Reflections on AIBATO: Seeing the Beauty of the Moment Between Probability and Inspiration

Hello, this is Karabee.

Actually, did you know that an event like this was held at the AI Expo in Makuhari Messe this week? (　・ω・)ノ

It's an event that has been heard of here and there recently, aiming to turn AI illustration into an e-sport, but this is said to be the first time it has been held on this scale in Japan.
I had also entered, and to my surprise, I received notification that I had passed the preliminaries,
and was suddenly invited to the venue to participate in the main tournament.

The result—somehow, I was the runner-up (；ﾟДﾟ)

Detailed footage of the tournament will apparently be streamed somewhere at a later date, so I'll leave you to watch that... Today, I want to talk about the "essence of AI art" that I unexpectedly caught a glimpse of (or so I felt) during the matches.

……………I feel like I've set the bar quite high from the start, but let's just go with it (￣▽￣;)

・Selection and Dialogue

To begin with, the question of "Is AI imagery art?" is still a topic where opinions are sharply divided. For now, let's set aside those fundamental aspects and focus on the fact that I believe there are broadly two routes to creating works with image generation.

1. The [Selection Method]: Drawing a vast and inexhaustible supply of ideas from LLMs (AIs like ChatGPT, Claude, or Gemini) and adopting the excellent ones that align with your intent.

2. The [Dialogue Method]: Directly inputting the scene you envision as a prompt and refining it based on the results.

I think the AIBATO participants this time were also divided into these two types.
Of course, there are people who use both, but at least I am the type that sticks exclusively to the second one.
The first is often compared to a director, but the second might be closer to Shogi. Even before generating the first image, the finished image is already formed in my brain, and in fact, several generation results that the AI might produce are already floating in my mind. However, what the AI actually produces might be prediction A or it might be prediction B. This is literally a world of probability. It's common for neither to appear and for C to come out instead, but I take that output and make my next move.

The generation I usually do has no restrictions and many convenient auxiliary functions, but in an art battle, those auxiliary functions are almost unusable, and I have to complete one picture in a super short time of 10 to 15 minutes.
Therefore, the aforementioned

1. Constructing the finished form in my brain while simultaneously imagining the AI's output prediction in response to the theme

2. Translating the finished form into a prompt

3. Generating

4. Considering revisions based on the output results and reflecting them in the prompt

I end up repeating steps 3 and 4 of this workflow with almost no time in between.
Then, something interesting happens: at some point, my hands suddenly start typing words that seem to have no relation at all. I am no longer thinking with my left brain; I am in a state where I am typing prompts before I can even verbalize them in my mind.

In fact, in the finals, I received this comment from the judges (from around 2:36 in the video on the left).

The word "chemical reaction" comes up, but personally, I think "mutation" might be more accurate.
However, the continuous changes in the picture that occur when you operate an image generation AI—a tool that removes pixel noise based on probability—according to the flashes of inspiration born from a brain running at full capacity. And as a result, a single image that is nothing like the original finished vision, but which I can assert has significantly higher quality.

―――That moment might have truly been "AI art."

Looking back on it now, I can't help but think so ( ˘ω˘ )