Before our eyes, artificial intelligence (AI) is revolutionizing image creation, allowing us to generate some stunning visuals and explore as far as our imaginations will take us. Given the nature of AI and how it generates images, I thought it would be useful to compile the best practices for leveraging AI models, such as DALL-E, to help you unlock the full potential of AI-generated images.

Understanding The Limitations
Before diving into the image creation process, it’s helpful to understand the capabilities and limitations of AI models. They have been trained on large amounts of data, including images from the Internet, and generate images based on learned patterns rather than a comprehensive understanding of the world. This creates a number of limitations.

First, they usually don’t have specific knowledge about real-world locations or unique buildings, unless widely depicted. For many landmarks, this means AI will produce more of an approximation than a photorealistic detail. In the images below, for example, note the difference in the spire of Smith Tower between the AI-generated image and a photo.

AI-generated image of Smith Tower in Seattle, iteration 2.
The AI-generated Smith Tower spire
Smith Tower in Seattle, Washington
The real Smith Tower spire

It’s also limiting in terms of generating coherent and meaningful textual content. DALL-E, for instance, leverages visual patterns learned from its training data to generate images; it doesn’t understand or generate text phrases. While it can generate character shapes as part of the image, requesting DALL-E to specifically generate accurate and meaningful textual content is not its primary function. In the images of Pike Place Public Market, you can see the odd characters DALL-E added.

AI-generated image of Pike Place Public Market in Seattle, iteration 1.
Pike Place Public Market in Seattle, iteration 1
Pike Place Public Market in Seattle, iteration 2
Pike Place Public Market in Seattle, iteration 2

If generating accurate and meaningful text is a crucial requirement, you’ll want to explore other models, like GPT, specifically designed for natural language processing (NLP) tasks.

Provide Clear and Specific Instructions
To improve your chances of getting the images you want, be sure to provide clear and specific instructions. Vague terms can lead to ambiguous interpretations. Instead, use detailed and descriptive language to articulate your vision. For example, instead of saying “cityscape,” you could specify “a vibrant cityscape at sunset with towering skyscrapers and reflections on the water.” By providing more details, you give the AI model a clearer understanding of your intentions. Bing’s Image Creator provides a handy framework: Adjective + Noun + Verb + Style.

Image of Bing's input formula of adjective, noun, verb and style.
The request template on Bing Image Creator.

Iterate and Experiment
Generating images with AI usually involves a fair amount of trial and error. If the initial results don’t match your expectations, don’t be discouraged. Experimentation is the key to success. Try different phrasings, word order, or variations in your text prompts. For instance, if you’re not satisfied with the first attempt at generating a forest scene, you can refine your prompt by specifying the types of trees, the lighting conditions, or even the mood you want to evoke. Through iteration, you can fine-tune the output and gain a better understanding of the model’s capabilities.

For comparison, I’ve generated two images of the Space Needle. The first uses the simple input of the “the space needle”. The second uses a more descriptive input, “the space needle on a sunny day, vibrant colors, dramatic, digital art.”

AI-generated image of the Space Needle using the input, the space needle
Output of “the Space Needle”
AI-generated image of the Space Needle using detailed input
Output of “the space needle on a sunny day, vibrant colors, dramatic, digital art”

In the sequence below, you can see how the iterations of Smith Tower evolve its details and environs, using the same input: “Smith Tower in downtown Seattle on a cloudy night, photorealistic.”

AI-generated image of Smith Tower in Seattle, iteration 1.
AI-generated image of Smith Tower in Seattle, iteration 2.
AI-generated image of Smith Tower in Seattle, iteration 4.
AI-generated image of Smith Tower in Seattle, iteration 5.
AI-generated image of Smith Tower in Seattle, iteration 6.
AI-generated image of Smith Tower in Seattle, iteration 7.

Use Reference Images or Real-World Examples
If you have a specific image or real-world example in mind, consider using it as a reference. Incorporating reference images helps the AI model grasp the specific details or style you desire. By providing visual examples, you enhance the model’s comprehension and increase the likelihood of generating images that align with your vision. For instance, if you’re aiming for a specific architectural style, share reference images showcasing that style to guide the AI model’s interpretation.

Output of “an outdoor patio of a futuristic house on a sunny day, vibrant colors, contemporary architecture”
Output of “woman laughing, colorful, vibrant, cubist style”

Balance Creativity and Realism
One of the most interesting aspects of AI-generated images is the ability to combine creativity and realism. AI models can produce unexpected and abstract results. However, striking a balance between creativity and realism can be a subjective choice. You can experiment with different prompts to achieve the desired aesthetic or level of abstraction. For example, if you want an AI-generated landscape, you can specify the desired elements like mountains, lakes, or forests while leaving room for the AI model’s creative interpretation.

Image of an abandoned mansion in a forest
Output of “abandoned medieval mansion in a fairytale forest”
Image of an indigo dahlia
Output of “indigo dahlia, macro photography”

Create Ethically
As creators, we must be mindful of the legal and ethical considerations surrounding AI-generated images. Just as with AI-generated text, we have to respect copyright laws and avoid misusing images obtained through AI models. While you may own the rights to an image you created with AI, if it’s derived from someone else’s work, you may be violating copyright law. Indeed, there are a number of intellectual property questions raised by AI that will ultimately need to be sorted out legislatively and, more likely, by the courts. When in doubt, check with your attorney.

Image creation with AI is an exciting opportunity for content creators to exercise their imaginations. By following these best practices, you can maximize your chances of generating stunning and perfect images for your project.