What is Guidance Scale? — one analogy

1The analogy

Picture shouting an order through a megaphone in a noisy, crowded market. Turn the volume too low and the words get lost in the noise, so whoever's listening catches only a rough idea and hands over the wrong thing. Turn the volume up to a reasonable level and the order comes through clearly.

But crank the volume all the way up and the sound distorts, and it actually gets harder to understand. The voice got louder, but the words turned to mush. Guidance scale is that volume dial.

One more thing: a megaphone only makes a voice louder — it doesn't fix what's being said. If the order itself is vague, shouting it louder just sends the same vague order across, only louder.

2In detail

Turn the volume down and the request gets lost

At a low value, the generator only gets pulled toward the request a little. The rest drifts along whatever plausible-picture habit it picked up during training. So the picture itself often looks natural, and unexpected layouts can turn up — but the very thing that was requested is sometimes missing from the frame, or it's swapped for something else. Words describing color or mood get ignored outright fairly often too.

A low value has its uses. It's good for keeping the request loose and generating several pictures to scan through the range of directions. The usual move is to find a direction that looks promising, then raise the value and generate again from there.

If a picture comes out hazy and lacking punch, a low value can be the reason. At a low value, the pull at each step is weak, so shapes never quite finish forming and end up blurry. Adding more steps doesn't clear that haziness up much either.

Turn the volume up and it sticks to the request

Raise the value and the pull toward the request gets stronger at every step. The requested subject shows up without fail, and instructions about color or layout get followed closely too. In exchange, several pictures generated together start looking alike. Shouting the same request loudly pulls every answer toward the same spot.

Looking alike cuts both ways. When several pictures need to feel like a matching set, that consistency is convenient; when the point is to browse a wide range of candidates, there's nothing left to choose between. It's worth setting the value differently depending on what stage of the work you're at.

The point where the sound distorts

Keep raising the value and at some point the picture starts looking wrong. Colors blow out like a fluorescent light, bright areas wash out to white, and outlines turn thick and stiff. The subtle texture goes first. Pushing this hard at every step shoves the result out past the range where plausible pictures actually sit.

Where that distortion starts differs from tool to tool. Hold the seed and the sentence fixed, step the value up like a staircase, and generate five or six pictures side by side — the value where colors start to blow out becomes obvious. Just below that is usually the comfortable range to work in day to day.

A megaphone can't fix what's being said

Raise the value and the request gets followed more faithfully — it doesn't mean the outcome matches what was actually wanted. A vague request just gets followed vaguely, but faithfully. When the result isn't what you hoped for, reaching for the value first is usually wasted effort; rewriting the sentence to be more specific comes first.

The same dial also sets how hard what's written in the exclusion field gets pushed away. Raise the value and the push-away force rises right along with the pull-toward force, so a long exclusion list combined with a high value stacks up and breaks the picture faster.

It's tied to the step count too. A higher value means a bigger move at every step, so raising the value calls for raising the step count a bit as well, to keep things from getting rough.

3More precisely

At every step, the generator produces two predictions: one that treats the request as a condition, and one made with no condition at all. If something is written in the exclusion field, that text stands in for the unconditioned prediction instead. The gap between the two predictions is the direction the request pulls in, and guidance scale is the number that decides how much that gap gets amplified before the step is taken. Turned all the way down, the amplification is negligible and the request barely registers; turn it up, and the step moves further in that direction. Push it too far and the result lands outside the range where real pictures cluster, which is why colors blow out and shapes stiffen.

The analogy breaks down in one place. A megaphone keeps a voice's meaning intact no matter how loud it gets, but guidance scale actually reshapes the picture itself as the value rises. A distorted sound is a problem on the listening end; a picture broken by too high a value is something the generator itself produced that way. And where a megaphone's volume knob works about the same on any unit, guidance scale's comfortable range shifts quite a bit depending on the tool and the sampler.

4Try it yourself

Diffusion Explainer ailearn.space Compare how closely the picture sticks to the same sentence as only the guidance value changes

5Common misconceptions

It's easy to think turning the value up makes the picture better, but actually it only decides how closely the result sticks to the request, and pushed too far it falls apart instead.
It's easy to think maxing it out gets everything in the request, but actually colors blow out and shapes stiffen until even the requested subject becomes hard to recognize.
It's easy to think a low value is useless, but actually it's handy for scanning a wide range of candidates during the exploring stage.

7One-line summary

In shortGuidance scale is the volume dial that decides how hard a request gets pushed, and the picture gets lost when it's turned too low and torn apart when it's turned too high.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02