ブレ写真の生成にあえてチャレンジした話

文字通り目覚ましい成長というのはこういうことかと思わされる生成AI。私は興味はそれなりにあるとはいえ、「Adobe Firefly」は登場時からちょこちょこ試している程度です。商用可能になってからはチャンスが有れば仕事にも取り入れてみようと思い、音楽ライブのイメージ(先方のイメージとしては無料のストックフォト)の使用を求められた際に、試しに生成したら指の数があやふやなり、実用を断念したこともありました。そんな中でも最近では何度か実用にこぎ着けたものもあります。しかし、今回はそういったわかりやすく有用なテーマではなく、「うまくいかなさ」を記事にしています。なお、この実験は2026年の2月頃に行いました。

まずはGeminiによるブレ・ボケ生成のアドバイス

最近はスマホのカメラの性能も目覚ましく、写真の失敗が少なくなりました。特に手ブレ補正が進歩したのかブレブレの写真を目にする機会が減ったと思います。そんなカットこそあえて生成してみたい!と思い、以下の質問をGeminiに投げかけるとこんなアドバイスをくれました。

DO NOT USE NANO BANANA アドビのfireflyでシャッタースピードが遅くてブレているフィルムタッチの写真を生成したいので、そのためのプロンプトを英語で教えて。

Adobe Fireflyで「スローシャッターによる躍動感」と「フィルムの質感」を両立させるには、技術的なキーワードを組み合わせるのがコツです。
Firefly(特に最新モデル)は、具体的なカメラ設定の指示をよく理解してくれます。以下のプロンプトを試してみてください。

1. おすすめの英語プロンプト
そのままコピー&ペーストして使える3つのパターンを用意しました。
パターンA:都会的でエモーショナルな雰囲気
Prompt: A cinematic film still of a person walking through a busy Tokyo street at night, slow shutter speed, motion blur, intentional camera movement, 35mm film grain, muted vintage colors, nostalgic atmosphere, high contrast.
パターンB:自然体でノスタルジックな日常
Prompt: A candid film photograph of a woman dancing in a room, heavy motion blur, shutter drag effect, soft light leaking, Kodak Portra 400 aesthetic, grainy texture, blurry but artistic, 1990s aesthetic.
パターンC:抽象的でドリーミーな表現
Prompt: An abstract long exposure shot of city lights, intentional blur, trailing lights, vintage analog film style, light leaks, heavy grain, dreamy and hazy, retro film look.

2. 理想の画像に近づけるためのキーワード集
Fireflyのプロンプト欄に追加したり、調整したりする際に役立つ単語リストです。
カテゴリ
効果的なキーワード
ブレの表現
Slow shutter speed, Motion blur, Long exposure, Shutter drag
フィルム感
35mm film, Film grain, Analog photography, Vintage aesthetic
特定の色調
Kodak Portra 400 (温かみ), Fujifilm Superia (青・緑系), Muted colors
光の効果
Light leaks (光漏れ), Hazy (霞んだ), Soft glow

「DO NOT USE NANO BANANA」は果たして有効なのかわかりませんが、プロンプトを教えてほしいのにすぐ成果をだしたがるGeminiが画像を生成し始めてしまうからです。「英語で教えて」というのは以前Geminiに相談した際、Fireflyへの指示は英語のほうが効果的だから(大意)と回答を得たからです。そもそもGeminiにプロンプトを書かせる理由は何度もFirefly上で生成して直しても、効果的な調整ができずにクレジットばかり消費してしまうので、ある程度ディレクションをしながらプロンプト自体を詰めていくほうが、目標に近づきやすいという気付きを得たからです。それにしても自分の語彙には無いキーワードがたくさん出てきて勉強になります。

…なんて合理的なプロセスっぽく説明しておきながら、やってみると色々壁打ちを繰り返して冗長になってしまうのが現実です。そのような理由で生成結果を全て扱っていたら書く方も読む方もダルいので、このあとは実験内容をかいつまんでをご紹介します。まずはプロセスの序盤でGeminiがダメと言ってるのに生成してきたブレ画像と改めて出してもらったプロンプトの一部を御覧ください。

演出的なスローシャッターというより、スナップ写真でブレてしまったようなイメージで再考してください

Adobe Fireflyで生成した、意図的な手ブレとピンボケを伴う人物写真

DO NOT USE NANO BANANA 

失敗したスナップ写真風プロンプト
パターンA:歩きながら撮ってブレた日常(最もおすすめ)
Prompt: A blurry snapshot taken while walking, accidental camera shake, out of focus, subject moving too fast, grainy 35mm film, low quality disposable camera aesthetic, light leak, candid moment, authentic messy street photography.
パターンB:暗い場所でシャッタースピードが足りなかった風
Prompt: A dark indoor party scene, underexposed and blurry, shutter lag, heavy motion blur from hand-held camera, messy composition, vintage film grain, harsh flash but blurry edges, nostalgic and raw.
パターンC:ピントが合わず流れてしまった人物
Prompt: A candid photo of a friend laughing, failed focus, motion smear, shaky hands effect, messy 90s aesthetic, low-res film scan, accidental movement, authentic lo-fi photography.

なんか思っていたのと違うなあと思いつつ、自分の中での効果的なブレ写真として(文字通りの)中2のころの私が出会ったCDのジャケットを思い出しました。

より具体的なイメージを元に画像生成してみる

まずはイメージの元である現物は確認せず、ぼんやりとした記憶でこのイメージを生成してみることにしました。はじめに上記のプロンプトをベースに自分で日本語で書いたイメージを追加してFirefly(ソースはGemini 2.5 (w/ Nano Banana←結局Nano Banana使う)で生成します。

Firefly上で使用したプロンプト

昼間の路地裏でロックミュージシャンの日本人3人が並んでいる。楽器は持っていない。真ん中は金髪、左は黒髪長髪、右はパーマをかけている。A dark indoor party scene, underexposed and blurry, shutter lag, heavy motion blur from hand-held camera, messy composition, vintage film grain, harsh flash but blurry edges, nostalgic and raw.

Adobe Fireflyで生成した、意図的な手ブレとピンボケを伴う人物写真

ただのアー写っぽいし、うっかり生成してもらったテキストに「A dark indoor party scene」とか入っているのをそのまま採用しちゃった自分が悪いんだけど、イメージしていたシチュエーションは屋外でした。コピペで英文使って作るとちゃんと読ま(読め)ないで進めちゃう悪い例です。その辺も改めつつGeminiに再生成してもらいます。

パターンBに近いイメージで「昼間の路地裏でロックミュージシャンの日本人3人が並んでいる。楽器は持っていない。真ん中は金髪、左は黒髪長髪、右はパーマをかけている。」上記内容を含んだNanoBananaは使用しないでプロンプトだけ作ってください。

Adobe Fireflyで生成した、意図的な手ブレとピンボケを伴う人物写真

DO NOT USE NANO BANANA 

結局生成しちゃううっかりさんのGemini。生成結果はみんな棒立ちなのはイメージと違うのでついでに軌道修正します。

3人を直立不動ではなく、路地裏を歩いている様子を盛り込みたい。

ブラッシュアップした英語プロンプト
Prompt: A blurry candid snapshot of three Japanese rock musicians walking toward the camera in a narrow daytime back alley, no musical instruments. The middle man has blonde hair, the man on the left has long black hair, and the man on the right has permed hair. Walking while shooting, accidental camera shake, heavy motion smear on moving bodies, failed focus, 1990s disposable camera aesthetic, grainy 35mm film scan, messy composition, authentic lo-fi street photography, underexposed shadows, low-res scan, unintended motion blur.

追加・修正したポイント
Walking toward the camera: 「カメラに向かって歩いている」と指定することで、前方への動きを作り出します。
Walking while shooting: 「撮る側も歩きながら撮った」というニュアンスを加え、手ブレ(カメラ側のブレ)と被写体ブレを複合させます。
Heavy motion smear on moving bodies: 歩いている体のラインが「ドロっと流れる」ような、シャッタースピードが遅い時の独特な質感を指定しています。
Underexposed shadows: 路地裏の昼間特有の、明暗差が激しくて暗部が潰れた「素人写真っぽさ」を強調しました。

こうしてちゃんと自分で生成したプロンプトのポイントも紹介するのが親切ではあります。この英語プロンプトで生成したのがこれ↓です。

Adobe Fireflyで生成した、意図的な手ブレとピンボケを伴う人物写真

だいぶ進んだ気がするのですが、背景が日本の開店前の飲み屋街みたいなイメージになったことで、一旦自分でプロンプトに手を加えて生成を試みました。もう元ネタ要素を入れ始めています。

Firefly上で使用したプロンプト

Motion blur. A blurry candid snapshot of three Japanese rock musicians looks like Sparks GO GO walking toward the camera in a narrow daytime back alley, no musical instruments. The middle man has blonde hair, the man on the left has long black hair, and the man on the right has permed hair. Walking while shooting, accidental camera shake, heavy motion smear on moving bodies, failed focus, 1990s disposable camera aesthetic, grainy 35mm film scan, messy composition, authentic lo-fi street photography, underexposed shadows, low-res scan, unintended motion blur.

Adobe Fireflyで生成した、意図的な手ブレとピンボケを伴う人物写真

急にキャラ変…。でもフィルムっぽさとかブレみたいなものを表現しようとし始めました。筆者も業を煮やしてプロンプトにも名前を書いてしまっているのでそろそろ元ネタを披露しようと思います。

…そもそも並びが間違ってましたね。懐かしのN-2Bみたいなの着てる。この動きのあるイメージを作りたいのでもうちょっと頑張ってみます。

90年代のSPARKS GO GOのジャケットみたいな写真にしたかったので以下のようにしてみたけど、うまくいかなかった。アドバイスありますか?「Motion blur. …」(以下先程のプロンプトと同じ)

SPARKS GO GOの90年代の空気感、特にアルバム『JUNKFOOD JUNCTION』や『CIRCUS』あたりの、少しザラついていて、気取らない「バンドの日常の切り取り」のような質感を目指されているのだと理解しました。
Fireflyでうまくいかなかった原因として、**「AIが『Musicians』という言葉に引っ張られて、アー写(宣材写真)のように綺麗に整えようとしすぎている」**可能性が高いです。
以下の3つのポイントでプロンプトを修正してみてください。

1. 修正版プロンプト案(テキストのみ)
AIに「これは宣材写真ではなく、スナップの失敗作だ」と強く認識させるための構成です。
Prompt: A low-quality, blurry snapshot of three Japanese men walking in a narrow daytime back alley. They have a 1990s rock band aesthetic. The middle man has blonde hair, the man on the left has long black hair, and the man on the right has permed hair. Extreme camera shake, heavy subject motion blur, out of focus. Looks like a failed 35mm film scan from a cheap point-and-shoot camera. Messy, raw, unpolished, grainy. Overexposed sky, deep shadows, authentic nostalgic 90s street photography. No musical instruments.

2. なぜうまくいかなかったか?(アドバイス)
「Sparks GO GO」という固有名詞: AI(特にFirefly)は日本の特定のバンド名を詳細に把握していないことが多く、逆に「有名人=綺麗に撮るべき」というバイアスがかかって、ブレを抑制してしまうことがあります。名前を出すより、**「1990s rock band aesthetic」「Grunge style」**と指定した方が、当時の質感が伝わりやすいです。
「Musicians」より「Men」: AIにとってMusiciansは「カッコよく撮る対象」です。あえて 「Three Japanese men」 と呼び、補足としてバンド風のスタイルを指定することで、より「不意に撮られた感」が出ます。
「Extreme(極端な)」の追加: Fireflyは安全性を重視するため、中途半端な指定だと「少しソフトフォーカスな綺麗な写真」にまとめてしまいます。「Extreme camera shake(極端な手ブレ)」 などの強い形容詞が必要です。

3. Fireflyの設定(サイドパネル)での工夫
プロンプト以外に、操作画面で以下を試してみてください。
コンテンツタイプ: 必ず「写真」に。
ビジュアルの強さ: あえて下げ気味に設定すると、AIの「作り込み」が減り、ラフな質感になります。
参照画像を使用しない: もし綺麗な写真を参照している場合は外してください。

アルバム名の指定はしてないのに特定しはじめました(1つ目のは存在しないというか惜しい名前の曲はある)。すごい。うまくいかないことを伝えると、何に原因があるのか一応推測してくれるのえらい。さあ、だいぶこちらの意図を理解し始めている感じはします。さてこれでどうでしょう?

Adobe Fireflyで生成した、意図的な手ブレとピンボケを伴う人物写真

後ろを向いてしまった…。このあたりで修正指示を出し始めてからも、Geminiは度々画像を生成しようとするので、続きはChatGPTにお願いしてみることにしました。

相談相手をChatGPTに変えて再開

まずはこれまでの経緯と最後のプロンプトを添えてご相談です。

GeminiといっしょにAdobe Fireflyのための英語プロンプトをブラッシュアップしてきたのだけど、すぐその場でNanoBananaで画像生成しようとするので、ChatGPTに来ました。以下のプロンプトはかなり完成度が上がってきたのだけど、背景の路地裏は日本ではなくロンドンっぽくしたい。「A blurry, candid eye-level snapshot of three Japanese men walking towards the camera in a narrow daytime back alley. They are facing forward, their faces visible despite the heavy blur. The middle man has blonde hair, the man on the left has long black hair, and the man on the right has long spiky hair. Accidental camera shake, extreme motion blur, out of focus. 1990s Japanese rock band aesthetic. Low-quality disposable camera look, grainy 35mm film, messy and unpolished, unintended motion smear, nostalgic lo-fi street photography. No musical instruments.」

内容としてはある程度煮詰まってきたことと、ChatGPTはノリがいいというか調子がいいことで内容を紹介するのが正直しんどいので、以下プロンプトと画像をいくつか紹介します。

Firefly上で使用したプロンプト

A badly framed, accidental eye-level snapshot of three ethnically Japanese men from Japan walking towards the camera in a narrow London back alley during daytime. The image is unintentionally tilted, off-balance, and poorly composed as if shot by mistake. Their distinctly Japanese facial structure and East Asian features are slightly visible despite the heavy motion blur. The middle man has dyed blonde hair, the man on the left has long straight black hair, and the man on the right has long spiky hair. Slim Japanese body proportions.

Weathered red brick walls and damp pavement under flat, overcast northern European daylight. Quiet residential alley with no visible landmarks. Natural environment, not staged.

Severe camera shake, extreme motion blur, completely out of focus, partial motion smear across faces. Cropped heads, awkward framing, uneven horizon. Cheap disposable camera aesthetic, harsh direct flash, strong grain, color distortion, slight greenish tint, underexposed shadows, low dynamic range. Messy

Adobe Fireflyで生成した、意図的な手ブレとピンボケを伴う人物写真

プロンプトが長すぎて切れてしまったことが功を奏したのか、イメージの外側の結果が出ました。これはこれで嬉しい失敗です。3兄弟?

実質的なゴールが生成される

前のプロンプトの文字数を調整してもらって生成したのが以下の画像です。結果としてここが今回の実験のピークで実質的なゴールになりました。

Firefly上で使用したプロンプト

A badly framed, accidental tilted snapshot of three ethnically Japanese men from Japan walking toward the camera in a narrow London brick alley in daytime. Uneven horizon, awkward composition, partially cropped heads.

They have clearly distinct Japanese faces and different features:
– the middle man has a sharp jawline and dyed blonde hair
– the left man has a longer face with straight black hair
– the right man has a rounder face with long spiky hair
Different expressions and slightly different heights and builds.

Overcast northern European light, damp pavement, worn red brick walls. No landmarks.

Strong camera shake, extreme motion blur, heavy grain, harsh flash, slight color shift, low-quality disposable camera look, unintentional motion smear, raw candid street photo. No musical instruments.

Adobe Fireflyで生成した、意図的な手ブレとピンボケを伴う人物写真

ゴールの先の失敗写真探求

以下に掲載するのは粘って煽るChatGPTとの悪あがきの結晶です。こんな感じのもできるんだな程度にご覧いただければ幸いです。

Firefly上で使用したプロンプト

An accidental, badly timed snapshot of three ethnically Japanese men from Japan in a narrow London brick alley during daytime. The camera is tilted and off-balance, as if taken by mistake while moving. One head is partially cut off, another face blurred beyond recognition. They have clearly different Japanese facial structures: one with a sharp jaw and dyed blonde hair, one with a longer narrow face and straight black hair, one with a rounder face and long spiky hair. Different heights and builds.

Flat overcast light, damp pavement, worn red brick walls. No landmarks, no staging.

Severe motion blur, heavy grain, harsh direct flash, partial overexposure, color shift, light leak, finger slightly covering the lens. Completely out of focus, messy, unintended disposable camera photo. No musical instruments.

Adobe Fireflyで生成した、意図的な手ブレとピンボケを伴う人物写真

調子に乗ってどんどん壊していく流れになって生成された↑の画像ですが、当時の筆者が大切なツッコミを入れていました。

めっちゃ面白いけど、元々のセンター金髪、左ロン毛、右スパイキーの指示は消えちゃってたことに生成してから気づきました。「light leak / finger」のセルフィーっぽさは要らないので指は無くても良い。みんな
がセルフィーをしていない90年代に撮られた写真っぽさを出したい。

文面が完全にチャッピーに引っ張られてますね。その後も調整し、ゴールに近づいては遠ざかるの繰り返しになり、最終的には完全に酒か何かが入っていそうな人々が生成されてしまう中タイムアウト→終了に向かいます。

Firefly上で使用したプロンプト

A badly mistimed 1990s snapshot taken by a friend of three ethnically Japanese men from Japan in a narrow London brick alley during daytime. The camera tilts sharply to one side as the shutter is pressed too late. The frame is chaotic and poorly composed.

The middle man (dyed blonde hair) is mid-step and heavily motion blurred. The man on the left (long straight black hair) is partially cut off by the edge of the frame. The man on the right (long spiky hair) turns slightly away, his face smeared by motion. They are clearly different Japanese individuals with distinct face shapes, but none of them are cleanly captured.

Flat grey overcast light, damp pavement, worn red brick walls fading into blur. No staging, no posing, unaware of the camera.

Severe motion blur, uneven horizon, accidental flash misfire, partial overexposure, underexposed corners, heavy grain, color fading, low dynamic range. Looks like a failed disposable film photo from the mid-1990s. Messy, unintended, almost ruined image. No musical inst

Adobe Fireflyで生成した、意図的な手ブレとピンボケを伴う人物写真

ごきげんですね。楽しそうで何よりです。

Firefly上で使用したプロンプト

Badly mistimed mid-1990s in-camera exposure of three ethnically Japanese men in a narrow London brick alley. Direct capture, not a print, no borders.

The camera jerks as the shutter fires too late, horizon severely tilted. Chaotic framing.

Middle: dyed blonde hair, heavily motion blurred.
Left: long straight black hair, partly cut off.
Right: long spiky hair, face smeared by motion.
One nearly walking out of frame. Clearly different individuals, none in focus.

Flat overcast light, damp pavement, red brick walls fading into blur. No staging.

Extreme motion blur, near focus failure, heavy grain, uneven exposure, harsh flash. Almost ruined image. No musical instruments.

Adobe Fireflyで生成した、意図的な手ブレとピンボケを伴う人物写真

最後の画像はとても良いのですが、典型的な東アジア系の見た目ではなくなってしまいました。指示を整理していくうちにロケーションに引っ張られてしまうのでしょうか。

最後に気づいたこと

素晴らしい結果が出たわけではないですが、フィルム時代のテイストみたいなものも学習されているようで、ブレ・ボケを表現することもそれなりにできることはわかりました。そしてシチュエーションに人物が引っ張られてしまうことがあるので、次にトライするとしたら、人物と背景をそれぞれ別に生成して、最終的に見せたいテイストに落とし込むのが合理的なプロセスなのかもしれないと書きながら気付きました。またその気になったらやってみます。最後までお付き合いいただきありがとうございました!