研究＆習作 / Research ＆ Study (Pixiv Fanbox)

過去に作った３DCGをAIで２Dイラストへと変換し、ペイントソフトで加筆・修正。

The 3DCG I created in the past was converted into a 2D illustration using AI, and then added and modified using paint software.

１：この画像のようにキャラクターが一人で写っているシンプルな構図はAIとの相性がよいみたいだ。

1: A simple composition with a single character, as in this image, seems to work well with AI.

２：かなり苦戦した１枚。「体は後ろを向いていて、顔は振り向いている」描写はAIにとって超苦手分野らしい。

2: This is one image that I struggled with quite a bit. It seems that "the body is facing backward and the face is turning around" is a very difficult subject for AI to depict.

３：表情が素晴らしい。

3: Great facial expressions.

４：正面構図はAIの得意分野。いい感じのイラストができたと思う。

4: Frontal composition is AI's strong point. I think I got a good illustration.

傾向としてわかってきたのは…

The trend I found is as follows...

人数→ AIは人物１人の構成が得意。人数が増えるほどクオリティが怪しくなり、別人物の体が融合したりする。たぶんAIは人間と人間の境界を理解するのが苦手なんだと思う。逆にいうと、パッと見で判断できる人間の目（脳）がすごいのかも？

エロを作るならこの問題は絶対に克服する必要があるので、工夫して何とかしたい。

Number of people → AI is good at composition of one person. The more the number of people, the more the quality becomes doubtful, and the body of another person may be fused. Perhaps AI is not good at understanding the boundary between human and human. On the other hand, maybe the human eye (brain) that can judge with a quick glance is amazing?

In order to create erotic works, this problem absolutely needs to be overcome, so I would like to devise a way to solve it.

手→ AIは基本的に手の描写が苦手なので、手作業での修正が必須。時には手首から先を全て描き直すことも。でも時々完璧に描写するので、パターンを探っていきたい。

Hands → AI is basically not good at depicting hands, so manual corrections are necessary. Sometimes I have to redraw everything from the wrist to the finger end. But AI sometimes describes them perfectly, so I would like to explore patterns.

構図→ AIは正面構図が得意。「体は後ろを向いているけど、顔はこっちを向いてる」みたいな構図はAIが混乱するのか、クオリティが怪しくなる。

でも正面構図だけ作るわけにもいかないので、ここも工夫して何とかしたい。

Composition → AI is good at frontal composition. Complicated compositions such as "the body is facing backward, but the face is facing this way" confuse AI and tend to degrade the quality of the illustration.

However, it is not possible to create only frontal compositions, so I would like to devise a way to solve this problem as well.