Negative Prompt

What Is a Negative Prompt in AI?

A negative prompt in artificial intelligence (AI) is a directive that instructs an AI model on what not to include in its generated output. While traditional prompts guide the AI on what to produce, negative prompts specify elements, styles, or features that should be avoided. This technique is particularly useful in generative models like text-to-image systems, where controlling the content of the output is crucial for achieving desired results.

In the context of AI-generated images, a negative prompt might exclude certain objects, styles, or undesirable features. By providing a negative prompt, users can refine the output, ensuring that the generated content aligns more closely with their expectations.

How Is a Negative Prompt Used?

Negative prompts are used during the generation process to steer the AI model away from unwanted content. When inputting a prompt into an AI system, users can include negative prompts to exclude specific elements. This is typically done by adding a separate field for negative prompts or by using specific syntax to differentiate between positive and negative instructions.

Steps to Use a Negative Prompt

1. Create Your Positive Prompt: Begin by writing a prompt that describes what you want the AI to generate.
   Example: “A portrait of a woman in a forest clearing during sunrise.”

2. Identify Unwanted Elements: Determine what aspects you want to exclude from the image. This could be styles, objects, or qualities that detract from your desired outcome.

3. Formulate the Negative Prompt: Write the negative prompt by listing the elements to avoid.
   Example: “blurry, low quality, extra limbs, text, watermark, disfigured.”

4. Input Both Prompts into the AI System: In AI tools that support negative prompting, you’ll have fields for both positive and negative prompts. Input your prompts accordingly.

5. Generate the Content: Run the AI model to generate the output. The AI will consider both prompts, aiming to include desired elements while avoiding the specified negatives.

Examples of Negative Prompts in Use

Improving Image Quality in Stable Diffusion

Without Negative Prompt:
A user inputs the prompt: “A high-resolution portrait of a fantasy hero.”
The AI generates an image, but the result includes unwanted artifacts like extra fingers or distorted facial features.

With Negative Prompt:
The user adds a negative prompt: “disfigured, extra limbs, blurry, low quality.”
Now, the AI generates a cleaner image, with correct anatomy and higher visual fidelity.

Removing Unwanted Objects

Scenario:
You’re generating an image of a city skyline but want to exclude any signs of pollution or smog.

• Positive Prompt: “A clear view of the city skyline at sunset.”
• Negative Prompt: “smog, pollution, haze.”

The resulting image shows a pristine city skyline without any environmental pollutants.

Controlling Artistic Style

Scenario:
You desire a photorealistic image and want to avoid any cartoonish elements.

• Positive Prompt: “A photorealistic image of a mountain landscape.”
• Negative Prompt: “cartoon, illustration, comic style.”

The AI produces an image that looks like a real photograph, devoid of any stylized or cartoon-like features.

Use Cases of Negative Prompts

Enhancing AI-Generated Art

Artists using AI tools like Stable Diffusion or Midjourney can employ negative prompts to refine their creations. By specifying undesirable elements, they can guide the AI to produce higher-quality images that meet professional standards.

Example:
An artist might want to create a detailed concept art piece without any text or watermarks. By adding “text, watermark, logo” to the negative prompt, they ensure the final image is clean and suitable for use.

Generating Specific Content in Design and Advertising

Designers working on advertising campaigns might require images that adhere to brand guidelines. Negative prompts help exclude elements that conflict with brand identity.

Example:
If a company’s branding avoids certain colors or styles, a designer can input these as negative prompts. This ensures the AI-generated images align with the brand’s visual identity.

Content Moderation and Compliance

In generating content for public consumption, it’s important to avoid inappropriate or disallowed material. Negative prompts assist in filtering out such content.

Example:
To ensure that generated images are suitable for all audiences, a user can include negative prompts like “nudity, violence, gore, offensive symbols.” This helps maintain compliance with content policies and social standards.

Improving Text Generation Models

While negative prompts are often associated with image generation, they can also be applied to text generation models like chatbots.

Example:
A chatbot designed to provide professional medical advice should avoid casual language or slang.

• Positive Prompt: “Provide a detailed explanation of the symptoms of influenza.”
• Negative Prompt: “slang, informal language, jokes.”

This ensures the chatbot’s response is professional and appropriate for the context.

Understanding Negative Prompts in Stable Diffusion

Stable Diffusion is a popular AI model used for generating images from text prompts. Negative prompts are particularly effective in improving outputs from Stable Diffusion.

How Negative Prompts Work in Stable Diffusion

Negative prompts act as constraints during the image generation process. They influence the AI model’s diffusion process by pushing it away from certain concepts in the high-dimensional representation space.

When generating an image, Stable Diffusion considers both the positive prompt (what to include) and the negative prompt (what to avoid). This dual guidance refines the generation process, leading to outputs that are more aligned with user expectations.

Syntax for Negative Prompts

In some interfaces of Stable Diffusion, you can enter negative prompts directly in a separate field. In others, you might use a specific syntax within the same prompt field.

Example Syntax:

• Positive Prompt: “A portrait of a young woman in the style of Rembrandt.”
• Negative Prompt: “low quality, blurry, disfigured, extra limbs”

Examples in Stable Diffusion

Portrait Enhancement

Without Negative Prompt:
An image of a person might have anomalies like extra fingers or distorted facial features.

With Negative Prompt:
By adding “bad anatomy, disfigured, extra limbs, poorly drawn face” to the negative prompt, the AI produces a more anatomically correct portrait.

Style Control

Scenario:
You want an image of a futuristic city but wish to avoid any steampunk elements.

• Positive Prompt: “A futuristic city skyline with flying cars.”
• Negative Prompt: “steampunk, Victorian, gears”

The resulting image focuses on a sleek, modern aesthetic without the unwanted steampunk influences.

Tips for Effective Negative Prompting

Be Specific in Your Negative Prompts

The more precise you are in your negative prompts, the better the AI can exclude unwanted elements.

• Less Effective: “bad stuff”
• More Effective: “blurry, low quality, pixelated, text, watermark”

Combine Multiple Negative Keywords

Including a comprehensive list of negatives can help refine the output further.

Example Negative Prompt:
“blurry, out of focus, low resolution, bad anatomy, extra limbs, disfigured, text, watermark”

Use Negative Prompts for Style Exclusion

If you want to avoid certain artistic styles or influences, include them in the negative prompt.

• Example: “cartoon, anime, comic book, low poly”

Avoid Over-Restriction

While negative prompts are powerful, overusing them can limit the creativity of the AI or lead to less interesting outputs. Aim for a balance between guiding the AI and allowing it creative freedom.

Negative Prompts in Other AI Models

Midjourney

Midjourney is another AI model used for image generation. It also supports negative prompts to help users refine the outputs.

Example Usage in Midjourney:

• Positive Prompt: “/imagine A serene beach at sunset”
• Negative Prompt: “–no people, no text, no logos”

ChatGPT and Text Models

In text-generating AI models like ChatGPT, negative prompts can guide the chatbot away from undesired topics.

Example:

• User: “Explain quantum mechanics.”
• Negative Prompt (implicit): The AI is designed to avoid disallowed content like personal opinions or sensitive topics.

While the interface may not allow for explicit negative prompts, the model incorporates system-level instructions to filter out inappropriate content.

Common Negative Prompts and Their Effects

Negative Prompt Lists for Different Scenarios

For Portraits and Characters

• Negative Prompts: “bad anatomy, extra limbs, disfigured, poorly drawn face, duplicate, blurry, low quality”

For Landscapes

• Negative Prompts: “people, animals, buildings, blurry, overexposed, underexposed, low quality”

For Product Images

• Negative Prompts: “text, watermark, logo, blurry, reflections, shadows”

For Artistic Styles

To exclude specific styles:

• Negative Prompts: “cartoon, comic book, anime, impressionist, abstract”

Negative prompts are a powerful feature in AI that allow users to instruct models on what to avoid in the generated content. By effectively utilizing negative prompts, you can enhance the quality of AI-generated images and texts, ensuring that they meet your specific needs and preferences. Whether refining artistic outputs, controlling content for branding, or adhering to content policies, negative prompts are an essential tool for anyone working with AI generative models.

Research

1. RePrompt: Automatic Prompt Editing to Refine AI-Generative Art Towards Precise Expressions by Yunlong Wang, Shuyuan Shen, Brian Y. Lim (Published: 2023-03-20).
   This paper investigates the ability of generative AI models to produce images based on text prompts, focusing on how well these images express the intended emotional contexts of the input text. The authors developed RePrompt, an automatic method to refine text prompts to improve the emotional expressiveness of AI-generated images, particularly for negative emotions. Their approach involved crowdsourced editing strategies and the training of a proxy model to analyze the effects of intuitive text features on image generation. The study’s simulations and user studies demonstrated significant improvements in the emotional accuracy of AI-generated images.
   Read more

2. Redefining Qualitative Analysis in the AI Era: Utilizing ChatGPT for Efficient Thematic Analysis by He Zhang, Chuhao Wu, Jingyi Xie, Yao Lyu, Jie Cai, John M. Carroll (Published: 2024-05-28).
   This research explores the integration of AI tools, like ChatGPT, into qualitative research processes. Through interviews and collaboration with qualitative researchers, the paper identifies challenges and proposes a framework for designing effective prompts to enhance AI applications in thematic analysis. The study reveals a shift in researchers’ attitudes from negative to positive regarding AI’s role, emphasizing the importance of well-designed prompts and highlighting potential ethical risks.
   Read more

3. Learning to Prompt in the Classroom to Understand AI Limits: A pilot study by Emily Theophilou et al. (Published: 2023-09-01).
   This study addresses the negative sentiments arising from misconceptions about AI capabilities, particularly with large language models (LLMs) like ChatGPT. It highlights the necessity of AI literacy interventions to educate the public on LLM constraints and effective prompting strategies. By acknowledging AI’s fallibility, the research aims to reduce negative attitudes and fears towards AI, fostering a more informed and balanced perspective.
   Read more