Dall-E2 can create astonishing pictures with just words. Nevertheless, when using these pictures as illustrations for a book or comic, the characters and art style should be consistent.
It is currently not possible to upload a character in Dall-E2 and tell the AI to use that character when drawing new pictures. The following tips will help you achieve consistency, at least to some degree.
Your prompt should contain the same keywords
Make sure you repeat your chosen art style in your prompts over and over again once you have settled on one.
When I directed the children’s book “The indecisive chameleon”, I used “Oil painting, art for children”. These keywords helped unify the art style and character design.
Batch matters
Dall-E2 seems to take into account prompts made in the same batch. As an example, let’s look at my chameleon. Each of these drawings above was made during the same batch and their style looks the same.
A few days later, these drawings were made using the same keywords (oil painting, children’s art). They are consistent, but slightly different from the previous attempts.
Characters and context matters
Dall-E2 cannot (or don’t want to) work with specific characters or settings. Dall-E2 should therefore be used in a medium that tolerates some inconsistency. I created with Dall-E2 a children’s book starring a chameleon, so by nature it looks slightly different every time. So, for a children's book about a chameleon, the end result seems consistent. Here’s a short video of the book in full.
Choose your AI
There are different styles for different models. Dall-E2 has the most “style agnostic” design. Because its outcomes are so diverse, they cannot be attributed to a single source. Models like Midjourney, however, seem to have a signature style. Keeping consistency with these models is easier, though the style may be overused very soon. Take a look at this wonderful video clip done in MidJourney.
Refining prompts
Dall-E2 has been trained on millions of photos, many of which are stock photos.
Therefore, itis more comfortable showing images it has been trained on, which may be biased. Here are DallE2 images for “A portrait of a kindergarten teacher” — Only women here.
When I tried making Dall-E2 show a “woman with her 3 months old baby in a cafe” DallE had problems, while “A young family with a 3-month-old baby in a cafe” was much more accurate.
Dall-E2 will therefore do the job when the prompt makes sense.
Here’s an example of a monkey eating a banana. No problems there.
A banana eating a monkey, however, didn’t go as planned.
If you want Dall-E2 to draw a banana eating a monkey, you’ll have to describe it in detail. “A giant banana with a big mouth is taking a bite from a monkey”
This is closer to what I had in mind, but still not perfect.
By telling DallE2 the style you’d like to see the picture in, and ensuring it’s in line with the prompt, you’ll get more accurate results. “A giant banana with a big mouth is taking a bite from a monkey, a surrealist digital art “
Having established a character and art style, we can continue this batch with the giant banana, incorporating a little more creativity from DallE2. In the end, no one likes to be micromanaged, and that seems to hold true for AI models as well. “A giant monster banana with a big mouth is the fear of all monkeys, a surrealist digital art “.
Community
Millions of prompts have been sent to Dall-E2, and the community is very generous. Dall-E’s official discord has many tips and tricks. This dictionary illustrates how Dall-E2 reacts to different filters, camera angles, and famous creators. This should come in handy when tying to envision how the prompt will be without actually spending it.
Here’s to a new era of art and a happy prompting to everyone!
— -
When not falling in love with AI I’m writing about games, game design and UX on my blog.
