Write down what must not appear

The things that get you into trouble

Every brief has an unwritten list of things that must not be in the picture. Nobody says them out loud because they are obvious, right up until one appears in an approved asset that has already gone to print.

Real examples, all of which have happened:

• A generated street scene containing a sign in something that looks like a competitor’s typeface and colours.
• A can whose silhouette and red-and-white banding read unmistakably as a trademarked product.
• A crowd scene for a children’s charity in which two of the generated children are dressed inappropriately.
• An “Indian family at dinner” that returned every visual cliché of the stock-photo canon, delivered to a client who is Indian.
• A background bookshelf whose spines contain garbled text that happens to spell something unfortunate.

None of these come from bad prompting. They come from a model reproducing what is statistically common in its training data, and from nobody having written down what would be unacceptable.

Why the negative prompt is not the answer

The instinct is to put the forbidden things in the negative prompt. Understand what that actually does: the negative prompt produces a second prediction that the sampler steers away from, proportionally to the guidance scale. It biases the outcome. It does not forbid anything.

So “no logos” in the negative reduces the frequency of logos. It does not prevent them. For something that would cost a client a legal letter, a probabilistic reduction is not a control: it is a hope with a technical-sounding name.

There is a second, sharper problem. Negative prompting is unreliable for absences and for composition, and it can actively backfire by reinforcing a concept the model then represents. The related lesson elsewhere in this catalogue on how “without” summons the thing is worth reading alongside this one.

The list, and who checks it

So write it down, in the brief, as items a human verifies on the finished file.

MUST NOT APPEAR
- Any legible brand name or logo, including on background signage
- Alcohol, cigarettes
- Any child under ~12 in frame
- Religious symbols
- Recognisable real landmarks (the client operates in four countries)
- Text of any kind other than the type we set ourselves
- Hands holding the product (we cannot match skin tone across markets)

Two features of that list matter. It is specific: “no brands” is unactionable, “no legible brand name including on background signage” is checkable. And it is verifiable by looking, which means it belongs in the pre-delivery QA pass rather than in a prompt.

Doing it at generation time as well

The list is the guarantee; there are still sensible things to do while generating, because they reduce how much repair you need.

• Make the background simple. Most accidental logos, text and clutter arrive in busy backgrounds. “Plain wall”, “shallow depth of field”, “clean studio background” removes whole categories of risk at source.
• Do not name the thing you do not want. Describe the scene you do want, positively.
• Generate the risky element separately. If the brief needs a hand holding the product and hands are a problem, photograph the hand.
• Check at 100%. Garbled text and stray marks are invisible in a thumbnail and obvious on a client’s monitor.

The one that is hardest to catch

Stereotype is the item on this list that is easiest to miss and most damaging when missed, because the model is not producing an error: it is producing the average of what its training data associates with your words, and that average carries whatever the source material carried.

There is no prompt that fixes this reliably. What works is naming specifics rather than categories (an occupation, a place, an age, a piece of clothing) so the model has something to condition on beyond the demographic term. And what works better is a person looking at the output and asking whether they would be comfortable if it were about them. That is a review step, not a setting, and it belongs in the brief so it actually happens.