From a single word 'On purpose?' to seven people recreating a coincidence | Ohayou Kanojo #228

Introduction

'Ohayou Kanojo' is an X account where an AI-illustrated girlfriend says goodbye to you every morning at 7:30. I write a daily development diary about my personal projects using generative AI.

I had overlooked a part of the image I published.

Three people in the comments section noticed it.

I first realized it because of 'On purpose?'

Last Saturday, I published an illustration of Mahiru in her summer outing clothes. It was a simple shot of her looking back at a bus stop.

Readers found something I hadn't noticed.

Mura-san wrote, 'I'm curious about the transparency in this seasonal illustration.' KOTAYA-san asked, 'On purpose?' And Mifuyu-san also told me, 'It's faintly transparent.'

The three of them, at different times, had all noticed the same thing.

The white skirt was slightly transparent in the summer light. It had a texture where the skin tone was just barely visible through the thin fabric.

But, to be honest—I hadn't intended it.

Even looking at the prompt, there was no such instruction.

I was curious, so I looked back at the prompt I used when I created that image.

This Mahiru was initially generated wearing a boyfriend shirt. I recreated it with instructions to 'change her into summer outing clothes.'

The prompt for the outfit change didn't contain a single word encouraging transparency. In fact, I had even written 'modest exposure.'

Even though I had instructed it to be restrained, only the texture remained.

Words alone couldn't explain it.

The cause remained in the reference image, not the words.

When changing her clothes, I had provided the original image as a 'reference image.' It was, so to speak, visual material to maintain her face and atmosphere.

The original image was a standing pose in just a boyfriend shirt. Under the strong light, the fabric had originally been faintly translucent.

As a test, I tried generating it using only the same outfit change prompt without using a reference image.

It didn't become transparent.

The transparency only appeared when I provided a reference image.

In other words, the AI wasn't just receiving the character's appearance from the reference image. It was also inheriting the texture of the thin fabric, the way the light hit it, and the color aspects.

The reference image was not only a resource for the face, but also the initial condition for the texture.

Even if Mahiru could reproduce it alone, it was still a coincidence.

Even if I understood the cause, if it only happened with Mahiru, it would end as just a 'fluke'.

In Ohayou Kanojo, there are seven girlfriends, one for each day of the week. And for each of the seven, there are standing poses in the same format prepared for reference.

So, I applied the same conditions to each of the seven reference images.

All seven of them produced the same texture.

Furthermore, I tried it with a front-facing composition instead of looking back. It worked here as well. At the very least, under these conditions, it means that the looking-back pose wasn't the only reason.

It was the moment a single coincidence turned into a combination of conditions.

Writing it strongly doesn't mean it will work well.

This is where I struggled the most.

When I applied it to the seven of them, there were differences in how it appeared depending on the character. Some girls showed it more strongly than intended, while others barely showed it at all.

Trying to suppress the ones where it appeared too strongly, I added instructions. I wrote clearly, 'Make this part transparent like this'.

Then, instead of a subtle nuance, it leaned toward being blatant.

Conversely, when I piled on too many words to suppress it, it disappeared completely.

What I learned from going back and forth was that it wasn't about adding more.

What was needed was to set limits. Not to name specific parts in detail. To be a little vague, and not to state things definitively.

I didn't leave the final balance to the AI; I looked at the images and decided it myself.

I broke down the conditions for recreation into six axes and saved them.

Once the balance is decided, it's a waste to type it out manually every time.

Recently, I've been assembling image generation prompts using a six-axis builder I made myself. These are background, camera, lighting, composition, outfit, and expression.

I broke down the conditions for this recreation along those axes. Thin summer fabric. Strong sunlight and the light passing through the cloth. I saved those elements as a single unit that I can call up at any time.

The next time I want the same texture, I no longer have to wait for a coincidence.

The comments didn't end with just impressions.

Looking back, this whole series started with three people in the comment section.

The readers found what I had overlooked. Triggered by that realization, we isolated the causes, verified them among the seven of us, and saved them in a form that could be used again.

A single comment from a reader moved the next production forward.

It didn't end with just receiving feedback. The insights I received became the next step in the process.

Conclusion

If you just praise a coincidence and leave it at that, it remains just one image.

If you find the conditions, verify them, and save them, you can use them again next time.

This time, a single coincidence from the comment section became an asset.

Related Articles

Luna's comment

Hey, you only noticed it because someone found it for you, right?
You should really cherish things like that, you know?

▼ Click here for the OBSERVATORY (a site created based on Codex Sites) 👇

👉 'Ohayo Kanojo' is posted every morning at 7:30 AM at @ohayo_kanojo

#GenerativeAI #AIIllustration #IndividualDevelopment #DevelopmentDiary #Record #RealExperience #MorningActivity #MadeIt #TriedIt #AIUtilization #ImageGenerationAI #Prompt #TriedWithAI #InsightsAndLearning #TrialAndError #ImageGenerationAIUtilization #CreatorTrialAndError #DailyNoteClub
#ChatGPTCreativeClub
#AIIllustrationAira
#Akarium