Currently experimenting to create cute anime-style images with Image V3.

As I write this article, for some reason the models available for use with high-speed generation off on this site are now only Image V3 (hereafter V3) and GPT Image 2... 😓

I think it should fix itself over time, but in case many other models remain unusable, I'm writing this to discover and share know-how for generating anime-style images with V3 that are just as cute as those from other anime-oriented models 🖋️

This article was written in a hurry, so the content may lack organization, but I hope it can serve as a reference for everyone, even just a little 🙏

【① "Very cute" is effective 😄】

V3 tends to generate monotonous anime-style images with simple prompts. I found that adding the prompt "very cute" makes it easier to produce adorable illustrations!

(From my experience, V3 is good at cute lolis)

Example 1: anime style, fubin-kawaii style, very cute girl, a girl dancing with her eyes closed and smiling, long white hair, black long-sleeved clothes with wide sleeves, white skirt, grassland, blue sky, sparkling, lens flare, high resolution

【② Making them close their eyes is also one technique 😌】

V3's anime-style images sometimes have unnaturally large eyes compared to other models (the pupils are also dilated, which can even be creepy 👀)

So by having them close their eyes, you can both avoid unnaturalness and increase the healing feeling—two birds with one stone! (Same as V2)

Example 2-1: anime style, very cute girl, sleepy expression, cute cheeks, cute little hands, fluffy sweater, long-sleeved pajamas, hugging a dog plushie, on the bed, sparkling, pastel colors, masterpiece, ultra detailed

Example 2-2: Anime CG, very cute girl, cute cheeks, long white hair, eyes closed with a smile, both hands placed adorably against her chin, sitting, stylish and cute black parka, three small white round robots floating, a cyber world composed of black and green circuit-like patterns

【③ The effectiveness of quality-enhancing prompts leaves room for verification 🤔】

Do prompts like masterpiece or ultra detailed also work on V3? To verify this, I conducted one experiment 🔍️

Below is a comparison generated with a shared prompt and shared seed value, with and without quality-enhancing prompts ⚖️

(By the way, there are surprisingly few people who demonstrate the effects of quality-enhancing prompts through controlled experiments like this)

〈Example 3-1 Without quality-enhancing prompts〉
Anime CG, very cute high school girl, smiling and energetic, dynamic pose, peace sign, sunset street corner, lens flare, sparkling, watercolor style
↓

〈Example 3-2 With quality-enhancing prompts〉
Anime CG, very cute high school girl, smiling and energetic, dynamic pose, peace sign, sunset street corner, lens flare, sparkling, watercolor style, masterpiece, ultra detailed
↓

What do you think?

Looking at this, it seems like there isn't much of an effect. However, this is just one example, so further experiments are needed to verify the effects 🤔

【In closing】

Looking at other people's works, I think V3 can generate very flavorful illustrations when using the reference image function.

But right now, I'm in the middle of exploring ways to create good illustrations without that function...

If anyone else knows know-how for creating cute anime-style images with V3, let's share them~