If you’re new to prompt engineering and don’t know much about working with generative AI systems, creating a prompt that generates beautiful images with DALL-E can be challenging.
Here are three tips for writing a great prompt that will yield a perfect DALL-E image.
Describe Your Image As if It Already Exists
A lot of new users make the mistake of writing instructions to DALL-E as if it’s a designer. For example, they’ll write something like “Can you make a photo of some fruit on a table?” The results can be kind of wonky.
DALL-E is not a chatbot; it’s designed to take a prompt and create exactly what that prompt describes, not to process and understand instructions.
Instead of telling DALL-E what to do, write your prompt as if you’re describing an image that already exists, for example, you could write: “A photograph of an ornate bowl of fruit on a table”
Describing your desired image as if it’s already there in front of you is the best way to generate exactly what you need.
Use Keywords to Control the Style
If you simply give DALL-E a prompt, like “A bowl of fruit on a table,” the system will try to guess the style you want. For that kind of prompt, it created the above still-life photograph, for example.
If you want to control the look of your final image, use style keywords in your prompt to control the output.
You can do this in a variety of ways. For example, you can say: “An oil painting of an ornate bowl of fruit on the table.”
If you want something that looks more corporate, you could say, “A stock illustration of an ornate bowl of fruit on a table.”
You can get very creative with these keywords! For example, you can attach something like “Steampunk charcoal drawing of a bowl of fruit on a table.”
Getting Photorealistic Results from DALL-E
If you want your end result to look like a real image, it helps to attach a descriptor like “Photorealistic.” If you’re trying to create a photo, you can still apply descriptors as if you were describing the technical details of the photo you want.
For example, you could say, “A photorealistic bowl of fruit on a table photographed with a Leica camera using a 70mm lens.”
Avoid Numbers, Text and Specifics
Although DALL-E is great at imagining completely new scenes, there are still areas where it struggles. The biggest one is in generating anything that is highly specific. For example, if you ask it to generate an image with text, it will probably yield something text-like, but words will be misspelled or formatted incorrectly.
Likewise, if you ask the system for a specific number of objects in your output image, it may or may not follow your instructions. To play to the system’s strengths, avoid asking for specific numbers of objects, as well as for text.
Here’s a result for the prompt “Five apples on a table with a sign reading Apples for Sale” that demonstrates both issues.
If you need text in your image, create the basic image first with DALL-E and then use another tool like Canva to add text after the fact.
If you need an image with a certain number of objects, one approach is to keep entering the same prompt until DALL-E generates an image with the right number. You can also ask DALL-E to create an image of one of the objects and then manually duplicate it in Photoshop.
Generative AI systems will probably get better at generating highly specific results in the future. For now, things like text and counting are not their strong suit. If you can work within those limitations, you’ll be happier with your output.
Conclusion
DALL-E is a powerful system, but as with any computer system, your outputs are only as good as your inputs. Use these tips to level up your prompt engineering game and get the most from this powerful generative AI system.