Short answer: a negative prompt is a second, separate list of things you do not want in the image, such as "blurry, extra fingers, watermark". The model steers away from those words while it follows the main prompt. Stable Diffusion and Midjourney (with --no) use one. Flux, Nano Banana and GPT Image have no negative field, so there you describe what you do want instead.
Every prompt in the Prompt Gallery comes with a negative prompt. Here is when to use it and when to fold it into the prompt itself.
How does a negative prompt work?
Diffusion models build an image in small steps. At each step the model predicts two directions: one towards your prompt and one towards a neutral "anything". Classifier-free guidance pushes the image away from the neutral prediction and towards your prompt.
A negative prompt replaces the neutral prediction with your list of unwanted things. Now each step pushes away from "watermark, blurry, extra fingers" as well as towards your subject. That is why it works best for things that tend to appear on their own (artefacts, text, a cluttered background) and less well for anything the main prompt asks for.
Which models have a negative prompt?
| Model | Negative prompt? | How to say "leave this out" |
|---|---|---|
| Stable Diffusion, SDXL | Yes, its own field | Comma-separated tags in the negative field |
| Midjourney | Yes, as a parameter | --no text, people at the end of the prompt |
| Ideogram | Yes, in the advanced settings | Short list in the negative field |
| Flux | No field in most tools | Describe the wanted result positively |
| Nano Banana (Gemini) | No field | Say it as an instruction: "keep the background empty" |
| GPT Image | No field | Say it in a sentence, and say what must not change |
Tools change their settings often, so check yours. If you see a box called "negative prompt", "exclude" or "avoid", it is the same idea.
In Midjourney, --no plants is the same as giving "plants" a weight of minus one half (plants::-.5). You can list several things after one --no, separated by commas.
Why can "no cars" in the main prompt backfire?
Writing "a street with no cars" in the main prompt often gives you a street with cars. Text encoders are good at nouns and weak at negation, so the word "cars" pulls cars into the picture and "no" barely registers.
Describe the result you want instead:
- Not "a street with no cars", but "an empty, quiet street at dawn".
- Not "no text", but "a plain, unmarked surface". (A real negative prompt for text is fine.)
- Not "no blur", but "tack-sharp, fine detail".
Google's own advice for Gemini image prompts says the same: describe the scene positively rather than listing what to leave out.
What should you put in a negative prompt?
Keep it short and specific. A long list of everything bad dilutes each word, and some words, such as "ugly", mean little to the model.
Good candidates are the failures that show up without being asked:
- Artefacts: blurry, jpeg artefacts, extra fingers, distorted hands, duplicate subject.
- Unwanted text: text, watermark, signature, logo.
- Wrong look: cartoon (when you want a photo), plastic skin, oversaturated.
- Job-specific failures: for a product shot, "redesigned packaging, misspelled label, floating product, missing shadow".
That last group is where a negative prompt earns its keep. The Product on White prompt in the gallery, for example, uses this one:
grey background, gradient background, colour cast, blown highlights on the
product, halo around the edges, redesigned packaging, misspelled label,
props, text, watermarkEach item is a thing that goes wrong in real packshots, not a generic quality word.
How do you write the exclusion for Flux, Nano Banana and GPT Image?
When your tool has no negative field, rewrite the important items as a positive rule at the end of the prompt. Take the list above:
Pure white background with no gradient, even light with no colour cast,
clean edges. The product is exactly as supplied: same shape, label and
lettering. No props and no added text.Two things make this work. The rules describe a state ("pure white", "exactly as supplied") instead of only naming the bad thing. And they sit in their own sentence at the end, so they do not mix with the description of the scene. Short "no" phrases like "no props" are fine at that point, because the subject is already fixed.
When does a negative prompt not help?
- The main prompt asks for it. "A busy market" with
--no peoplefights itself. Change the main prompt. - The model ignores it. If something keeps appearing, it is often part of the subject in the model's mind (a "chef" comes with a hat). Describe the subject differently: "a cook in a plain t-shirt".
- It is too long. Twenty items push in twenty directions. Keep the three to eight that matter for this picture.
Where can you check this?
What is the short version?
- A negative prompt pushes the image away from a list of things at every step.
- Stable Diffusion, Midjourney (
--no) and Ideogram have one. Flux, Nano Banana and GPT Image do not. - Without a field, describe the wanted result positively and add a short rule sentence at the end.
- List the failures that appear on their own, not quality words and not things the prompt asks for.
Every prompt in the Prompt Gallery shows its negative prompt next to the main one, ready to copy into the field or to rewrite as a rule. For the order of the main prompt itself, see how to write an image prompt that works, and the course lesson on using gallery prompts in Midjourney, Flux and Nano Banana shows where the negative prompt goes in each tool.
