I had AI draw fish and it created a swarm of creatures, so I'm posting them here as a tribute #RecentLearning #ThanksToGeminiAndChatGPT #Hobbies #AIArtAndPrompts #AIImageGeneration #CreativeRecord #WorkingMomVisitingHome

Good morning. I am Mari Kaneda.
I'm visiting my parents' home for Golden Week.

When I took my eldest son and my husband took our second son to our respective parents' homes, it actually became much easier for the adults, including the grandparents, in terms of physical stamina.

I'm not sure if it's the reaction from returning to work or just being completely relaxed during the visit, but my brain isn't working at all.

But I want to post something on Note, so I decided to share some rejected illustrations from the Sakana AI series.

Schools of fish were quite difficult, so please feel free to laugh at these unfortunate rejected images!

Gemini and the rejected images

It already has a strong ama diver vibe to it, doesn't it?
It's more like capturing aliens than fish.

The outfit, which is hard to tell if it's Japanese or Chinese, seems to further highlight the
ama diver vibe.

I think this applies to all SD (Stable Diffusion) based tools, but I've been feeling lately that the difficulty lies in not being able to find the model's unique quirks unless you use the model (Checkpoints, LoRA, etc.) repeatedly.

But I like the world-building and the composition, so I'll throw some rough prompts at Nano Banana 🍌 to see if I can fix it somehow.

女性の周りを魚の群れが自由に泳いでいるイラストにしたいです。
特に、
女性の左横の魚紛いの生き物
女性が握っている魚紛いの生き物
イラスト右下の魚紛いの生き物
この3つも魚に変えてくれますか？
今のイラストのトーンを徹底的に守り、魚以外は変更しないこと。

A big catch...

Shrimp and small fish are casually mixed in.

I thought about how I forgot to fix the fish-like creature in the top right,
or that I needed to change the aspect ratio,
and various other things, but

I started to wonder.
Actually, there's something a bit off about it.

Getting the fish to appear can be handled with prompts or other AI corrections, but I find this 'not quite working' feeling a bit interesting.

I used to think model quirks were things to be fixed, but

I thought for a moment.

Well, for now, the spirit of let's just do it!.
In the end, I chose an illustration with intense eyes generated in a 16:9 aspect ratio ▼

Doesn't it look nothing like a face predicting the future?
Setting that aside, it has personally become one of my favorite thumbnails.

After publishing the article, Gemini asked me what I would write next, so I continued the conversation.

It gave me a copy.
Thank you, banana🍌

I thought it would reply with something like“That’s great!”to the illustration I uploaded to the banana, but it actually sent an illustration back.

And what’s with the“Let’s meet again.”part? If Gemini had facial expressions, I thought it would probably have a smug look like,

I imagined it would look like that.

By the way, I also tried again with ChatGPT, which has significantly improved its image generation skills.

Personally, I might like the color palette of Gemini's revisions better.

By the way, ChatGPT gave me some advice on how to master multiple AIs.

It said I should be quite conscious of those points.

The thing is, the more strictly you constrain generative AI, the more 'correct' it seems, but it becomes characterless.

Conversely, it seems that leaving some imperfect output actually brings out the unique personality of the model.

Once again,I see!I learned something new.

The rejected ones that were close

Personally, I decide on thumbnail illustrations intuitively based on things like:

・Does it have strong eye-catching elements?
・Does it seem to fit the world of the article?
・Does the composition overlap with recent past articles?
・Does it have a sense of unity when lined up with past articles?

I decide based on these factors. This time, the composition overlapped with the past article below, so it was rejected.
It was rejected.

Since no fish were appearing at all, I forced a solution using prompts.

The hardest part was that instead of fish,birdskept appearing in swarms.

It seems the model I used this time had a tendency to produce birds more easily than fish. Since over 20 bird illustrations were generated, I've picked some out to share with you.

Finally, by usingemphasis syntaxmore strongly, I succeeded in generating something that looks like a fish.

I wrote about emphasis syntax in this article ▼

Tracing the history of the rejected drafts

For the first illustration, I asked Gemini,

and it gave me a prompt, so,

(masterpiece:1.2), (best quality:1.4), 1girl, from below, cinematic composition, looking up, awe and wonder, deep ocean, underwater,
(silhouette of a large, elegant school of fish:1.3) swimming across the sun,
(volumetric lighting:1.3), shafts of light piercing the water, god rays,
caustics patterns on face and clothing, depth of field,
(original character design:1.2, teal blue bun hair, unique multi-colored eyes, teal geometric earrings),
(subtle bioluminescent data particles:1.1) mixed within the fish school,
cyber-organic fusion, fluid motion, flow state, emergent behavior,
colors of deep cerulean, turquoise, with tiny hints of magenta and gold, soft pink highlights from image_0.png

here is the illustration created by pasting that prompt as is ▼

I generated about 120 illustrations like this.
80% of them are creature illustrations.

Even though I used about 4 or 5 of them as thumbnails or illustrations in three articles, quite a few illustrations end up never seeing the light of day.

Still, I realized that sometimes it's better to use them as they are rather than trying to fix them perfectly, which felt a bit like the philosophy behind Sakana AI.

Sakana AI doesn't aim to build one perfect model,
but rather follows the philosophy of combining imperfect things to make them work.

I'm very happy you've read this far.
Thank you for sticking with me!

Next time,

is what I plan to post about from the perspective of the client.

If you have any requests like 'I want to hear about this!',I will do my best to cover them in an article, so please feel free to leave a comment😊

Depending on the content, I might make it paid.
In that case, it will be included in the paid magazine below▼

Have a great Golden Week♪