ChatGPT画像ジェネレーターを100時間かけてマスターし、学んだことのすべて

result = client.images.edit(
    model="gpt-image-1.5",
    input_fidelity="high",  # アイデンティティ保持のための秘密の成分
    quality="high",
    image=[open("portrait.png", "rb")],
    prompt="Change the background to a sunset beach while preserving the person's exact appearance"
)

この組み合わせにより、要求された変更を適用しながら、元の被写体の保存を最大限に確保します。

Goal/Output:
- [Type of image: ad, UI mockup, infographic, photo, illustration]
- [Intended use and audience]

Scene:
- [Background/environment description]
- [Main subject with specific details]
- [Action or relationship between elements]

Style:
- [Medium: photograph, watercolor, 3D render, vector illustration]
- [Key textures: matte, glossy, grainy, smooth, organic]
- [Quality descriptors: realistic imperfections, stylized, minimalist]

Composition/Layout:
- [Camera position: close-up, wide shot, aerial view, eye-level]
- [Lighting: golden hour, studio strobes, overcast, dramatic shadows]
- [Element placement: centered, rule of thirds, negative space, margins]

Text (if any):
- "Exact text in quotes"
- [Font style, size, color, position]
- [Specify: render only once, no duplicates]

Constraints:
- Change ONLY: [specific element if editing]
- Preserve exactly: [elements that must stay unchanged]
- Negative: no watermark, no extra text, no logos, no [unwanted elements]

このフレームワークは、モデルが行わなければならないすべての視覚的決定に対して明確なコンテキストを提供します。

❌ BEFORE（曖昧）:
"A beautiful portrait of an old fisherman, very detailed, high quality, realistic"

✅ AFTER（写真撮影のマインドセット）:
"Candid documentary photograph of an elderly fisherman on a weathered wooden boat.
Weathered face with visible wrinkles, sun spots, and pores. Deep-set kind eyes.
Gray stubble. Faded traditional anchor tattoo on forearm. Salt-stained navy wool
sweater, worn cap.

Early morning coastal light, soft fog diffusing the sun. Medium close-up at eye
level, 50mm lens, f/2.8, shallow depth of field. Shot like 35mm film with subtle
grain, natural color balance.

Documentary style — honest, unretouched, capturing a real moment. No glamorization."

写真撮影のマインドセットは、曖昧な希望を、モデルが深く理解できる正確な視覚的仕様に変換します。

Create a photorealistic candid photograph of an elderly sailor standing on a small fishing boat.

Subject: Weathered face with visible wrinkles, sun spots, and pores. Deep-set kind
eyes with crow's feet. Gray stubble, a few days unshaven. Faded traditional anchor
tattoo on forearm. Salt-stained navy wool sweater, worn and pilled. Creased cap
with faded insignia.

Setting: Early morning on the water, soft coastal fog diffusing the light. Aged
wooden boat deck with peeling paint, fishing nets in background, coiled rope.

Technical: Shot like 35mm film photography, medium close-up at eye level, 50mm
lens, shallow depth of field with boat blurred behind him. Subtle film grain,
natural color balance without heavy grading.

The image should feel like a real moment captured by a photojournalist — honest,
unposed, with real skin texture, worn materials, and everyday imperfection. No
glamorization, no heavy retouching, no artificial perfection.

風化した肌、使い古した素材、剥がれたペンキなど、不完全さを明示的に要求していることに注目してください。現実にはテクスチャがあります。

Create a detailed infographic explaining how a coffee machine works.

Structure:
- Title at top: "The Journey of Your Morning Coffee"
- Vertical flow diagram showing: bean hopper → grinder → portafilter →
  grouphead → water heating → extraction → cup
- Each step has an icon and 1-2 sentence explanation
- Warm color palette (browns, creams, copper accents)
- Clean, modern design with plenty of white space
- Subtle coffee stain texture in background corners

Style: Professional print-quality infographic, vector-style icons, clear
hierarchy, readable at A4 size.

Typography: Clean sans-serif headings, readable body text, clear visual
hierarchy between title, section headers, and explanatory text.

No watermarks. No stock photo elements. Original illustration only.

高密度のテキストと複雑なレイアウトの場合は、常にquality="high"を使用して、テキストが鮮明で読みやすい状態を維持するようにしてください。

Create an original logo for "Field & Flour" — a local artisan bakery.

Brand personality: Warm, authentic, handcrafted, timeless. Not trendy or corporate.

Design requirements:
- Clean vector-style shapes with strong silhouette
- Balanced negative space
- Must read clearly from 16px favicon to large signage
- Flat design, minimal strokes, no gradients unless essential
- Earth-tone palette: warm wheat gold, deep brown, cream
- Could incorporate subtle wheat or grain element
- Text must be perfectly legible and properly kerned

Output: Single centered logo on plain cream background. Generous padding around
the design for flexibility.

No watermarks, no mockups, no 3D effects, no complex imagery. Simple, functional,
timeless design.

複数のバリエーションを生成するにはn=4を使用します。ロゴデザインは主観的なものです — 選択肢を用意しましょう。

Create a realistic mobile app UI mockup for a local farmers market app.

Screen content (from top):
- Simple header with market name "Riverside Market" and search icon
- Today's featured vendor carousel with square photos
- "Fresh Today" section with produce category chips (Vegetables, Fruits, Dairy, Baked)
- Vendor list with small photos, names, specialties, and distance
- Bottom navigation: Home, Map, Favorites, Cart, Profile

Design language:
- White background, subtle natural green accents
- Clear typography hierarchy (system fonts feel)
- Generous padding and touch-friendly targets
- Looks like a real shipped product, not a concept
- Uses realistic vendor names and produce photos

Frame: Place the UI inside an iPhone 15 Pro device frame, slight perspective
tilt, subtle shadow beneath.

レイアウト、階層、間隔、リアルなインターフェース要素に焦点を当てます。概念的または芸術的な言葉は避けてください。

Create a 4-panel vertical comic strip. Equal panel sizes, clear panel borders.

Panel 1: Pet owner walks out the front door, keys in hand. Through the window
behind them, we see their cat watching — paws pressed against glass, eyes wide
with apparent sadness. The house suddenly feels empty.

Panel 2: The door clicks shut. The cat slowly turns away from the window toward
the empty house. Its posture shifts from forlorn to interested. Eyes narrow with
possibility.

Panel 3: Total chaos. Cat sprawled across the forbidden couch like royalty.
Knocked over plant on the floor. Papers scattered. Sunbeam spotlighting the
scene of domestic crime.

Panel 4: Door handle turns. Cat sits perfectly upright by the entrance,
composed and innocent, tail wrapped neatly around paws. Not a hair out of
place. As if nothing happened.

Style: Warm illustrated style with expressive characters, clear visual
storytelling that reads without text. Consistent character design across
all panels.

No speech bubbles or text. Let the visuals tell the story.

各パネルを明確なアクションを伴う個別の視覚的ビートとして定義します。モデルはパネルレイアウトと視覚的な連続性を処理します。

Create a children's book illustration introducing the main character.

Character: Young forest hero, around 8 years old.
- Green hooded tunic (think woodland adventurer, not Robin Hood)
- Soft brown boots, well-worn
- Small belt pouch for collecting treasures
- Carries a tiny wooden bow (symbolic, for helping not hurting)
- Kind expression, bright curious eyes, brave but gentle demeanor
- Slightly oversized head for picture book proportions

Theme: This character protects and rescues small forest animals in trouble.

Style: Hand-painted watercolor look with soft outlines, warm earthy palette
with forest greens and autumn oranges. Whimsical, friendly, inviting for
young readers ages 4-8.

Composition: Character standing in simple forest glade, dappled sunlight,
leaving room for title text above. Character clearly showcased.

Original character design only. No text. No watermarks. No copyrighted
character references.

このキャラクター参照画像を保存してください — 後のイラストで一貫性を維持するために必要になります。

Create a realistic outdoor crowd scene in Bethel, New York on August 16, 1969.

Photorealistic, period-accurate clothing, staging, and environment.

Documentary photography style, shot on film, natural lighting.

モデルは言われなくてもこれがウッドストックであることを知っています。日付と場所だけで、ヒッピー、当時のファッション、フェスティバルの雰囲気を生成します。

Edit the image to dress this person in the provided clothing items.

MUST PRESERVE (do not change in any way):
- Face, facial features, expression, skin tone
- Body shape, proportions, and pose
- Hairstyle and hair color
- Background and environment
- Camera angle, framing, and composition
- Overall lighting direction and quality

CHANGE ONLY:
- Replace current clothing with provided garment images
- Fit garments naturally to body geometry
- Show realistic fabric draping, folds, and behavior
- Match lighting and shadows on fabric to original photo

REQUIREMENTS:
- Photorealistic integration — outfit should look worn, not pasted
- Maintain color temperature of original image
- No accessories, text, logos, or watermarks added
- Identity must remain clearly recognizable

バーチャル試着には、顔の類似性を確実に保持するために常にinput_fidelity="high"を使用してください。

Using the EXACT visual style of the reference image (Image 1), create:
A man riding a motorcycle on a winding mountain road.

STYLE ELEMENTS TO MATCH PRECISELY from reference:
- Color palette and saturation levels
- Line quality and weight
- Texture treatment and brushwork
- Lighting style and direction
- Level of detail vs. abstraction
- Overall artistic aesthetic

APPLY TO NEW CONTENT:
- Single subject (man on motorcycle)
- Clear composition with visual interest
- Mountain road environment with curves
- Sense of motion and freedom

The new image should look like it came from the same artist or series as
the reference. Maintain stylistic consistency exactly.

スタイル転送は、どのスタイル要素を保持し、どのコンテンツ要素を変更するかを具体的に指定した場合に最もよく機能します。

In this room photo, replace ONLY the white plastic chairs with
mid-century modern wooden chairs (walnut finish, tapered legs,
woven seat).

PRESERVE COMPLETELY:
- Camera angle and perspective
- Room lighting direction and quality
- All other furniture and objects
- Wall colors and decorations
- Floor material and shadows
- Overall image quality and color grading

INTEGRATION REQUIREMENTS:
- Chairs must match room's perspective exactly
- Wood grain should catch existing light realistically
- Contact shadows must be natural and match light source
- Scale must be accurate relative to table height
- New chairs should look like they belong in this room

Photorealistic result — should look like the original photograph.

インテリアデザインの視覚化は、最も商業的に価値のある編集アプリケーションの1つです。

Transform this hand-drawn sketch into a photorealistic image.

PRESERVE FROM SKETCH:
- Exact layout and proportions
- Perspective and viewing angle
- Element placement and relationships
- Implied depth and layering

ADD FOR REALISM:
- Appropriate real-world materials and textures
- Consistent natural lighting (interpret from sketch shading)
- Environmental context matching the implied setting
- Surface imperfections and wear appropriate to materials

CONSTRAINTS:
- Do not add new elements not present in sketch
- Do not add text or watermarks
- Treat the sketch as an architectural blueprint to follow exactly
- Fill in realistic details while honoring the original composition

モデルはスケッチの意図を解釈し、元の構図を尊重しながらリアルな詳細を埋めます。

Transform this daytime summer scene into a winter evening with snowfall.

CHANGE:
- Time of day: from afternoon to dusk (warm interior lights visible)
- Season: summer to deep winter
- Weather: clear to active snowfall
- Ground: grass to fresh snow coverage
- Trees: summer foliage to bare branches with snow
- Atmosphere: add visible breath if people present
- Surfaces: add frost on windows and metal

PRESERVE:
- Camera position and angle exactly
- All objects and their exact positions
- Architecture and structural elements
- People and their poses (update clothing appropriately)
- Overall composition and framing

Style: Photorealistic, natural atmospheric perspective, visible
snowflakes in air, cozy contrast between warm interior lights and
cold exterior. Should feel photographed, not filtered.

環境変換で最良の結果を得るには、input_fidelity="high"とquality="high"を使用してください。

I'm providing 2 images:
- Image 1: Beach scene with woman standing on shore at sunset
- Image 2: Golden retriever sitting in a studio setting

Task: Place the dog from Image 2 into the beach scene from Image 1,
positioned next to the woman, looking up at her.

MATCHING REQUIREMENTS:
- Dog's lighting must match beach sunset (warm golden light from left)
- Scale dog appropriately relative to woman's height
- Dog should cast shadow consistent with scene's sun angle
- Sand texture should show around and under dog's paws
- Fur should catch the same golden hour highlights as scene

PRESERVE FROM IMAGE 1:
- Woman's exact appearance, position, and pose
- Beach background completely unchanged
- Original photo's color grading and mood

The composite should look like a single photograph taken on location.
No visible compositing artifacts.

番号で画像を参照し、どの要素が転送され、どの要素が固定されたままであるかを明示してください。

Translate all text in this infographic from English to Japanese.

MUST PRESERVE:
- Exact layout, spacing, and positioning of all elements
- All visual elements, icons, illustrations, and graphics
- Typography hierarchy (headlines vs body text relationships)
- Color scheme and overall design aesthetic
- Font weights and relative sizes

TRANSLATION REQUIREMENTS:
- Accurate Japanese translation with natural phrasing
- Match visual weight and style to original fonts
- Adjust character spacing for Japanese typographic norms
- No text truncation or overflow outside original bounds

Do not modify any non-text elements. Only change the language.

このワークフローは、ゼロから構築することなく、マーケティング資料、UIスクリーンショット、パッケージング、インフォグラフィックを処理します。

Continue the children's book story using the character from the reference image.

New Scene:
The same young forest hero is gently helping a frightened squirrel out
of a fallen hollow tree after a winter storm. Snow on the ground, bare
branches above, warm light filtering through clouds.

CHARACTER CONSISTENCY (from reference):
- Same green hooded tunic, exact shade and style
- Same soft brown boots
- Same belt pouch
- Same facial features, proportions, and color palette
- Same gentle, heroic personality in expression
- Same children's book proportions

STYLE CONSISTENCY (from reference):
- Same watercolor illustration style
- Same soft outlines
- Same warm earthy color treatment
- Same whimsical, friendly aesthetic

New elements: winter forest environment, frightened squirrel, fallen
tree with hollow.

Do not redesign the character. Do not change the artistic style.
No text. No watermarks.

本全体で一貫性を維持するために、アンカー画像を参照し、主要なキャラクターの詳細を繰り返します。

Create a hyper-stylized 3D floating head portrait based on this person.

STYLE CHARACTERISTICS:
- Smooth skin with glossy vinyl-finish surface
- Strong highlighter on cheekbones and nose tip catching soft light
- Holographic, iridescent eyeshadow (purple to teal color shift)
- Thick hair sculpted in slick, glossy waves like polished acrylic
- Small metallic chrome nose piercing with brushed reflections

EXPRESSION:
Confident, slightly unimpressed look — half-lidded eyes, subtly
arched brow, the sophisticated "too cool" attitude.

TECHNICAL SPECIFICATIONS:
- Head floats isolated against plain white background
- Slight 15-degree tilt (premium product render feeling)
- Bright, diffuse studio lighting with no harsh shadows
- Emphasis on glossy, plastic, subsurface scattering effects
- Ultra-smooth textures throughout
- Close-up portrait angle, straight-on, 85mm lens feel

The result should look like a high-end 3D character render or
collectible figure — plastic perfection with personality.

このレベルの美的詳細は、さまざまな被写体に対して驚くほど一貫した結果を生み出します。

Transform this person into an adorable chibi-style character.

CHIBI PROPORTIONS:
- Tiny body (about 1 head-height tall)
- Oversized head (3x body proportions)
- Large, sparkling eyes with cute highlights
- Soft, rounded facial features
- Cheerful, expressive pose with personality

PRESERVE FROM ORIGINAL:
- Recognizable facial features (simplified but identifiable)
- Hairstyle, length, and hair color
- Distinctive clothing style or accessories
- Any notable characteristics (glasses, jewelry, etc.)
- Overall personality and vibe

STYLE:
- Smooth pastel shading
- Clean lines and simplified details
- Bright, expressive colors
- Collectible figure aesthetic

Background: Simple gradient or plain color to showcase character.

The result should feel like an irresistible chibi mascot that
clearly represents the original person.

ちび変換は、パーソナルブランディング、チームアバター、商品デザインに適しています。

Create a realistic highway billboard mockup featuring this product.

BILLBOARD CONTENT:
- Product bottle prominently displayed on left third
- Main headline on right (EXACT TEXT, render verbatim):
  "Fresh & Clean — Every Day"
- Tagline below headline: "Nature's Best Ingredients"
- Small logo placeholder area in bottom right corner

TYPOGRAPHY SPECIFICATIONS:
- Headline: Bold sans-serif, white text, high contrast
- Tagline: Light sans-serif, slightly smaller, same white
- Clean kerning, centered alignment within text area
- Text appears EXACTLY ONCE — no duplicates anywhere

SCENE:
- Billboard on highway overpass or roadside structure
- Sunset lighting creating warm, appealing atmosphere
- Photorealistic environment with motion-blurred vehicles below
- Professional advertising photography feel

No watermarks. No additional marketing copy. No logos unless
specified. Text must be perfectly legible and correctly spelled.

テキストを含むマーケティング資料には常にquality="high"を使用してください。最終使用前にスペルを確認してください。

Extract the product from this image for e-commerce use.

OUTPUT SPECIFICATIONS:
- Transparent background (RGBA PNG format)
- Crisp silhouette with clean edges
- No halos or color fringing around product
- All product labels and text perfectly preserved
- Exact product geometry and proportions maintained

OPTIONAL ENHANCEMENT:
- Add subtle, realistic contact shadow
- Shadow should be soft and natural, no hard edges
- Shadow works with the transparent background

CRITICAL CONSTRAINTS:
- Do NOT restyle or recolor the product
- Do NOT modify product appearance in any way
- Only remove background and add optional shadow
- Preserve every detail of the original product exactly

注意：現在のモデルは、透明度を示すために市松模様をレンダリングします — 真のアルファチャネルには後処理が必要な場合があります。

import os
import base64
from openai import OpenAI

client = OpenAI()

# Create output directory
os.makedirs("output_images", exist_ok=True)

def save_image(result, filename: str) -> None:
    """Save base64 image response to file."""
    image_base64 = result.data[0].b64_json
    with open(f"output_images/{filename}", "wb") as f:
        f.write(base64.b64decode(image_base64))

# Basic text-to-image generation
result = client.images.generate(
    model="gpt-image-1.5",
    prompt="Your detailed prompt here",
    quality="high",  # or "low" for faster iteration
    n=1  # number of variations
)

save_image(result, "output.png")

result = client.images.edit(
    model="gpt-image-1.5",
    input_fidelity="high",  # Essential for identity preservation
    quality="high",
    image=[
        open("input_images/source.png", "rb"),
        open("input_images/style_reference.png", "rb"),
    ],
    prompt="""
    Apply the artistic style from Image 2 to the subject in Image 1.

    PRESERVE: subject's identity, pose, and composition
    CHANGE: artistic style, color palette, texture treatment

    Do not add new elements. Maintain subject likeness exactly.
    """
)

save_image(result, "styled_output.png")