生成AI画像を実写っぽくするには?ノイズ・手ブレ・動画風の4パターンを比較
同じ共通プロンプトに「写真のノイズ」「手ブレ」「動画の切り抜き」をそれぞれ単独で追加しました。少し寄った通学シーンで4枚を見比べ、どの指定が画像にどう現れたかを記録します。
変えたのは末尾の一節
基準画像は共通文だけ。残り3枚には、共通文の末尾へそれぞれ別の一節を追加しました。ノイズ・手ブレ・動画風を同時には指定していません。
| 作例 | 追加した指定 | 今回の見え方 |
|---|---|---|
| ① 基準 | なし | 輪郭が明瞭 |
| ② ノイズ | 暗部や路面の細かな粒子 | 変化は控えめ |
| ③ 手ブレ | 腕と画面端の局所的なブレ | ブレが視認しやすい |
| ④ 動画風 | スマホ動画の一コマ、圧縮感と露出の揺らぎ | 違いは微妙 |
4枚の生成結果
1. 指定なし
共通文のみ。撮影質感の追加指定はありません。
実際に入力したプロンプト全文
Use case: photorealistic-natural. Create a wide 16:9 image of a fictional Japanese woman age 22, clearly an adult university student, walking to university on a quiet ordinary residential street in Japan on an overcast morning. She has a cute natural face, long straight dark brown hair falling below her shoulders to mid-back, soft bangs, and a relaxed small smile while looking slightly away from the camera. She wears modest adult preppy fashion inspired by a school uniform: navy blazer, white collared shirt, small burgundy ribbon, dark pleated skirt, and a black shoulder bag. No school emblem. One main person in the foreground. Medium-close framing from mid-thigh to above the head, with her full face, shoulders, hands, bag and some of the street visible. Eye-level camera, subject slightly left of center; Japanese low-rise houses, utility poles, pavement and distant campus buildings behind her. Natural daylight, realistic skin texture and ordinary colors. Keep face and hands readable and anatomically correct. No text, watermark, logo, nudity or sexualized pose.2. 写真のノイズだけ
追加した一節:
Additional capture texture: visible fine high-ISO photographic noise and subtle sensor grain in shadow areas, on the dark jacket and on the road surface. Keep edges and the face sharp.
実際に入力したプロンプト全文
Use case: photorealistic-natural. Create a wide 16:9 image of a fictional Japanese woman age 22, clearly an adult university student, walking to university on a quiet ordinary residential street in Japan on an overcast morning. She has a cute natural face, long straight dark brown hair falling below her shoulders to mid-back, soft bangs, and a relaxed small smile while looking slightly away from the camera. She wears modest adult preppy fashion inspired by a school uniform: navy blazer, white collared shirt, small burgundy ribbon, dark pleated skirt, and a black shoulder bag. No school emblem. One main person in the foreground. Medium-close framing from mid-thigh to above the head, with her full face, shoulders, hands, bag and some of the street visible. Eye-level camera, subject slightly left of center; Japanese low-rise houses, utility poles, pavement and distant campus buildings behind her. Natural daylight, realistic skin texture and ordinary colors. Keep face and hands readable and anatomically correct. No text, watermark, logo, nudity or sexualized pose. Additional capture texture: visible fine high-ISO photographic noise and subtle sensor grain in shadow areas, on the dark jacket and on the road surface. Keep edges and the face sharp.3. 手ブレだけ
追加した一節:
Additional camera movement: slight handheld camera shake with localized directional motion blur on the swinging hand and frame edges, while keeping her face readable.
実際に入力したプロンプト全文
Use case: photorealistic-natural. Create a wide 16:9 image of a fictional Japanese woman age 22, clearly an adult university student, walking to university on a quiet ordinary residential street in Japan on an overcast morning. She has a cute natural face, long straight dark brown hair falling below her shoulders to mid-back, soft bangs, and a relaxed small smile while looking slightly away from the camera. She wears modest adult preppy fashion inspired by a school uniform: navy blazer, white collared shirt, small burgundy ribbon, dark pleated skirt, and a black shoulder bag. No school emblem. One main person in the foreground. Medium-close framing from mid-thigh to above the head, with her full face, shoulders, hands, bag and some of the street visible. Eye-level camera, subject slightly left of center; Japanese low-rise houses, utility poles, pavement and distant campus buildings behind her. Natural daylight, realistic skin texture and ordinary colors. Keep face and hands readable and anatomically correct. No text, watermark, logo, nudity or sexualized pose. Additional camera movement: slight handheld camera shake with localized directional motion blur on the swinging hand and frame edges, while keeping her face readable.4. 動画の切り抜きだけ
追加した一節:
Additional capture context: a single frame extracted from an ordinary smartphone video, with subtle phone-video compression texture and slightly uneven automatic exposure. Keep her face readable.
実際に入力したプロンプト全文
Use case: photorealistic-natural. Create a wide 16:9 image of a fictional Japanese woman age 22, clearly an adult university student, walking to university on a quiet ordinary residential street in Japan on an overcast morning. She has a cute natural face, long straight dark brown hair falling below her shoulders to mid-back, soft bangs, and a relaxed small smile while looking slightly away from the camera. She wears modest adult preppy fashion inspired by a school uniform: navy blazer, white collared shirt, small burgundy ribbon, dark pleated skirt, and a black shoulder bag. No school emblem. One main person in the foreground. Medium-close framing from mid-thigh to above the head, with her full face, shoulders, hands, bag and some of the street visible. Eye-level camera, subject slightly left of center; Japanese low-rise houses, utility poles, pavement and distant campus buildings behind her. Natural daylight, realistic skin texture and ordinary colors. Keep face and hands readable and anatomically correct. No text, watermark, logo, nudity or sexualized pose. Additional capture context: a single frame extracted from an ordinary smartphone video, with subtle phone-video compression texture and slightly uneven automatic exposure. Keep her face readable.この4枚から分かったこと
今回の出力では、③の手ブレ指定が画面左端と腕に最も分かりやすく反映されました。②のノイズ指定は入れたものの、縮小表示では差が見えにくい程度です。④の動画風も、静止画から動画の一コマと判別できるほどの強い特徴は出ませんでした。効かなかったように見える指定も結果として残しています。
「リアルさ」は好みも含む評価です。粒子やブレを増やせば必ず良くなるわけではなく、顔や服まで崩れると読みにくくなります。今回の手ブレ版では、顔を読める状態に保つ指定も同時に入れました。
比較の限界
4枚の入力文は共通ですが、seedも参照画像も固定できない独立生成です。顔立ち、人物の位置、街並みも変わっているため、見た目の差を追加語句だけの因果効果とは断定できません。各条件1枚なので再現性も測っていません。厳密に調べるには、同じ基準画像やseedで複数回ずつ試す必要があります。
画像はCodex組み込みの画像生成ツールによるAI生成です。ChatGPTのWeb画面やローカルモデルの結果ではありません。正確な基盤モデル名とseedは取得できていません。人物と背景は架空で、実在の学生や場所を撮影した写真ではありません。
肌・髪・色・服の短い指定は実写感を増す一言プロンプトの検証で試しました。全身の通学風景へ構図を変えた作例は、生成AI画像を実写風に見せるコツにまとめています。





