Skip to content

Learn

What Is a Negative Prompt? How It Works

A negative prompt is a second text input that tells an AI image or video model what to leave out, such as 'blurry, extra fingers, watermark'. It works through classifier-free guidance: at every denoising step the model is steered away from the negative prompt as well as toward the main prompt.

Start on the free Starter plan.

Negative Prompt - FunAI
Negative Prompt - FunAI

Steer Toward the Prompt, Away From the Rest

A negative prompt gives guidance a second direction. The image is pulled toward what you described and pushed away from what you listed, step by step, from noise to final picture.

Open Stable Diffusion 3.5 Large

A second prompt

A separate field for objects, styles and defects you want the model to avoid.

Guidance, not deletion

It reduces the chance of a concept appearing; it does not erase anything from an existing image.

Model-dependent

Many open image and video models accept it. Others are built differently and only read the main prompt.

Models with a negative prompt on FunAI

These image and video models list a Negative Prompt among their inputs. Open a model page to see the full input list and price per run.

  • •
    Stable Diffusion 3.5 Large:Text-to-image with Negative Prompt, Seed and Guidance scale (CFG) inputs.
  • •
    Qwen Image:Prompt-faithful text-to-image, strong with text in images, with a Negative Prompt input.
  • •
    HiDream I1 Full:Open-source model for concept art and illustration with a Negative Prompt input.
  • •
    Wan 2.2:Text-to-video with Negative Prompt, Seed and Guidance scale inputs.
  • •
    Veo 3.1:Google DeepMind's video model, with a negative prompt to guide the generation.
  • •
    Kling Video v3 Pro:Text-to-video with optional sound, a Negative Prompt and a CFG Scale.
Browse all image models
Negative Prompt Models

Negative Prompt explained

The main prompt describes what you want. The negative prompt lists what you do not want: unwanted objects ('people, cars'), unwanted styles ('cartoon, 3d render') or common defects ('blurry, low quality, distorted hands, text'). It became popular with Stable Diffusion, where a few defect words were an easy way to clean up results, and it is now an input on many open image and video models.

It works through classifier-free guidance. At each step the model predicts the image once with your prompt and once without it, then pushes the result in the direction of the difference. A negative prompt replaces the 'without' prediction: instead of moving away from a blank prompt, the model moves away from your negative text, so the concepts in it become less likely. Because it is a direction rather than a filter, it lowers the odds of something appearing instead of guaranteeing it is gone, and its effect grows with the guidance scale.

Use a negative prompt when a model keeps adding something you did not ask for, or to keep a style clean. Not every model has one. On FunAI, models such as FLUX.1 [dev], Nano Banana 2 and GPT Image 2 have no negative prompt input; for those, describe what you want in positive terms ('an empty street at dawn' instead of 'no people').

Tips for writing negative prompts

  • Never write exclusions in the main prompt: 'no hats' can add hats, because the model still reads the word 'hats'. Put 'hats' in the negative prompt instead.
  • Keep it short. A long copied list of defect words can flatten style and detail, and many of those words do nothing on newer models.
  • Target what you actually see in the results, not every problem you have read about.
  • Avoid contradictions, such as negating a concept the main prompt depends on; the two directions cancel out and the image gets muddy.
  • Remember the guidance scale amplifies the negative prompt too. If a negative seems too strong, lower the guidance before deleting words.
  • For video, describe motion problems you see, such as 'flicker' or 'warping', alongside the usual quality terms.

How to Use a Negative Prompt on FunAI

Feature visual preview 1
Feature visual preview 2
Feature visual preview 3
Feature visual preview 4
01

Choose a model with the input

For images, Stable Diffusion 3.5 Large, Qwen Image and HiDream I1 Full take a negative prompt. For video, Wan 2.2, Veo 3.1 and Kling Video v3 Pro do. The model page lists it among the inputs.

02

Get the main prompt right first

Run your positive prompt once without any negative prompt, so you know which problems actually need removing.

03

Add short, specific negatives

List what you saw go wrong as comma-separated words, such as 'text, watermark, extra limbs, motion blur'. Concrete nouns work better than vague ones like 'bad'.

04

Compare on a fixed seed

Keep the seed the same and run again with the negative prompt, so the only difference is your list. Trim it until it fixes the problem without dulling the image.

No public creations to show yet.

Everything you need to know