Consistent Theming Using Generative A.I.

This article is part of a series about designing a board game using generative A.I. specifically DALL-E2, which is the A.I. we used to design our first game, The Rings of Saturn.

For a recap of what generative A.I. is, see here.

This article will focus on the difficulties of getting a consistent theme across multiple pictures and ways to negate that randomness.

In the following sections, unless otherwise noted, all pictures have been generated by DALL-E2.

The Problem

In The Rings Of Saturn, the theme of the game is space. To be more specific, when The Rings Of Saturn was first an idea, the theme was colorful, cartoonish, retro space. Below are a couple images that inspired the beginning thought process. This was 5 years ago so I am unable to attribute them.

Additionally, the theme was supposed to be congruent, meaning the same group of astronauts, spaceships, habitats, rovers should appear throughout the artwork.

When using a generative A.I. there are a couple issues you run into:

How do I describe in words what I want to see? This includes the color, the setting, the action occurring, the style and all the other miniscule details
How do I keep the same images propagating thru all my generations ?

To illustrate the first issue, consider the astronaut from the middle picture above eating a sandwich on Mars (left picture), how would you describe that? Lets start extremely basic and put this prompt into DALL-E.

Astronaut eating sandwich on Mars

As you can see DALL-E provides 4 images to our prompt and the left-most one is actually pretty good/close. In full transparency, I was not expecting this example to match as well as it does. However, we got lucky, the other three are not what we are looking for.

For the second issue, lets extend the theme to another scene.

Astronaut drinking tea on Mars

When comparing the two sets of prompts, they are close but Mars is slightly different between the two sets. Also, the astronauts have different suits on. This might be fine if your setting is a poor mining colony on Mars where everyone is cobbling together their spacesuits, but if your setting is a singular astronaut eating lunch on Mars this will not work.

Mitigating The Issue

There is not a sure fire way to fix these issues but there are a couple things we can do to get a more consistent end result.

Use the same style for all generations
Be extra descriptive
Seed your description with a picture
Iterate

Use The Same Style

There are all sorts of artistic styles you can use and DALL-E will give a more consistent end result. These can be classical styles such as the ones found here, but they also can be styles from pop culture such as steampunk or vaporwave. You can also use a specific artist as a style.

Below are a couple examples of our first prompt with different styles. Note, I chose vaporwave as an example because I think it is a close match to the Mars inspiration poster.

Astronaut eating sandwich on Mars, vaporwave

Astronaut eating sandwich on Mars, surrealism

Astronaut eating sandwich on Mars, rococo

Be Extra Descriptive

The more descriptions you provide, the closer the end result will be to your vision. In the example below I used descriptions of the area instead of using the proper noun Mars.

Astronaut eating sandwich on red sand, surrounded by red rocks with the night sky behind them, hyper realistic vaporwave

A benefit of the extra descriptions is now we can apply the same descriptions to another action and receive a similar result.

Astronaut drinking tea on red sand, surrounded by red rocks with the night sky behind them, hyper realistic vaporwave

Seed Your Prompt

DALL-E has a couple different ways to introduce consistency via seeding.

Editing
Outpainting
Generate Variations

To edit, you can upload a photo and edit it to remove the parts you don’t want. DALL-E will replace the empty section of the image with a new image as evidenced below.

Astronaut in red scratched spacesuit eating sandwich

Astronaut walking dog

To outpaint, you can upload a picture and expand that picture. DALL-E will take the prompt and continue it in a new section. You can also shrink a picture via another image editing software so your end result is the same as all the other prompts. Below, I left the frame in the picture so you can see what DALL-E used in its prompt.

Dog with leash, outpainting

On DALL-E, you can generate variations. This is handy if an image is close but not quite right. The end result will provide images similar but not the same.

Astronaut eating sandwich on Mars, variations

Iterate

Iterating is key in the process of developing a consistent theme using DALL-E. Iterating is a fancy way of saying try a lot of things and take the lessons from each generation. We have been iterating through out this article and have found a couple takeaways such as describing Mars instead of writing the word ‘Mars’.

To illustrate the iteration process one more time lets work on improving the space suit. I like to add each description separately so that I can see what effect the new description has on the generation.

Astronaut in turquoise spacesuit eating sandwich on red sand, surrounded by red rocks with the night sky behind them, hyper realistic vaporwave

As you’ll see with the next prompts, there doesn’t seem to be any difference due to the red highlights or red faceplates. Sometimes, it doesn’t work.

Astronaut in turquoise spacesuit with red highlights eating sandwich on red sand, surrounded by red rocks with the night sky behind them, hyper realistic vaporwave

Astronaut in turquoise spacesuit with red faceplate eating sandwich on red sand, surrounded by red rocks with the night sky behind them, hyper realistic vaporwave

Conclusion

Using the techniques described above will let you generate a consistent theme across all your A.I. generations.

As a final farewell, enjoy some Astronaut Pizza!

Astronaut in turquoise spacesuit eating pizza on red sand, surrounded by red rocks with the night sky behind them, hyper realistic vaporwave