Good Practices: Prompts for Generative AI Art

Understanding the importance of well-crafted prompts in AI-generated art is pivotal in harnessing the full potential of diffusion models like Dall-E or StabilityAi's SDXL. These AI models are designed to interpret text-based inputs and transform them into visual content. The quality and clarity of your prompts directly influence the output, determining not only the accuracy but also the creativity and uniqueness of the generated art. By mastering the art of prompt creation, you can guide these models to produce more consistent and sometimes even rare expressions of creativity, pushing the boundaries of what AI-generated art can achieve.

For instance, a poorly defined prompt might lead to generic or unexpected results that do not align with your vision. On the other hand, a well-constructed prompt can unlock the model’s ability to generate intricate and meaningful art that resonates with your intended message or aesthetic. This guide will walk you through the essential techniques for crafting effective prompts, helping you to achieve your desired outcomes more reliably.

While AI has advanced a lot, it still can not guarantee the exact imagination that you possess. We can try to get as closer as possible by explaining in as much detail as possible.

Clarity and precision are the cornerstones of successful prompt creation. When you provide a clear and concise prompt, the AI model has a better chance of generating an output that closely matches your expectations. For example, consider the prompt "A Siberian tiger wearing a leather jacket." This prompt is specific—it describes not only the subject (a Siberian tiger) but also what it is doing or wearing (a leather jacket). This level of detail helps the AI focus on the exact image you want.

In contrast, a vague prompt like "Cats raining from the sky" is open to broad interpretation. The model might generate anything from cats literally falling from clouds to a surreal abstract pattern featuring cats, which may differ significantly from what you envisioned. Therefore, always aim for specificity. Think about the key elements you want to see in the image—whether it’s the subject, action, environment, or style—and ensure that your prompt conveys these elements clearly.

• Good Example: "A Siberian tiger wearing a black, leather jacket."

  • This prompt is specific and clear, describing the subject (a Siberian tiger) and what it is doing or wearing (a black, leather jacket).

• Bad Example: "Dogs falling from heaven."

  • This prompt is vague and open to broad interpretation, leading to unpredictable results.

Descriptive language is your tool for guiding the AI to produce visually rich and engaging content. When crafting a prompt, include as many relevant details as possible. This might involve specifying colors, textures, styles, or even emotional tones. For instance, instead of saying "A sunset," you could describe "A fiery red and orange sunset over a calm ocean with soft pink clouds." The added details provide the AI with more information to create a vivid and visually appealing image.

Similarly, if you want to evoke a certain mood or style, be explicit. A prompt like "A serene, pastel-colored landscape in the style of impressionist art" gives the AI clear directions about both the color palette and the artistic style, helping it to produce an image that closely aligns with your artistic vision. Additionally, if you have a particular trait or feature you want to emphasize, you can tag these separately on our platform to further refine the output.

• Good Example: "A fiery red and orange sunset over a calm ocean with soft pink clouds."

  • This prompt provides detailed information that allows the AI to create a vivid and visually appealing image.

• Bad Example: "Sunset."

  • This prompt is too general and lacks the detail necessary for creating a specific and compelling image.

The process of generating AI art is inherently experimental, and one of the most powerful techniques at your disposal is the use of prompt variations. By tweaking and experimenting with different prompts, you can discover unexpected and unique results that may not emerge from your initial ideas. For example, starting with a base prompt like "A futuristic city at night" and then modifying it with variations such as "A futuristic city at night with neon lights" or "A futuristic city at night in a cyberpunk style" can lead to vastly different outputs.

This approach not only enriches your creative assets but also helps you learn how the model interprets different inputs. Sometimes, even small changes in wording or the addition of a new element can significantly alter the output, revealing new creative possibilities that you might not have considered. Keep in mind that some of the most striking AI-generated art comes from this iterative process of exploration and refinement.

• Good Example: "A futuristic tech city at night with neon lights."

  • This prompt builds on a base idea and adds specific elements, leading to a distinct and creative output.

• Bad Example: "Cityscape."

  • This prompt is overly broad and will likely result in a generic image.

Contextualising your prompts by incorporating references or specific genres can significantly enhance the quality and relevance of the generated art. AI models like Dall E or SDXL respond well to prompts that include clear references to known art styles, genres, or themes.

In addition to artistic styles, you can leverage themes or broader contexts to steer the AI’s creative direction. If you’re working on a collection of images with a "retro" or "steampunk" theme, including these keywords in your prompts will help the AI generate consistent and thematically cohesive artwork. For example, a prompt like "A steampunk-inspired airship floating above a Victorian-era city" clearly conveys both the theme and the desired aesthetic, leading to more accurate and visually aligned outputs.

• Good Example: "A portrait of a woman holding an umbrella in the style of 19th-century impressionism."

  • This prompt clearly guides the AI in terms of a known artistic style, helping to set the tone, composition, and color scheme.

• Bad Example: "Portrait."

  • This prompt is too generic and lacks the context needed to produce a meaningful image.

Creating AI-generated art is often an iterative process. Rarely will your first prompt yield the perfect result. Instead, view each output as a step closer to your final vision. After each generation, take time to evaluate the results—what worked well, and what didn’t quite meet your expectations? Use this feedback to refine your prompt, making adjustments to the details, language, or structure as needed.

For instance, if your initial prompt "A futuristic cityscape at dusk" results in an image that’s too dark or lacks detail, you might refine it to "A brightly lit futuristic cityscape at dusk with glowing skyscrapers and bustling streets." This iterative process allows you to gradually hone in on the exact output you desire. Additionally, experimenting with different wordings or adding/removing elements can help you explore a broader range of creative outcomes, leading to potentially surprising and delightful results!

• Good Example: "A brightly lit futuristic cityscape at dusk with glowing skyscrapers and bustling streets overlooking a moonlit lake."

  • After refining an initial prompt, this version includes more specific details to guide the AI towards the desired outcome.

• Bad Example: "A picture of my cat."

  • This prompt is vague and lacks the necessary details to produce a specific result. No matter how many times or even with context you try this prompt, a generic diffusion model will give you no better results that just writing "Cat"

Pitfalls to watch out for when creating prompts for AI art.

• Avoid specific proper nouns like names and places that are harder to contextualize like lesser known cities, celebs or TV shows.
  
  Avoid prompts like "Jerry singing along the choir in Neverland"

• Avoid copyrighted content like names and places that are trademarked.
  
  Avoid prompts like "Luke's lightsaber in Star Wars: Return of the Jedi"

• Avoid political content. Some of the models specifically have guard rails against controversial content while others just don't perform so well
  
  Avoid prompts like "Trump v Harris"

• Avoid asking the model to generate textual content. Diffusion models are not good at generating text unless they have been fine tuned
  
  Avoid prompts like "Generate a textual poster for my coffee shop named 'The Java Roast Corner'"

• Avoid NSFW prompts. All the models would either reject the prompt or just simply not generate NSFW content

There are several resources to advance your knowledge and skills on prompt engineering. These are some of the most useful ones we could gather:

• OpenAI Prompt Engineering Guide

• DigitalOcean Prompt Engineering Best Practices

• Brex's Prompt Engineering Guide

• Techniques to improve readability

• Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing

• Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

• Prompt Engineering 201: Advanced methods and toolkits

Enjoy exploring the vast creative possibilities of AI-generated art. By mastering prompt creation, you can unlock new realms of artistic expression, discovering unique and unexpected visuals that might have been impossible to conceive otherwise.

Happy creating!