start.
