Negatives and must-not requirements

A useful exclusion can be checked by another person. “Not weird” cannot. “No extra fingers, duplicated limbs, fused hands, or cropped fingertips” can. “No clutter” is vague; “plain wall with no shelves, posters, cables, or visible appliances” defines the expected background.

Group exclusions by purpose:

• Content: no weapons, alcohol, crowds, animals, or real people.
• Brand: no logos, trademarks, packaging, uniforms, or recognizable product designs.
• Layout: no text in the image, no objects in the reserved copy area, no cropped subject.
• Quality: no duplicated objects, broken geometry, illegible labels, or inconsistent reflections.
• Safety: no identifying documents, graphic injury, sexualized depiction, or dangerous behavior.

Keep the list short enough to prioritize. If ten exclusions are essential, the task may need multiple generation stages, compositing, or manual editing rather than one overloaded prompt.

Pure negation leaves an empty space the model must fill. Tell it what should appear instead.

Instead of:

No busy background.

use:

A simple matte cream studio wall with a faint soft shadow; no furniture or wall decoration.

Instead of:

No dramatic lighting.

use:

Even, diffused daylight with gentle contrast and natural skin tones; no hard spotlight or colored gels.

Instead of:

No text.

use:

Blank unprinted label area; all typography will be added later. No letters, numbers, signatures, or watermarks.

The positive direction constrains the replacement. It also makes the request easier to transfer between tools that support negatives differently.

Prompts frequently contain hidden conflicts:

• “Crowded street” and “no people.”
• “Authentic vintage magazine cover” and “no text.”
• “Mirror reflection” and “no duplicate person.”
• “Dynamic motion” and “everything tack sharp.”
• “Exact product package” and “no brand elements.”

When outputs repeatedly violate a must-not, inspect the brief before blaming the model. Decide which requirement has priority. You may change “crowded street” to “dense street architecture with parked bicycles but no pedestrians,” or replace a magazine cover with an “editorial illustration inspired by print layout, with blank geometric text blocks.”

Negation can also backfire because mentioning an unwanted object still activates its visual concept. If “no red car” repeatedly yields red cars, omit the car concept and positively describe the desired empty road. Behavior varies by model, so test rather than treating this as a universal rule.

Negative prompts can reduce stereotypes, but “no stereotypes” is not a sufficient fairness strategy. Define respectful, varied representation positively. If a scene depicts a profession, avoid relying on demographic defaults. Specify a contextually appropriate mix only when identity is relevant, and do not use protected traits as decoration.

Similarly, “no copyrighted characters” does not prove an output is clear of protected expression or trademarks. Ask for original attributes rather than a close substitute: describe silhouette, palette, mood, and function without naming a living artist, franchise, or branded character. Then review the actual result for resemblance.

For safety-critical imagery, prevention and inspection are both required. A prompt showing laboratory work should specify gloves, eye protection, tied-back hair, and correct equipment. A knowledgeable reviewer should still check whether the depicted procedure is safe. The model may satisfy “wearing goggles” while inventing a hazardous setup elsewhere.

Before generating, label every constraint:

Hard must-not: A violation makes the asset unusable or unsafe.
Soft avoid: A violation reduces quality but may be repairable.
Post-process: A detail is better handled after generation.

For a nonprofit website hero, that might become:

• Hard: no identifiable real people, no logos, no generated text, clear empty left third.
• Soft: avoid overly saturated colors and staged stock-photo expressions.
• Post-process: add approved logo and headline in the design system.

After generation, inspect at full resolution. Look at hands, eyes, repeated patterns, reflections, text-like marks, background faces, and object joins. Thumbnail appeal can hide defects. If the image will be cropped, inspect every intended crop too.