미드저니 프롬프팅 실험 - 배치, 스타일, 파라미터
Midjourney
(미드저니 실험 내용은 영어 공부 겸 미친척하고 미디엄에 올려보려고 합니다 ㅎ...)
요약
1. 특정 단어의 경우, 긍/부정의 뉘앙스가 강하게 박혀있어 이를 중화시키는 표현을 추가하더라도 영향이 크지 않을 수 있다.
- 예를 들어, 'surprised'는 부정적인 뉘앙스를 가지고 있어, 이를 약화시킬 positively를 사용하더라도, 여전히 부정적인 얼굴의 표현이 지속된다.
2. 이미지가 이미 만들어진 것처럼 객체화하여 배치를 요구하는 프롬프트는 모델이 이해하기 어려울 수 있으며, 표현하려는 대상의 특성으로 구체화시켜 표현해야 의도한 이미지가 생성될 수 있다.
- '남자는 이미지 좌측에 위치해있다.'보다 '남자는 우측 상단을 바라보고 있다.'라는 표현을 해야 의도에 가까운 이미지가 생성된다.
3. 특징 레퍼런스 파라미터(--cref)와 스타일라이즈(--stylize) 파라미터를 동시에 사용할 경우 서로 간 충돌적인 영향을 미칠 수 있는 것으로 보인다.
- 이미지의 고급화를 수행하는 스타일라이즈 파라미터를 높게 잡을 경우, 사진과 비슷한 퀄리티로 이미지를 생성하게 되는데, 이는 레퍼런스로 삼은 이미지의 영향을 어느 정도 약화시키는 것으로 추측된다.
Summary
The adjective ‘surprised’ has strong negative nuances on the model in spite of adding positive adverb ‘positively’ in front of it.
The prompt which regards the image as an object seldom work, so you need to make subject-specific prompt for effective composition and arrangement.
It seems that the model has a conflict between — cref and — stylize, in the point that it produced an image looking like a photography, so to be sure of its medium and style, you might have to declare it on the prompt explicitly along with — cref
What I intended to create:
- Referencing Andy-Warhol-related to image(https://s.mj.run/Bnqi1o-yAc0) as style medium such as color and theme
- A man laughing and smiling happily with his eyes and mouth wide open
1st prompt:
pink and red color, the face of a man who is positively surprised at something, smiling, the face is located at the left on the image, his gaze is from left to right, wide open eyes and mouth — cref https://s.mj.run/Bnqi1o-yAc0 — ar 16:9 — stylize 700
Result: Creepy
Takeaway
I translated the expression of ‘happily laughing’ into ‘positively surprised’, but it seems that the model accepted the meaning in a negative way → I speculate that the adjective ‘surprised’ has strong negative nuances on the model
The composition and arrangement of a figure is hard to set, as [the face is located at the left on the image, his gaze is from left to right] is not properly reflected on the generated image
2nd Prompt
light pink and red color, happy, smiling, the man is located at the left side, open eyes and mouth — cref https://s.mj.run/Bnqi1o-yAc0 — ar 16:9 — stylize 700
Change on color: ‘light’ pink and red color
Removal of ‘positively surprised’, ‘gaze’ phrase
Result: Creepy again
Takeaway
Though it has character reference( — cref) parameter, it seems that the model has a conflict between — cref and — stylize, in the point that it produced an image looking like a photography
To be sure of its medium and style, I might have to declare it on the prompt explicitly along with — cref
3rd prompt
a man happily lauging, pop art style of Andy Warhol, light pink and red color, the man is located at the left side — cref https://s.mj.run/Bnqi1o-yAc0 — ar 16:9 — stylize 700
Addition of ‘Andy Warhol style’ prompt
Result: Not bad, but why are they making their head upward…?
Takeaway
Arrangement phrase isn’t taken into account — [the man is located at the left side]
4th Prompt
pop art style of Andy Warhol, a man happily laughing, the man is staring right upward, light pink and red color, the man is located at the left side — cref https://s.mj.run/Bnqi1o-yAc0 — ar 16:9 — stylize 700 -
Changes the position of ‘Pop Art Style’ prompt to the front
Changes of composition phrase into subject-specific expression
Result: Satisfactory
Takeaway
Subject-specific phrase for arrangement and composition works
