Creating a new flip is not a hard process, but sometimes a good consistent story just doesn't come into mind, especially when it is bounded by a couple of words you have you use in it. I had an idea how to accelerate this process a little bit with some AI techniques that currently exist and are easily available for everyone. The main idea is to present a type of a story I want to get to a GPT-like model and use its response as a backbone for a flip’s story.

A good thing there is GPT-J — a 6 billion parameter, autoregressive text generation model that is openly available at https://6b.eleuther.ai. After some experimentation with it, I ended up with a prompt that looks like this:

Story 1: A four part consistent story told in four images built around these two words: Chimney — flue; Hotdog — frankfurter.

“Santa climbed into the chimney and then placed a gift under the christmas tree. In the morning a kind looks at the gift and sees that there is a hotdog inside.”

Image 1: An image of santa climbing into chimney
Image 2: An image of santa placing a gift under the christmas tree
Image 3: An image of a kid looking at the christmas gift
Image 4: An image of christmas gift containing a hotdog

Story 2: A four part consistent story told in four images built around these two words: Disk — Disc, flat circular plate; Decoration — Ornament, ornamentation.

“

The last quote sign is an important part of the prompt. Do not delete it, if you would like to use this prompt yourself. «Disk — Disc, flat circular plate; Decoration — Ornament, ornamentation.» part should be replaced with your key words.

The result I got:

A kid comes to his parent and says that he wants a disc for his christmas decoration. His parents ask him what kind of disc. He says a disk for a flat circular plate. His parents get him a disc for a flat circular plate. He decorates a christmas tree with it.

Image 1: An image of a kid showing a flat circular plate to his parent
Image 2: An image of a kid looking at his christmas tree without decoration
Image 3: An image of a kid looking at the flat circular plate
Image 4: An image of the kid putting the christmas decoration in a christmas tree

As you may see it is a consistent story with a proper storyline that indeed can be described in four images. It does not always work perfectly. Usually you have to run the prompt 4-5 times before getting any appropriate result. The only part left is to find such images and make a flip out of them. Here is where the problem lies… It usually not possible to find the exact images described and the story itself might need some tweaks to work. But yet this could be a good source of inspiration for a flip creationing process. For this case the flip I actually submitted looked like this:


This little experiment might make you think if creation of good flips could be fully automated? And my answer to this is almost certainly yes. As well as solving them. It might not be the easiest thing to do, but with the current state of AI it is definitely possible. Is Idena dommed than, cause the only reason it exists is to fight bots and distinguish people from machines? Here, I would say no. There are still plenty of things that only humans can do, although the number of such things goes down day by day. For example there is «The Abstraction and Reasoning Corpus» that proven to be extremely hard to solve with any AI techniques that exist today. So while flips might become absolite in the future, the idea behind Idena is very important for our digital society and if it will overcome the problems it faces it will no doubt live and thrive but with somewhat different ground underneath.