When is there a negative prompt?

The Negative Prompt field appears for every image and video model that takes a prompt and is hidden for audio and avatar models; what you type is sent to the provider, but some models do not read it, so their prompting guide tells you to describe what you want instead.

Last updated:

When you see it

The field sits under the main prompt in the Workspace. It is shown whenever the selected model accepts a prompt and is not an audio or avatar model. That covers most image and video models.

It is hidden for:

  • audio models — speech and music take a script or a description, not exclusions;
  • avatar / lip-sync models — they animate a picture to a sound file and have no use for it;
  • models that take no prompt at all, such as image-to-3D rows.

What happens to the text

Whatever you type is sent to the provider as negative_prompt. Whether it is honoured depends on the model:

  • Models whose parameter list on the model page includes negative_prompt read it as intended.
  • Models without that parameter ignore it — and a few (Grok Imagine, Nano Banana) have no negative channel at all, so "no rain" in any field can produce rain. Their prompting guides say to write the scene you want instead.

Habits that work everywhere

  • Describe presence, not absence: "clear sky" rather than "no clouds".
  • Use the negative field for model families that document it, and keep it short.
  • Let the feather check tell you when a model would misread your exclusion.

When you switch models

Your negative prompt is carried over as long as the new model takes prompts. Moving to a model that takes none clears both fields; the text is not restored when you switch back.

Frequently asked

Does every model obey it?

No. Diffusion-style models use it; several newer models have no such input and ignore it or, worse, read the words as things to include.

Is my negative prompt kept when I switch models?

Yes, while the new model accepts prompts. Switching to a model that takes no prompt clears both fields.

Where do I learn a model's stance on it?

On the model page: if negative_prompt is in its parameter list the provider reads it; if not, the prompting guide explains how to steer away from things in prose.

More questions