Practical theory of prompt design in portrait generation AI | Sugiyama Nobutsugu

- Structural decomposition and complete prompts to ensure reproducibility -

table of contents

Introduction
Prompts are “blueprints” rather than “text”
1. ■Basic structure of portrait prompt
Three factors that reduce reproducibility
Design principles to improve reproducibility
Complete Practical Level Prompts
The essence of this prompt
application：How to use prompts correctly
summary
Detailed explanation of the above prompts
① Bias in learning data
② Accuracy difference in token decomposition
1. For English
2. For Japanese
③ English is the standard language for photography terms
■Case where Japanese is suitable
●Recommended structure
●Example
●① Completely in English + internal design in Japanese
●② “Intentional translation” from Japanese to English

Introduction

In portrait generation、There is a problem that many people face。
It is the instability that ``a different image appears every time even with the same prompt''。

The problem is not simply because the prompt is short.。
The essence is、Prompts are “unstructured”It is located in。

In this article、Breaking down the prompts in portrait generation into a “production process”、Design methods that ensure reproducibility and、Complete guide to practical prompts。

Prompts are “blueprints” rather than “text”

A typical prompt tends to look like this:。

beautiful woman
cinematic
high quality

But this、In terms of the production process, it is synonymous with "please take the photo in a good way."。
In other words、state of no control at allです。

What's important is、What is to do is to break down the prompt as below。

■Basic structure of portrait prompt

Subject design (Subject)
Styling
Pose/Composition / Composition）
Lighting
Environment
Camera settings (Camera)
Color/tone
Style
Emotion/Meaning (Mood)
Quality control

This structure is、It matches the production process of the live-action film.。
So what is a prompt?、

Text of shooting direction

です。

Three factors that reduce reproducibility

① Frequent use of abstract words

Words such as “beautiful” and “cinematic” can be interpreted too broadly.、The result changes every time。

② Undefined physical conditions

light position、camera distance、If the focal length etc. are not specified,、The composition is blurry。

③ Ignoring noise factors

Background, hair, facial expressions, etc.、If there are too many parts that the AI interprets freely, it will become unstable.。

Design principles to improve reproducibility

■1. Quantify as much as possible

Distance (1.5m)
Angle (45°)
Color (#code)

■2. Fix swaying elements

Expression (neutral)
sight (camera)
Hairstyle (length/parting)

■3. Prevent deviations with negative prompts

anime / cartoon exclude
Eliminate over-correction
background noise elimination

Complete Practical Level Prompts

Below is、Prompts designed to maximize reproducibility。

ultra-realistic portrait of a 26-year-old Japanese woman,
height 165cm, slim build, narrow shoulders, long neck,
oval face shape, small chin, straight nose bridge, slightly wide-set almond eyes,
dark brown iris (#3b2f2f), clear sclera, natural eyelashes,
light smooth skin with subtle pores, no blemishes, no freckles,

neutral calm expression, lips slightly closed, no smile,
eye contact directly to camera,

hair: medium-length (shoulder length), straight with slight inward curl at ends,
natural black (#1a1a1a), center part, slight volume at crown, no stray hair,

outfit: white silk blouse, matte texture, no patterns, slightly loose fit,
top button open, soft fabric folds, no accessories,

pose: seated upright on a chair, spine straight but relaxed,
hands resting gently on thighs, fingers naturally curved,
shoulders slightly angled (15 degrees to camera),

camera position: eye-level, 1.5 meters distance, centered framing,
framing: chest-up portrait, head near top margin,

lens: 85mm prime lens,
aperture: f/1.8,
depth of field: shallow, sharp focus on eyes, background fully blurred,

lighting setup:
single soft light at 45° camera left, slightly above eye level,
soft shadows on opposite side, subtle reflector fill (10%),

environment:
indoor studio, plain warm gray background (#d6d1cc), no objects,

color grading:
warm tone, low contrast, soft highlights, natural skin tones,

style:
editorial fashion photography, realistic, non-stylized,

negative prompt:
cartoon, anime, bad anatomy, extra fingers, blur, noise,
overexposed, harsh shadows, plastic skin, busy background,

--ar 2:3 --q 2 --style raw --seed 12345

The essence of this prompt

The value of this prompt is not "length"。

What's important is、

leaving no room for interpretation
fixed variables
Completely verbalizes the production process

That's the point。

application：How to use prompts correctly

This complete version is not a "finished form" but a "base"。

For example：

Change only the lighting
Just change the hairstyle
Change only the lens

By replacing things like、

You can design your desired variations.

It will look like this。

summary

Prompt design in portrait generation is、

rather than adding words
Control structure design

です。

and most importantly、

Which process should be left to AI?、Where should humans design?

That's the point of view。

AI is good at “generating”、“Design of intent” is not possible。
The person responsible for the design is、This is a prompt。

What you need is not technique、
Breaking down the production process、perspective to reconstructです。

Detailed explanation of the above prompts

This prompt is not just a "detailed description"、Control specifications designed to suppress fluctuations in generated results"is。

In the following、Each blockwhat to fix、How do we control the degree of freedom of AI?of、I will break it down and explain it from a practical perspective.。

■Understanding the overall structure (most important)

This prompt consists of three layers:：

① Shape definition (Geometry)

→ 人物の物理的特徴・構図

② Optical definition (Optics)

→ 光・レンズ・色

③ Constraints

→ AIの逸脱防止

👉 この3つを揃えることで、“再現性のある生成”が成立します

■① 被写体設計（Subject Design）

ultra-realistic portrait of a 26-year-old Japanese woman,
height 165cm, slim build, narrow shoulders, long neck,
oval face shape, small chin, straight nose bridge, slightly wide-set almond eyes

●Role

人物の「平均化」を防ぐ
骨格レベルでのブレを抑制

●Points

「26歳」→ 若すぎず老けすぎない中間固定
「narrow shoulders / long neck」→ シルエット制御
「face shape」→ 顔の輪郭を固定

👉AIは顔より“輪郭”の方がブレやすい

■② 目・肌の詳細定義（微細ディテール制御）

dark brown iris (#3b2f2f), clear sclera, natural eyelashes,
light smooth skin with subtle pores, no blemishes, no freckles

●Role

「不気味の谷」回避
Texture stabilization

●Points

Specify color code → Improve color reproducibility
sclera (pewter of eyes) → Prevents cloudiness
pores → Prevent excessive AI correction

👉“Skin is not specified” and made into plastic

■③ Facial expression/gaze (the most variable element)

neutral calm expression, lips slightly closed, no smile,
eye contact directly to camera

●Role

Fix emotional fluctuations

●Points

no smile → Eliminate subtle changes in facial expressions
eye contact → prevention of line of sight deviation

👉When your gaze shifts, you look like a “different person”

■④ Hair (control of noise generation source)

medium-length, straight, inward curl, center part, no stray hair

●Role

Suppressing the biggest cause of generation failure

●Points

Specify all length + shape + parting
Stray hair elimination → noise reduction

👉Hair is the most unstable part for AI

■⑤ Costume (light reflection control)

white silk blouse, matte texture, no patterns

●Role

Stabilize the behavior of light

●Points

silk × matte → prevent excessive reflection
no patterns → recognition noise reduction

👉Patterns cause AI to misrecognize

■⑥ Pose (skeleton of composition)

seated upright, hands on thighs, shoulders angled 15°

●Role

Prevention of collapse of human body structure

●Points

Hands designation → Measures against crooked fingers
Angle specification → Avoid unnaturalness that is too frontal

👉If you do not specify the move, there is a high probability that it will collapse.

■⑦ Camera position (fixed viewpoint)

eye-level, 1.5m distance, centered framing

●Role

Stabilizing the perspective

●Points

Specify distance → Prevent facial distortion
eye-level → natural impression

👉Distance not specified = wide-angle distortion occurs

■⑧ Lens/depth of field (the core of photography)

85mm, f1.8, shallow depth of field

●Role

Determining photo-likeness

●Points

85mm → portrait standard
f1.8 → background separation

👉This is the turning point between “CG feeling” and “photo feeling”

■⑨ Writing (most important)

single soft light at 45°, slightly above eye level

●Role

Creating a three-dimensional effect

●Points

45° → Optimum face shadow balance
From above → Natural light reproduction
fill 10% → Contrast fine adjustment

👉Not controlling light = everything collapses

■⑩ Environment (noise isolation)

plain warm gray background (#d6d1cc), no objects

●Role

背景の暴走防止

●Points

単色指定 → 認識安定
no objects → 不要要素排除

■⑪ カラーグレーディング

warm tone, low contrast, soft highlights

●Role

印象の統一

👉 撮影後の「現像工程」に相当

■⑫ スタイル指定

editorial fashion photography, realistic, non-stylized

●Role

出力の方向性固定

👉 “photorealistic”より具体性が高い

■⑬ Negative prompt (constraint)

cartoon, anime, bad anatomy, extra fingers...

●Role

Preventing AI from running out of control

●Points

Anatomy → Preventing shape loss
style → manga prevention

👉Often more important than positive

■⑭ Parameter

--ar 2:3 --q 2 --style raw --seed 12345

●Role

final control
ar → composition ratio
seed → core of reproducibility

👉Complete reproduction is impossible without a seed

■The essence of this prompt

This prompt is

It's not because it's detailed that it's good.
Stable because “variables are crushed”

The design is。

■Important understanding

This structure remains、

Planning → Concept → Photography → Development

This is consistent with the live-action production.。

■Summary

The essence of this prompt is：

eliminate ambiguous expressions
into physical conditions
Eliminate noise factors in advance

In other words、

Instead of leaving it to AI、Designed to control AI

です。

■Why English is more advantageous

① Bias in learning data

Many generative models、

Images with English captions
English-speaking datasets (e.g. LAION)

is being learned in。

👉 つまり、

概念とビジュアルの結びつきが英語で最適化されている

② Accuracy difference in token decomposition

AIは文章をそのまま理解しているのではなく、トークン（意味単位）に分解I am doing。

For English

“soft light”
“85mm lens”

👉 意味単位で安定して分解される

For Japanese

「柔らかい光」
「85mmレンズ」

👉 分解が不安定になりやすい
（文脈依存・曖昧性が高い）

③ English is the standard language for photography terms

For example：

aperture（絞り）
depth of field（被写界深度）
rim light
cinematic lighting

👉 これらは英語で学習されている前提語彙

■では日本語は使えないのか？

in conclusion：使えるが、用途を分けるべき

■Case where Japanese is suitable

① Concept/Emotion

「静かな朝の雰囲気」
「孤独感のある表情」

👉 抽象概念は日本語でも機能する

② Rough generation

Idea generation
雰囲気確認

👉 精度より方向性重視

■日本語が不利なケース

① Physical and technical designation (fatal)

光の角度
lens
距離

👉 ここは英語でないと崩れる

② When reproducibility is required

👉 同じ画像を出したいなら英語必須

■実務的な最適解（重要）

👉ハイブリッド運用

●Recommended structure

[英語：物理・構造]
+
[Japanese：コンセプト・感情]

●Example

ultra-realistic portrait, 85mm lens, soft lighting, shallow depth of field,静かな朝の空気感、少し内省的な表情、Calm atmosphere

👉 これが最もバランスが良い

■さらに精度を上げる方法

●① Completely in English + internal design in Japanese

設計 → 日本語で考える
出力 → 英語に変換

👉 実務ではこれが主流

●② “Intentional translation” from Japanese to English

単純翻訳はNG：

❌ "Soft light" → soft light
✔ “Diffused light from the window” → diffused window light

👉Important to translate into photographic terminology

■Common misconceptions

❌"No problem if it's a model that supports Japanese"
→Partially correct but not enough

reason：

I can understand the language
butEnglish is dominant when it comes to visual connections

■Conclusion

It can also be generated in Japanese.
but
English is superior in terms of accuracy, reproducibility, and controllability.

■Essence

The problem is not the language、

To what extent is the description “controllable”?

です。

However, in the current model、

Its control language is optimized for English

That is the reality。

*Use the prompt above、I tried generating images with Gemin and ChatGPT。(every time、(It may not be the same photo)

ChatGPT has created a face that is commonly seen in AI images.。For Gemini,、It looks natural, like a real person。But、Gemini ignores the request to shake your shoulders 15 degrees.。There are many places where Gemini does not listen to your instructions.。