AI画像プロンプトは長いほど良い？｜短文から長文まで同一Seedで比較する実験計画

この記事で分かること

• 既存の「要素を外す」アブレーションと対になる Forward Ablation（足していく）の設計
• Prompt長を文字数・token相当・意味ブロック数で測る記録方法
• Minimum Effective Prompt の定義と、必須／推奨／状況依存／過剰の分類枠
• いまは 検証準備中。採点結果・グラフ・採用率は画像取得後に更新する

要点

既存 Making（prompt-ablation-same-seed）は完成Promptから要素を外す。本稿はその逆で、最小構成から要素を足し、どこで品質・忠実度が飽和し、どこから競合で崩れるかを測る。

「長いほど良い」を断定しない。データが無い段階で結果を書かない。

なぜ検証するか

複数Seed再現性 では「詳細なプロンプトほど良い」ではないと既に示唆されている。prompt-ablation-same-seed の条件Gは品質タグ削減でも明確な劣化が無かった。

一方、読者の実務では「どこまで書けば十分か」が未整理のまま残る。除去実験だけでは「足す順序」と「過剰の境界」が見えない。

研究質問

主質問:

> AI画像生成では、Promptを長く・詳細にするほど画像品質は向上するのか？

副次（断定禁止。「今回の5 Seedでは」と限定予定）:

• Prompt長と指示忠実度
• Prompt長と構図安定性
• Prompt長とキャラクター一貫性
• Prompt長と不要要素発生率
• Prompt長とSeed間ばらつき
• Prompt長と採用率
• 指示同士の競合が始まる地点
• 品質向上の飽和点
• 性能が悪化し始める地点

仮説

実験条件

• 検証ID V11 / サービス TensorArtのみ / モデル WAI-illustrious-v17
• Sampler Euler a / Steps 20 / CFG 7 / 解像度 1280×768（16:9）
• Seed: 10101 / 20202 / 30303 / 40404 / 50505（最低5。可能なら後続で10へ拡張）
• Negative Prompt: V10 FULL と同一固定
• LoRA: Detail 系3種を共通基底として固定（条件差にしない）
• シーン: cafe-window-teacup（既存V09/V10と接続）
• 開示予定: AI生成（TensorArt）。水着・浴室・寝室・着崩れは未使用
• 採点: 複数Seed。1条件1枚で結論を出さない。5 Seedでは統計的有意差を主張しない

Prompt一覧（Level）

1girl, solo, japanese woman, adult woman, early twenties appearance, shoulder-length dark brown hair, brown eyes, sitting in cafe

1girl, solo, japanese woman, adult woman, early twenties appearance, shoulder-length dark brown hair, straight hair with soft ends, brown eyes, same eye color both eyes, natural makeup, calm facial features, opaque clothing, fully clothed, modest neckline, soft knit cardigan over blouse, sitting in cafe

1girl, solo, japanese woman, adult woman, early twenties appearance, shoulder-length dark brown hair, straight hair with soft ends, brown eyes, same eye color both eyes, natural makeup, calm facial features, opaque clothing, fully clothed, modest neckline, soft knit cardigan over blouse
BREAK
sitting at window table, both hands wrapped around teacup, head slightly tilted toward open book, relaxed upright posture, elbows near table
BREAK
landscape 16:9, upper body from side angle, camera at seated eye level, medium close distance, window as backdrop, table edge visible

1girl, solo, japanese woman, adult woman, early twenties appearance, shoulder-length dark brown hair, straight hair with soft ends, brown eyes, same eye color both eyes, natural makeup, calm facial features, opaque clothing, fully clothed, modest neckline, soft knit cardigan over blouse
BREAK
cozy cafe by the window, wooden table, open book, teacup, quiet afternoon interior
BREAK
sitting at window table, both hands wrapped around teacup, head slightly tilted toward open book, relaxed upright posture, elbows near table, feet planted under chair, weight centered over seated pelvis, palms wrapping cup, forearms on table edge
BREAK
landscape 16:9, upper body from side angle, camera at seated eye level, medium close distance, window as backdrop, table edge visible, shallow depth
BREAK
warm afternoon sunlight through window pane, soft golden rim light on cheek and teacup edge, soft golden glow on cheek and hands, midtones preserved

masterpiece, best quality, ultra detailed, highly detailed, 1girl, solo, japanese woman, adult woman, early twenties appearance, same eye color both eyes, brown eyes, shoulder-length dark brown hair, straight hair with soft ends, natural makeup, calm facial features
BREAK
cozy cafe by the window, wooden table, open book, teacup, quiet afternoon interior
BREAK
sitting at window table, both hands wrapped around teacup, head slightly tilted toward open book, relaxed upright posture, elbows near table, feet planted under chair, weight centered over seated pelvis, palms wrapping cup, forearms on table edge
BREAK
landscape 16:9, upper body from side angle, camera at seated eye level, medium close distance, window as backdrop, table edge visible, shallow depth
BREAK
warm afternoon sunlight through window pane, soft golden rim light on cheek and teacup edge
BREAK
content peaceful expression, gentle half-smile, soft relaxed eyes
BREAK
opaque clothing, fully clothed, modest neckline, soft knit cardigan over blouse, soft golden glow on cheek and hands, midtones preserved, exceptionally soft skin, soft tonal gradation, natural skin sheen
Narrative: Wrapped both hands around the warm teacup, she lifted her eyes from the page as warm afternoon light filtered through the cafe window.

masterpiece, best quality, ultra detailed, highly detailed, intricate details, extremely detailed eyes, perfect anatomy, award winning, cinematic masterpiece, 1girl, solo, japanese woman, adult woman, early twenties appearance, same eye color both eyes, brown eyes, emerald green contacts optional, shoulder-length dark brown hair, blonde highlights, straight hair with soft ends, wavy hair tips, natural makeup, heavy eyeliner, calm facial features, soft smile, serious expression, looking at viewer, looking at book, looking out window
BREAK
cozy cafe by the window, wooden table, marble table, open book, closed laptop, teacup, coffee cup, latte art, pastry plate, flower vase, quiet afternoon interior, busy cafe crowd in soft bokeh, rainy night city outside, sunny afternoon inside
BREAK
sitting at window table, both hands wrapped around teacup, one hand on cheek, hand holding book open, leaning forward, leaning back, crossed legs, feet planted under chair, weight centered over seated pelvis, palms wrapping cup, forearms on table edge, arms crossed
BREAK
landscape 16:9, portrait close-up, full body standing, upper body from side angle, camera at seated eye level, bird's eye view, medium close distance, window as backdrop, table edge visible, shallow depth, deep focus
BREAK
warm afternoon sunlight through window pane, cold blue moonlight, soft golden rim light on cheek and teacup edge, neon pink rim light, volumetric god rays, harsh flash
BREAK
content peaceful expression, gentle half-smile, soft relaxed eyes, dramatic tearful eyes, mysterious smirk
BREAK
opaque clothing, fully clothed, modest neckline, soft knit cardigan over blouse, leather jacket, school blazer, soft golden glow on cheek and hands, midtones preserved, exceptionally soft skin, soft tonal gradation, natural skin sheen, glossy plastic skin
Narrative: Wrapped both hands around the warm teacup, she lifted her eyes from the page as warm afternoon light filtered through the cafe window, while also preparing to leave for a midnight train and remembering a summer festival.

Prompt長の記録方法

各Levelで次を記録する（生成前にJSONへ記入可）:

• 文字数（空白含む / 除く）
• token相当数（使用UIの表示、または概算）
• 意味ブロック数（BREAK分割＋Narrative有無）

評価方法

各画像を0〜5で採点（詳細Rubricは prompt-quality-score）:

• Prompt adherence
• Character consistency
• Composition accuracy
• Pose accuracy
• Background accuracy
• Lighting accuracy
• Anatomy stability
• Artifact incidence（低いほど良い→記録時は発生率）
• Unexpected object incidence
• Seed variance（条件内）
• Adoption rate（overall≥4 かつ致命フラグなしを基本）

結果

未取得。 ここに平均点・採用率・グラフを書かない。

画像配置予定パス: /images/forward-ablation/20260811/L{n}-{seed}.webp

Prompt長と品質の関係 / 飽和点 / 逆効果

データ取得後に記入。現時点の記述は禁止。

Minimum Effective Prompt（枠のみ）

意味:

> 狙った画像を十分な確率で生成できる最小限のPrompt構成

分類枠（結果後にLevelを割当）:

実践ルール（データ後に確定。現時点は手順のみ）

1. まず Minimal でSeedを固定し、動作の成功を確認する

2. Character → Composition → Lighting の順で1ブロックずつ足す

3. 改善が頭打ちになったら Full へ飛ばず、戻って削る

4. Overloaded の症状（矛盾指示・色汚染・人数増）が出たら長さではなく競合を疑う

5. 最終採用は最高点ではなく採用率とWorstを見る

今回の限界

• 画像未生成のため仮説判定不可
• 1モデル・1シーン・5 Seedに限定予定
• 人間単独採点のバイアス
• token数はUI依存で厳密比較が難しい場合がある

次の検証課題

• Seedを10へ拡張
• Instruction Conflict（両手カップ）との接続（ instruction-conflict-both-hands ）
• LoRA数を増やしたときの飽和点変化（ lora-conflict-stack ）
• PQSカードのMaking一覧表示

関連Making

• prompt-ablation-same-seed
• アブレーション研究
• 複数Seed再現性
• instruction-conflict-both-hands
• prompt-quality-score

関連Knowledge

• minimum-effective-prompt
• アブレーション比較の始め方
• 複数Seed比較の読み方
• prompt-quality-score
• prompt-competition