Images, video, voice and music
Negative prompt
In short
A negative prompt is text that tells an image generator what to leave out, such as ‘blur’ or ‘watermark’. It is a separate instruction from the positive prompt describing what the picture should contain12.
- Also called
- Negative text, Negative prompting
- Topic
- Images, video, voice and music
- Level
- Beginner
An image prompt describes the desired picture. A negative prompt describes unwanted objects or qualities. Some systems pass the two inputs separately, then use both when guiding the generation process12.
For example, an image request can ask for a bowl of fruit while the negative prompt contains ‘bananas’. Amazon’s Nova documentation says to put the unwanted item in its `negativeText` field rather than writing ‘no bananas’ in the main prompt1.
A negative prompt is not a guarantee or a universal filter. Support and behaviour vary by model. NVIDIA describes the field as an advanced feature, and its documentation says changing the seed with other inputs fixed changes the output2.
Negative prompts appear in image-generation interfaces and APIs, including systems that expose fields called `negativeText` or `negative_prompt`12. They are most useful for a specific omission, not as a long list of generic quality words.
Example
You ask for a photo of a mountain cabin at sunrise and put ‘people, text, watermark’ in the negative prompt. The generator will try to avoid those elements, but you still need to inspect the result because the instruction is not a guarantee12.
Why it matters
Negative prompts give image makers another way to steer a result when the model keeps adding an unwanted element. They work as model instructions, not as an editing eraser or safety promise12.
Questions people ask
What should I put in a negative prompt?
Name the objects or qualities you want to reduce, such as ‘watermark’ or ‘bananas’. Amazon advises using the unwanted word itself rather than phrases such as ‘no’ or ‘without’1.
Is a negative prompt the same as saying ‘no’ in a prompt?
Not necessarily. Amazon says its model does not understand negation reliably in the main prompt, so ‘a fruit basket with no bananas’ can signal bananas. It recommends putting ‘bananas’ in the separate negative field1.
Do negative prompts always work?
No. They are model-specific guidance, and results can still contain the unwanted element. NVIDIA labels negative prompts an advanced feature rather than a guaranteed exclusion2.