GPT Image 2 — Complete Prompting Guide

A Human-Friendly Reference for Creative Professionals and Enthusiasts

Welcome: What Is GPT Image 2 and Why Should You Care?

Imagine briefing a talented art director who not only listens to your idea, but pauses to think it through — sketching the layout in their head, checking that the typography will fit, looking up a reference if they're unsure, and catching their own mistakes before they hand you the final piece. That is the closest analogy to what GPT Image 2 does every time you press generate.

GPT Image 2 is OpenAI's flagship image generation and editing model, released on April 21, 2026 and built on the GPT-5.4 backbone. Inside ChatGPT it appears under the consumer name ChatGPT Images 2.0; in the API it is the model gpt-image-2. It replaces both DALL·E 3 and the interim GPT Image 1.5, and it is the first OpenAI image model with reasoning built in. Before it commits to a single pixel, it can plan composition, search the web for real-world references, and self-check its own output against your instructions.

For creative professionals — designers, marketers, art directors, illustrators, product teams — this changes how you write prompts. You no longer reach for keyword soup and legacy "magic words." You write the way you would brief a skilled collaborator: full sentences, clear intent, a named deliverable, and a short list of things that must not change. The clearer you are about what you want and why, the more the reasoning layer has to work with.

This guide is your mentor for that conversation. It covers the raw technical specs (for when you need them), the mental model behind the reasoning engine, copy-paste-ready prompts drawn from OpenAI's own cookbook, and a workshop that builds one great prompt from scratch. Whether this is your first image model or you're migrating from DALL·E 3, Midjourney, or a diffusion tool, you'll find what you need here.

Quick Start — Jump Right In

Not ready to read the whole guide yet? Here are three copy-paste-ready prompts you can use immediately. Each one demonstrates a different strength of the model — photorealism, dense text layout, and clean design.

Quick Start Prompt 1 — Photorealistic Portrait

A photorealistic candid photograph of an elderly sailor standing on a small fishing boat at dawn. Weathered skin with visible wrinkles, pores, and sun texture; a thick wool sweater worn at the cuffs. Shot like a 35mm film photograph, medium close-up at eye level, 50mm lens. Soft coastal daylight, shallow depth of field, subtle film grain, natural color balance. The image should feel honest and unposed — real skin texture, worn materials, everyday detail. No glamorization, no heavy retouching.

Quick Start Prompt 2 — Infographic (text-heavy)

Create a clean, modern infographic explaining how an automatic espresso machine works, from bean hopper to grinder to boiler to extraction. Show the flow as a left-to-right diagram with five labeled stages, each with a simple line icon and a one-line caption. Dark navy background, white Inter font, warm amber accent lines. Title at top: "FROM BEAN TO CUP". 16:9, high quality for legible small text.

Quick Start Prompt 3 — Logo / Brand Mark

Create an original, non-infringing logo for a bakery called "Field & Flour". Warm, simple, timeless. Flat design, minimal strokes, no gradients. A single centered wordmark with a small wheat-sheaf icon above it, plain background, generous padding. Render the text exactly as "Field & Flour". No watermark, no extra text.

Once you're comfortable with the basics, come back and read the full guide — the reasoning layer rewards a more deliberate prompt than you might be used to.

Table of Contents

1. Model Identity
2. Technical Specifications
3. Core Architecture Insight — How Reasoning Changes Image Generation
4. Golden Rules of Prompting
5. Prompting Frameworks
6. Text Rendering
7. Creative Director Controls
8. Character and Identity Consistency
9. Multi-Image and Editing Workflows
10. Resolution, Quality, and Output Settings
11. Thinking Mode in Practice (Advanced)
12. GPT Image 2 vs. GPT Image 1.5 vs. DALL·E 3 — When to Use Which
13. Common Pitfalls
14. Production-Ready Example Prompts
15. Common Mistakes → Better Versions
16. Prompt Building Workshop
17. Cheat Sheet — Quick Reference

1. Model Identity

GPT Image 2 sits at the top of OpenAI's image stack. The single most important thing to understand about it is in the last row: it is a thinking image model. Everything else in this guide flows from that one fact, so we'll unpack it carefully in Section 3.

2. Technical Specifications

What These Numbers Mean in Practice

On resolution: Forget "4096×4096." The real ceiling is a total pixel budget of 8,294,400 — which is exactly 4K UHD (3840×2160) in a wide frame, or roughly 2880×2880 if you want a square. Both edges must be multiples of 16, and no single edge can reach 3840px, so the model nudges your requested size to the nearest legal dimensions. For most work, 2K (around 2048px on the long edge) is the practical baseline; reach for the full budget only for large print or billboard mock-ups.

On quality tiers: quality is your single biggest cost-and-latency lever. Use low for high-volume ideation and drafts — it's fast and cheap, comparable to previous-generation output. Step up to medium or high for the things that punish errors: small or dense text, detailed infographics, close-up portraits, and identity-sensitive edits. A good habit is to explore at low, then re-run your chosen direction at high.

On cost: Because pricing is by output tokens, a small low-quality image is genuinely cheap — a batch of fifty exploratory 1024² drafts is a rounding error. The expensive combination is large size × high quality, which is exactly where you want to spend when an image is going to print or in front of a client. Streaming partials costs a little extra (≈100 tokens each) but lets you bail early on a bad generation.

On transparency: Unlike some prior models, gpt-image-2 does not output a true alpha channel. If you need a transparent PNG for compositing, generate on a clean opaque background (pure white or a flat color) and run a background-removal step afterward.

3. Core Architecture Insight

How Reasoning Changes the Way You Write Prompts

Most image generators treat your prompt as a bag of words — they activate concepts probabilistically and paint, with no real "understanding" of what you meant. GPT Image 2 is the first OpenAI image model that works differently.

It runs a reasoning pass — a "thinking" step — before it generates. Riding on the GPT-5.4 backbone, it interprets your instructions the way a language model does: it plans the composition, weighs what you said against what you almost certainly meant, and can even call web search to ground the image in real facts (a real skyline, a current logo, today's data). After drafting, it can self-check the result against your constraints — did the text spell correctly? did it keep the face you asked to preserve? — and correct before handing the image back.

This has three practical implications for how you prompt:

1. The prompt is an art-direction brief, not a tag list. Name the deliverable ("a pitch-deck slide," "a billboard mock-up," "a children's book illustration") and the model infers the right level of polish, layout, and convention. Telling it the job is often worth more than ten style keywords.
2. It fills gaps sensibly, so you control by constraint. Leave something unspecified and the reasoning layer picks a reasonable default rather than producing noise. That's great for speed — but it means the way you take control is by stating constraints and invariants explicitly, not by piling on adjectives.
3. Coherent prose beats fragments. Because it reasons over language, "A ceramic mug of steaming coffee on a worn wooden table, photographed in soft morning light" gives it far more to work with than "coffee, mug, wood, warm, bokeh, 4k." The first has intent; the second is a word cloud.

4. Golden Rules of Prompting

These seven rules distill OpenAI's published prompting guidance for GPT Image 2 into the habits that matter most. Master these before reaching for any advanced technique.

Rule 1 — Brief the Job, Not Just the Subject

Name the deliverable. "A pitch-deck slide," "a billboard mock-up," "a children's book illustration," "a product hero shot for a landing page" — each sets a different mode, and the reasoning layer adjusts polish, layout, and convention accordingly.

Why this works: "A phone on a table" could be a snapshot, a meme, or a catalog image. Naming the job tells the model the audience and the standard, so it makes the dozens of small compositional choices the way a professional would for that use case.

Rule 2 — Be Concrete About Materials, Shape, and Medium

Vague subjects produce average images. Name the material, the shape, and the visual medium — photo, watercolor, 3D render, line illustration. Concrete nouns give the reasoning layer something specific to commit to.

Why this works: Specific materials and a stated medium establish a voice for the image. The more particular your description, the more the model leans into your vision instead of defaulting to the average of everything it has seen.

Rule 3 — Say "Photorealistic" When You Mean It

For photographic results, put the word photorealistic directly in the prompt — it strongly engages the model's photographic mode. Then reinforce it with real-world texture: "visible pores, fabric wear, fine imperfections."

The flip side: if you want a candid, documentary look, avoid words that imply studio staging. Use "candid," "unposed," "honest" instead, and skip cinematic color-grading language, which pushes the result toward a polished, art-directed look.

Why this works: The model has distinct registers — glossy studio, candid documentary, illustrated, rendered. Naming the register explicitly is far more reliable than hoping it infers the one you want from the subject alone.

Rule 4 — Direct the Camera and the Light

Think like a photographer. Specify framing, viewpoint, and lighting — they are compositional and emotional choices, not just technical ones.

• Framing: "close-up", "medium shot", "wide", "top-down flatlay"
• Viewpoint: "eye-level", "low angle", "45-degree overhead"
• Lighting: "soft diffuse window light", "golden-hour backlight", "hard single key light"
• Optics: "50mm lens, shallow depth of field (f/2.0)" — but keep these high-level; the model treats exact camera specs as a look, not a physics simulation

Why this works: A low-angle, shallow-focus close-up feels powerful and intimate; a high, deep-focus wide shot feels observational. Specifying these puts the emotional decision in your hands instead of leaving it to chance.

Rule 5 — Separate What Changes From What Stays

This is the single most important habit for editing and iteration. Always say what should change and what must remain exactly the same — and restate the "keep" list on every follow-up.

Example follow-up: "Change only the jacket to charcoal wool. Keep her face, hair, pose, and the background identical."

Why this works: Without an explicit invariant list, edits drift — the model "helpfully" adjusts the face or lighting to match the change you asked for. Naming the invariants gives it permission to leave everything else alone, and the reasoning pass actively checks the result against that list.

Rule 6 — Exclude by Naming the Desired State

Where you can, describe what you want rather than what to avoid. But GPT Image 2 also handles a small set of explicit, concrete exclusions well — "no watermark," "no extra text," "no logos or trademarks" — which are worth stating outright on commercial work.

Why this works: Positive descriptions are unambiguous — there is nothing to imagine-then-suppress. Reserve hard negatives for short, concrete production guardrails like watermarks and stray text.

Rule 7 — Iterate in Small, Single-Change Steps

Long prompts work, but they're hard to debug. Start from a clean base prompt, then refine with small, single-change follow-ups. If a generation is 80% right, edit it — don't re-roll from scratch.

5. Prompting Frameworks

These are the five core shapes a prompt can take, depending on what you're doing and what inputs you have. Each has a formula and a worked example.

5.1 Text-to-Image — The Art-Direction Brief

The most common case: you have an idea and want an image. OpenAI's recommended structure is a brief written in a consistent order.

Formula:

[Context / use case] + [Scene / background] + [Subject] + [Key details] + [Style / medium] + [Constraints] + [Exact copy]

Why this works: Each slot answers a question the model would otherwise guess at: what is this for, where are we, who/what is the subject, what are the telling details, what does it feel like, what must not happen, and what text appears verbatim. Together they form a complete brief — nothing critical is left to default.

Example:

A polished campaign image for a youth streetwear brand called Thread. A confident model in an oversized varsity jacket and wide trousers stands in a sunlit concrete underpass, mid-stride, looking off-frame. Energetic, contemporary, tasteful — like a real seasonal lookbook cover. Shot on 35mm, natural light, slight film grain. Render the tagline exactly once, clearly and legibly, integrated into the layout: "MADE TO MOVE". No extra text, no watermarks, no unrelated logos. 4:5, high quality.

5.2 Conversational Editing

You want to change an image — one the model generated, or one you uploaded. The discipline here is the change/keep split from Rule 5.

Formula:

[What to change] + [How to change it] + [What to keep exactly the same]

Why this works: The model already knows the image. You only need to describe the delta — and, crucially, the invariants. Restating the "keep" list on every turn is what prevents identity, geometry, and layout from drifting across iterations.

Examples:

• "Remove the flower from the man's hand. Do not change anything else." — surgical object removal
• "Change the background to a neon-lit Tokyo street at night. Keep the subject, pose, and lighting on her face identical." — background swap with explicit preservation
• "Make it overcast and rainy — wet pavement, soft grey light — but keep the building geometry, camera angle, and people exactly as they are." — weather/lighting change with locked structure

5.3 Multi-Image Compositing

You have several input images and want to combine them. The cardinal rule: reference each input by index and describe how they interact.

Formula:

Image 1: [what it is]. Image 2: [what it is]. [How they should interact] + [what to preserve].

Why this works: Naming each image and its role removes the guesswork. Without it, the model has to infer which image is the subject, which is the background, and which is just a style reference — and it will sometimes guess wrong.

Example:

Place the dog from Image 2 into the setting of Image 1, right next to the woman. Use the same style of lighting, composition, and background as Image 1, and match the dog's shadow and color temperature to the scene. Do not change anything else about Image 1.

5.4 Style Transfer and Virtual Try-On

Two specialized edits that live or die on a clean "preserve vs. change" split.

Style transfer — keep the content, change the rendering. Be specific about which aspects of the style you want: palette, brushwork, line character, grain.

Recreate this photograph of a city street as a Van Gogh-style oil painting — thick impasto brushstrokes, swirling navy-and-gold sky, vivid saturated color. Keep the composition, the buildings, and the perspective; change only the rendering style.

Virtual try-on — change the garment, lock the person. This is the canonical identity-preservation edit.

Edit the image to dress the woman in the clothing from the provided product images. Do not change her face, facial features, skin tone, body shape, pose, or identity in any way. Replace only the clothing, fitting the garments naturally to her existing pose and body with realistic fabric behavior. Match the lighting, shadows, and color temperature of the original photo so the outfit integrates photorealistically.

Why this works: The reasoning layer can separate "what is in the image" (content, identity) from "how it's rendered" (style, garment). Telling it exactly which layer to hold fixed is what keeps a style transfer from mangling the composition, or a try-on from quietly changing the model's face.

5.5 Web-Search Grounding

Because GPT Image 2 reasons over your request, it can call web search to ground an image in real, current facts before drawing. This makes data-driven and reference-accurate images far less generic.

Formula:

[What to look up] + [How to use the data] + [How to render it visually]

Why this works: Grounding an image in real numbers or real references makes it immediately credible. A market-cap infographic built from looked-up figures is more accurate and more useful than one invented from memory — and the three-part formula ensures you get the right data, processed the right way, rendered the way you want.

Example:

Search for the current top 5 companies by market capitalization. Build a clean horizontal bar chart infographic showing each company's name and market cap. Dark navy background, white Inter font, one accent color per bar, Bloomberg-terminal aesthetic. 16:9, high quality so the numbers stay legible.

6. Text Rendering

Text-in-image is one of GPT Image 2's headline strengths: roughly 99% character-level accuracy across Latin, Japanese, Korean, Chinese, Hindi, Bengali, and Arabic scripts. That makes it genuinely usable for posters, packaging, infographics, and UI mockups — but only if you prompt the text deliberately.

The Rules of Text in Prompts

1. Put literal copy in quotes (or ALL CAPS). Quoting tells the model "render this exactly." "HAPPY NEW YEAR", "Sale Ends Sunday", "Field & Flour". Unquoted text invites the model to paraphrase.
2. Demand it verbatim. For anything that must be exact, say so: "render this text exactly, verbatim, with no extra characters."
3. Specify the typography. Don't just say "text" — say "bold white sans-serif," "thin letter-spaced caps," "elegant Garamond italic," and give placement ("centered," "top-right with generous padding").
4. Spell out tricky words. For brand names or uncommon spellings, spell them letter-by-letter in the prompt to raise character accuracy.
5. Use medium or high quality for dense text. Small type, multi-font layouts, and information-dense panels are exactly where low quality drops or smears characters. Step up.
6. Multilingual is reliable — name the language. Ask explicitly: "render the headline in Korean," or request several languages at once (English, then Arabic right-to-left, then Korean).

Example: Multi-Typography Layout

A glossy beauty product shot of a minimalist nude-colored moisturizer jar on a pale stone surface. Next to the product, render three lines of text, each verbatim: top line "GLOW" in a flowing elegant brush script; middle line "10% OFF" in a heavy blocky condensed font; bottom line "Your First Order" in a thin minimalist sans-serif. Below them, add the same offer translated into Korean. High quality, 1:1.

7. Creative Director Controls

This section is your vocabulary reference — the specific terms that give you fine-grained control over the look and feel of an image.

7.1 Lighting

Lighting is the highest-leverage control in any image. Getting it right turns a competent render into a striking one.

7.2 Camera, Lens, and Focus

Use photographic language for framing and feel — but keep it high-level. GPT Image 2 treats detailed camera specs as a look, not a physical simulation, so name them for composition and mood rather than expecting exact optical behavior.

• Shot type: "close-up", "medium shot", "full body, feet included", "wide", "top-down"
• Lens feel: "35mm documentary", "50mm natural", "85mm portrait compression", "24mm wide-angle", "macro"
• Depth of field: "shallow depth of field, soft bokeh" (f/1.4–f/2.8 feel) vs "deep focus, everything sharp" (f/8–f/16 feel)
• Angle: "eye-level", "low angle", "45-degree overhead", "dutch tilt"

7.3 Color Grading and Mood

Color is mood in shorthand. Film-stock and grading references are a fast way to set an emotional target.

7.4 Materiality and Texture

Name materials explicitly — it's the difference between a generic object and a believable one, especially for product, fashion, and architecture.

• Fabric: "raw silk dupioni", "weathered canvas", "cashmere ribbed knit", "distressed denim"
• Metal: "brushed steel", "hammered copper", "anodized matte black aluminum", "aged patina bronze"
• Ceramic / glass: "matte stoneware glaze", "frosted borosilicate glass", "iridescent glaze"
• Natural: "reclaimed oak", "Carrara marble", "woven rattan", "polished obsidian"
• Skin / realism: "visible pores, fine wrinkles, natural skin texture" — the antidote to plastic, over-retouched faces

8. Character and Identity Consistency

Keeping the same person or character recognizable across multiple images is a core production need — for storyboards, children's books, brand mascots, and casting-style series. GPT Image 2's reasoning and automatic high-fidelity input handling make this reliable when you work in two steps.

Step 1 — Anchor the Identity

Generate (or supply) one clear reference and describe it precisely: appearance, outfit, proportions, and personality. This becomes your anchor.

Design a children's book character: a small raccoon named Pip, wearing a red scarf and round brass spectacles, with a curious, gentle personality. Soft watercolor illustration style, warm palette. Front-facing, full body, plain background. This is the definitive character reference.

Step 2 — Reuse With an Explicit Keep-List

For each new scene, vary the environment, pose, and narrative — but restate the identity invariants and tell the model not to redesign the character.

Using the character reference, show Pip in three scenes: 1) reading under a giant mushroom in a rainy forest; 2) gazing up at a starry sky from a hilltop; 3) stirring a bubbling pot in a cozy burrow kitchen. Keep the same red scarf, the same round brass spectacles, the same facial features, proportions, and gentle personality. Do not redesign the character. Same watercolor style and warm palette across all three. 3:4, high quality.

Why this works: Specific, repeatable anchors (the red scarf, the brass spectacles) give the reasoning layer strong identity cues, and "do not redesign" plus a restated keep-list lets it vary scene without drifting the character. For human likeness, the same pattern applies — name the face, hair, and build as invariants on every turn.

9. Multi-Image and Editing Workflows

Beyond the basic frameworks, here's a reference table of the editing jobs GPT Image 2 handles well and the approach that gets each one right. The through-line is always the same: state what changes, state what stays.

Referencing Multiple Images

When you pass more than one image, label each by index and give it a role — exactly as in Section 5.3. A useful vocabulary:

• "Image 1: the product to feature — keep its exact shape, color, and label."
• "Image 2: the background environment."
• "Image 3: the color-palette reference — use it for overall tone only."

Mask-Based Edits

For precise, region-locked edits, the Image API's edit endpoint accepts a mask. The mask must share the input image's format and size and carry an alpha channel; the transparent area marks where the model may paint. When you supply multiple images, the mask applies to the first one.

10. Resolution, Quality, and Output Settings

Size and Aspect Ratio

You request a pixel size (or auto), and the model snaps it to the nearest legal dimensions — both edges multiples of 16, longest edge under 3840px, aspect ratio no wider than 3:1, total pixels between 655,360 and 8,294,400.

Quality Tiers

Output Settings

• Format: png (default), jpeg, or webp. For jpeg/webp you can set output_compression (0–100).
• Background: opaque or auto. True transparency isn't supported on gpt-image-2 — render opaque and remove the background downstream if you need an alpha channel.
• Count: n generates multiple images per request — handy for logo or concept variations (e.g. n=4).
• Streaming: set partial_images (0–3) to receive intermediate renders, so you can abort a generation that's clearly going wrong.

11. Thinking Mode in Practice (Advanced)

The reasoning pass isn't a setting you toggle — it's always on. But you can prompt in ways that use it well instead of fighting it. Here's how to get value from the thinking layer.

Give It a Goal and Constraints, Not Micromanagement

Because the model plans before it paints, it handles long, structured briefs better than older generators — which means you can describe a complex layout in prose and trust it to lay things out sensibly. State the goal, the must-haves, and the must-nots; let it make the taste-driven decisions in between.

Design a one-page event poster for a jazz night. Must include, verbatim: the title "BLUE ROOM SESSIONS", the date "Fri Apr 24", and the venue "The Cellar, Seoul". Mood: intimate, late-night, warm. You choose the layout, type pairing, and imagery — make it feel like a real, well-designed gig poster. 2:3, high quality.

Lean on the Self-Check

The model can review its own output against your instructions before returning it. You make that more effective by stating checkable constraints explicitly — exact text, a preserved face, a required object. The more concrete the constraint, the more the self-check has to verify against.

Let It Ground in Reality

When browsing is available (in ChatGPT, or via tools in the Responses API), the reasoning layer can fetch real references — a real skyline, an actual product, current data — and build them into the image. This is the difference between a plausible-looking chart and a correct one (see Section 5.5).

Multi-Step Refinement in the Responses API

For programmatic workflows, the Responses API exposes image generation as a built-in tool and keeps context across turns. Chain edits with previous_response_id (or by passing the prior image generation call), and use the action hint — generate, edit, or auto — to steer whether a turn makes a new image or modifies the last one. This is the API-side equivalent of the conversational "edit, don't re-roll" discipline.

12. GPT Image 2 vs. GPT Image 1.5 vs. DALL·E 3 — When to Use Which

GPT Image 2 is the right default for almost everything. The earlier models still exist for a few specific reasons.

The practical rule: Reach for gpt-image-2 by default — for quality, for editing, for text, for reasoning-grounded work. Use gpt-image-1.5 only when you specifically need a true transparent asset or its fixed-size behavior. Use DALL·E 3 only for legacy compatibility. (There's also gpt-image-1-mini for ultra-cheap, rapid ideation when fidelity isn't critical.)

13. Common Pitfalls

14. Production-Ready Example Prompts

This is the most expanded section of the guide. For each category you'll find ready-to-use prompts adapted from OpenAI's cookbook patterns, each with a short note on the key choices.

14.1 Product Photography

Skincare — Minimalist Hero Shot

A photorealistic commercial hero shot of a frosted glass serum bottle with a matte gold dropper, centered on a smooth travertine surface. Soft directional daylight from a single large window on the left, casting one clean shadow. A dried white wildflower rests beside it, slightly out of focus. Shallow depth of field. Natural color balance, no heavy retouching. 1:1, high quality.

Key choices: "photorealistic" plus "natural color balance, no heavy retouching" keeps it from looking like a plastic 3D render; one window + one shadow reads as real photography; high quality protects the fine glass and droplet detail.

Tech — Floating Perspective

A clean product mockup of a slim wireless charging pad in matte white, shot at a 45-degree overhead angle with a soft floating shadow beneath it on a pure white background. A slim phone rests on the pad, screen glowing softly. Even, diffused studio light, no harsh highlights. Minimal, premium tech aesthetic. 1:1, high quality.

Key choices: the 45-degree angle plus soft floating shadow is the standard "hovering product" look; pure white background makes the asset easy to drop into marketing layouts; diffused light avoids distracting hotspots on the matte surface.

Spirits — Dark & Moody

A dramatic product shot of a single-malt whisky bottle on an oiled mahogany bar. Hard single key light from above creates a warm pool of amber and deep falloff shadows. Two heavy crystal tumblers flank the bottle, one with ice and light condensation. Shot like an 85mm photograph, shallow depth of field. 3:2, high quality.

Key choices: hard single key light and deep shadow match the premium, masculine positioning of whisky; condensation and crystal add tactile realism; "like an 85mm photograph" sets the compression without over-specifying optics.

14.2 Photorealistic Portrait

Documentary — Occupational Portrait

A photorealistic candid documentary portrait of a woman in her 50s at a busy fish market at dawn, apron on, surrounded by ice and fresh catch. Calm, proud expression, looking slightly off-camera. Weathered hands, real skin texture with visible pores. Flat early-morning light. Shot like a 35mm reportage photograph, natural color balance, no glamorization. 3:4, high quality.

Key choices: "candid documentary" plus "calm, proud expression" (not "smiling") prizes dignity over posing; "visible pores" and "no glamorization" fight the plastic-skin default; flat dawn light is honest rather than flattering.

Editorial Fashion

A fashion editorial portrait of a young woman with short silver hair in an oversized charcoal wool blazer over a sheer white blouse, leaning against a concrete pillar in an abandoned industrial space. Golden-hour light through broken windows casts warm geometric shadows. Low angle, shallow depth of field, Kodak Portra 800 grain. 3:4, high quality.

Key choices: a named film stock (Portra 800) sets the warm, grainy editorial tone in one phrase; the broken-window shadows give the frame structure; low angle adds presence and drama.

14.3 Infographics and Data Visualization

This is where GPT Image 2's text accuracy and reasoning shine. Use high quality so small labels stay crisp.

Climate Data — Area Chart

A clean modern infographic titled, verbatim, "Global Temperature Rise 1880–2024", showing an area chart with a clear upward trend. Dark navy-to-midnight gradient background. Y-axis: temperature anomaly in °C; X-axis: decades. The area fills from pale yellow to deep orange-red as temperature climbs. Annotate 1980 and 2023 as milestone markers. White Inter font, small grey source credit at the bottom. 16:9, high quality.

Key choices: quoting the exact title prevents the model from rewording it; the warm color ramp visually encodes the data's meaning; calling out specific milestone years gives the reasoning layer concrete anchors to label.

Process — Supply-Chain Diagram

A clean editorial infographic of a 5-step coffee supply chain: 1. Farm (Colombia), 2. Processing & Drying, 3. Quality Grading, 4. Roasting, 5. Packaging. Show each step as a labeled icon in a rounded square, connected left-to-right by arrows. Terracotta icons, pale gold arrows, white background, dark charcoal Inter labels. Title, verbatim: "From Farm to Cup". 16:9, high quality.

Key choices: numbered, named steps give the model exact content to lay out; a tight three-color palette keeps the diagram clean and on-brand; high quality keeps the multi-word labels legible.

14.4 Logos and Brand Marks

Minimalist Wordmark

Create an original, non-infringing logo for a bakery called "Field & Flour". Warm, simple, timeless. Flat design, minimal strokes, no gradients. A centered wordmark rendering the text exactly as "Field & Flour", with a small wheat-sheaf mark above it. Plain background, generous padding. No watermark, no extra text. 1:1.

Key choices: "original, non-infringing" steers away from existing brands; quoting the exact name protects the spelling and the ampersand; "no watermark, no extra text" is a concrete commercial guardrail. Run with n=4 for variations.

Emblem / Badge

A circular badge-style logo for a mountain coffee roaster called "Summit Roast". Vintage outdoor aesthetic: a simple two-color design (forest green and cream), a stylized mountain peak in the center, and the text "SUMMIT ROAST" curved along the top edge and "EST. 2026" along the bottom, both verbatim. Clean vector look, flat, plain background. 1:1, high quality.

Key choices: a two-color constraint reads as a real printable logo; curved top/bottom text is a classic badge layout the model executes well; quoting both text elements keeps them exact.

14.5 UI and Product Mockups

Mobile App Screen

A realistic mobile app UI mockup for a local farmers market, shown inside a clean iPhone frame. A simple header, a short list of vendors with small thumbnail photos and category tags, a compact "Today's Specials" section, and a location-and-hours row. White background, subtle natural accent colors, clear typography, minimal decoration. It should look like a real, well-designed app. 4:5, high quality.

Key choices: naming concrete UI sections gives the reasoning layer a real information architecture to render; "inside an iPhone frame" produces a presentation-ready mockup; high quality keeps the small UI labels readable.

Web Landing Hero

A web landing-page hero mockup for a calm productivity app. Left side: a headline, verbatim, "Focus, finally." with a one-line subhead and a single rounded call-to-action button reading "Start free". Right side: a clean app screenshot floating with a soft shadow. Generous whitespace, muted sage-and-stone palette, modern sans-serif. Desktop proportions. 16:9, high quality.

Key choices: quoting the headline and button label keeps the copy exact; specifying a left-text / right-visual split gives a real landing-page structure; whitespace and a muted palette read as a polished modern product site.

14.6 Editing and Composites

These assume you've supplied one or more input images. The discipline is always change/keep (Section 5).

Virtual Try-On

Edit the image to dress the woman in the jacket from the second image. Do not change her face, facial features, skin tone, body shape, pose, or identity in any way. Replace only the jacket, fitting it naturally to her pose with realistic fabric folds. Match the lighting, shadows, and color temperature of the original photo.

Key choices: the long, explicit "do not change" list is the whole point — it locks identity so only the garment moves; "match lighting and color temperature" is what makes the swap look photographed, not pasted.

Surgical Object Removal

Remove the parked car on the left side of the street. Reconstruct the road and curb behind it naturally. Do not change anything else — keep the buildings, people, light, and camera angle identical.

Key choices: naming exactly what to remove and "reconstruct behind it naturally" tells the model to inpaint, not just erase; the explicit keep-list stops it from quietly restyling the rest of the scene.

Weather / Lighting Change

Change this daytime street photo to a rainy evening — wet reflective pavement, warm shopfront glow, soft overcast-to-dark sky. Keep the building geometry, the storefronts, the camera angle, and the people exactly as they are. Only the weather, lighting, and time of day should change.

Key choices: scoping the change to "weather, lighting, time of day" and locking geometry/angle/people prevents the model from rebuilding the scene; "wet reflective pavement" and "warm shopfront glow" give it concrete cues for a believable rainy mood.

15. Common Mistakes → Better Versions

Learning to prompt is partly learning to recognize weak patterns and replace them. Here are eight before/after comparisons covering the most common failure modes.

Mistake 1 — Keyword Soup

Why the weak version fails: "masterpiece" and "ultra detailed" are legacy diffusion hacks that mean nothing to a reasoning model. Without narrative structure it has no spatial relationships or intent. The strong version gives it a scene, a mood, and a composition.

Mistake 2 — Negative-Only Framing

Why the weak version fails: Describe the desired state, not a pile of absences. Positive description is unambiguous; reserve one or two concrete negatives ("No vehicles") for the things that truly must not appear.

Mistake 3 — Vague Subject

Why the weak version fails: "A chef cooking" could be anyone doing anything. Age, appearance, specific action, specific food, and environment create a real, particular moment.

Mistake 4 — No Camera or Light Direction

Why the weak version fails: "Dramatic" is subjective. The strong version defines what makes it dramatic — the low angle, the foreground rocks, the lit peak. Camera and light direction are decisions, not optional detail.

Mistake 5 — Unquoted Text

Why the weak version fails: Unquoted text invites paraphrase and spelling errors. Quoting the copy and specifying typography per line yields a legible, designed poster.

Mistake 6 — Contradictory Style

Why the weak version fails: "Photorealistic" and "Studio Ghibli animation" are opposite rendering goals — the model averages them and satisfies neither. Pick one register and commit.

Mistake 7 — Re-Rolling Instead of Editing

Why the weak approach fails: Starting over discards the work the model already got right. A targeted edit preserves the keeper and fixes one thing — and with session context, the reasoning model is built for exactly that.

Mistake 8 — Editing Without an Invariant List

Why the weak version fails: With no keep-list, the model "helpfully" recolors clothing and skin to match the winter mood. Naming the invariants is what keeps the edit surgical.

16. Prompt Building Workshop

The most valuable prompting skill is starting from a vague idea and refining it into a production-ready brief. Here's a complete walkthrough — five iterations on a single concept, with the thinking at each step.

Starting Idea: "I want a product shot of a new cold brew can."

A fine starting point. Let's build it into something real.

Iteration 1 — Apply the Basic Brief

A product shot of a cold brew coffee can. Photorealistic.

What we have: A valid prompt — the model will produce something. But it's generic: average can, average background, average light.

What's missing: What does the can look like? What's the surface and setting? What's the light? What's it for?

Iteration 2 — Specify the Subject

A photorealistic product shot of a slim matte-black aluminum cold brew can with a minimalist cream-and-copper label. Condensation beads on the surface. Plain background.

What we added: Material (matte-black aluminum), label style (minimalist cream-and-copper), and a tactile detail (condensation) that signals "cold and fresh."

What's still missing: A real setting, deliberate light, and a camera decision.

Iteration 3 — Build Environment and Light

A photorealistic product shot of a slim matte-black aluminum cold brew can with a minimalist cream-and-copper label, beaded with condensation. It stands on a wet dark slate surface beside a few scattered coffee beans and a single eucalyptus sprig. Soft directional daylight from the left, one clean shadow. Plain pale background.

What we added: A specific surface (wet slate), supporting props that reinforce the story, and a single, real light source with a defined shadow.

What's still missing: Camera framing, color treatment, and purpose.

Iteration 4 — Add Camera and Color

A photorealistic product shot of a slim matte-black aluminum cold brew can with a minimalist cream-and-copper label, beaded with condensation, on a wet dark slate surface beside scattered coffee beans and a eucalyptus sprig. Soft directional daylight from the left, one clean shadow. Shot slightly above eye level, 85mm feel, shallow depth of field so the background falls away. Natural color balance, no heavy grading. Plain pale background.

What we added: Camera height and an 85mm compression feel, shallow depth of field to isolate the can, and "natural color balance, no heavy grading" to keep it believable rather than over-stylized.

What's still missing: Purpose, exact label text, and output settings.

Iteration 5 — Add Purpose, Text, and Polish

A premium e-commerce hero shot for a new cold brew brand. A slim matte-black aluminum can with a minimalist cream-and-copper label, beaded with fresh condensation, on a wet dark slate surface beside scattered coffee beans and a eucalyptus sprig. The label reads, verbatim: "NORTH" in bold copper caps and "Cold Brew" in thin cream serif below. Soft directional daylight from the left, one clean shadow. Shot slightly above eye level, 85mm feel, shallow depth of field. Natural color balance, no heavy grading. Plain pale stone background, generous negative space on the right for headline overlay. 4:5, high quality.

Final additions: Purpose ("e-commerce hero shot") orients every trade-off; quoted, per-line label text uses the model's text strength; "negative space on the right for headline overlay" is a production-aware touch that makes the asset drop-in ready; 4:5 and high quality lock the deliverable.

The five iterations compared:

The lesson: A production-quality prompt is usually 100–150 words. You don't start there — you build there, one layer of intent at a time. Each layer makes the image more specific, more useful, and easier for the reasoning layer to get right.

17. Cheat Sheet — Quick Reference

Everything you need on one page.

The Core Brief

[Use case / deliverable — name the job]
[Scene / background with details]
[Subject + specific materials and shape]
[Lighting]
[Camera: framing + viewpoint + lens feel]
[Style / medium / color treatment]
[Constraints + exact text in "quotes"]
[Output: aspect ratio + quality]

Non-Negotiable Rules

Size and Quality Quick Reference

Text Rendering Checklist

• Wrap literal copy in "quotes" and say "verbatim"
• Name the typography (weight, style, placement) per text element
• Spell out tricky brand names letter-by-letter
• Use high quality for small or dense text
• Name the language for non-English text; multilingual is reliable
• Proof logos and long strings before production use

Sources

This guide was compiled from OpenAI's official documentation:

• GPT Image Generation Models — Prompting Guide (OpenAI Cookbook)
• Image generation — OpenAI API guide
• gpt-image-2 — OpenAI model reference