GPT Image 2 Prompt Guide: How to Write Prompts for ChatGPT Images 2.0

GPT Image 2 (gpt-image-2) is OpenAI's current flagship image model and the default in ChatGPT since DALL-E 3 was retired on May 12, 2026. It's the most precise text-renderer of the current crop of image models (98.5% typo accuracy in head-to-head tests), the most reliable for identity-preserving edits, and the slowest of the four primary models at ~4,200ms per generation. The reason to use it is control, not speed.

This guide is built on OpenAI's official cookbook prompting guide (the first-party docs published when the model launched) plus the patterns that hold up in production. Below: the 7 fundamentals that make every prompt work, the 5 production-ready patterns, and the mistakes that silently break results.

How GPT Image 2 prompting differs from older OpenAI models

If you used DALL-E 3 or gpt-image-1.5, the prompt structure is similar but stricter. GPT Image 2 takes your prompt more literally, follows constraints more precisely, and fails harder on ambiguity. That's the upside (when the prompt is right, the result is right) and the downside (when the prompt is sloppy, the result is sloppy).

Three things to internalize before writing any prompt:

Order matters. Always go background/scene → subject → key details → constraints. The model reads prompts in sequence and weights earlier elements more heavily.
State the use case. "For an ad" or "for a UI mockup" sets the model's mode and level of polish. A photo prompt and a UI prompt with the same visual elements produce very different results.
State what to preserve. For edits, "change only X" + "keep everything else the same" works. For generations, the constraint list (no watermark, no extra text, no logos) prevents unwanted additions.

The 7 fundamentals (from OpenAI's official cookbook)

1. Structure + goal

Write prompts in a consistent order: background/scene → subject → key details → constraints. Include the intended use ("for an ad," "UI mock," "infographic") to set the mode and polish level. For complex requests, use short labeled segments or line breaks instead of one long paragraph.

Scene: editorial magazine spread, white background
Subject: typography-forward quote poster
Details: large serif headline, small caption, sidebar with bullet points
Use case: editorial print poster
Constraints: every word legible, professional magazine layout

2. Prompt format

Use whatever format is easiest to maintain. Minimal prompts, descriptive paragraphs, JSON-like structures, instruction-style prompts, and tag-based prompts all work. For production systems, prioritize a skimmable template over clever prompt syntax. The "clever" prompt that gets 10% better results isn't worth the maintenance burden.

3. Specificity + quality cues

Be concrete about materials, shapes, textures, and visual medium ("photorealistic," "watercolor," "3D render," "oil painting"). Add targeted quality levers when needed: "film grain," "textured brushstrokes," "macro detail." For photorealism, include the word "photorealistic" directly — it engages the model's photorealistic mode more reliably than "real-looking" or "lifelike."

Camera specs (85mm, f/1.4, ISO 400) work as high-level look and composition hints, but the model may interpret them loosely. Use them for vibe, not for exact physical simulation.

4. Latency vs fidelity

For high-volume or latency-sensitive use cases, start with quality="low". The low setting is meaningfully faster and provides sufficient fidelity for many use cases. For small or dense text, detailed infographics, close-up portraits, identity-sensitive edits, and high-resolution outputs, compare medium or high before shipping.

In practice: iterate at low, lock the final at high.

5. Composition

Specify framing (close-up, wide, top-down), perspective (eye-level, low-angle), and lighting (soft diffuse, golden hour, high-contrast). If layout matters, call out placement: "logo top-right," "subject centered with negative space on left." For wide, cinematic, low-light, rain, or neon scenes, add extra detail about scale, atmosphere, and color — the model otherwise trades mood for surface realism.

6. People, pose, and action

For people in scenes, describe scale, body framing, gaze, and object interactions:

"Full body visible, feet included"
"Child-sized relative to the table"
"Looking down at the open book, not at the camera"
"Hands naturally gripping the handlebars"

These details fix body proportion, action geometry, and gaze alignment. Skipping them produces the classic AI-look where the eyes are slightly off-axis or the hands interact weirdly with objects.

7. Constraints (what to change vs preserve)

State exclusions and invariants explicitly: "no watermark," "no extra text," "no logos/trademarks," "preserve identity/geometry/layout/brand elements." For edits, use "change only X" + "keep everything else the same" and repeat the preserve list on each iteration to reduce drift. If the edit should be surgical, also say not to alter saturation, contrast, layout, arrows, labels, camera angle, or surrounding objects.

The 5 production-ready patterns

1. Photoreal portrait (LinkedIn headshot, editorial portrait)

Scene: a sunlit modern office, golden hour light through floor-to-ceiling windows
Subject: a 35-year-old woman, dark hair pulled back, wearing a cream blazer
Details: laughing, holding a coffee cup, soft smile, looking at the camera
Use case: professional LinkedIn headshot
Constraints: 85mm portrait lens, shallow depth of field, photorealistic, no watermark

Why it works: every variable is constrained. Scene, subject, action, lens, mood, use case. The model can't generalize because you've set the whole frame.

2. Product shot (ecommerce, packaging, brand asset)

Scene: minimalist white seamless backdrop
Subject: matte black ceramic coffee mug
Details: three-point softbox lighting, soft highlights, 45-degree elevated angle
Use case: ecommerce product page
Constraints: photorealistic, no text, no people, studio photography

Why it works: the lighting language ("three-point softbox," "soft highlights") is studio-photography vocabulary the model recognizes. The 45-degree angle is the ecommerce default for good reason. No people, no text, no distractions.

3. UI mockup (app screens, dashboards, prototypes)

Scene: a high-fidelity tablet screen on a clean white desk
Subject: a recipe app interface called "Cosmic Ramen"
Details: top-left nebula icon, "2300 AD" timestamp, three buttons (ADD, SAVE, COOK) with icons, ingredient list "3x Nebula Dust, 1x Moon Salt, 2x Void Sprouts"
Use case: futuristic app design mockup
Constraints: 2K render, sharp text, no watermark, no extra UI elements

Why it works: GPT Image 2 is one of the strongest models for UI generation because it follows layout and label instructions precisely. Quote the literal text you want, specify the hierarchy, lock the resolution.

4. Text-in-image poster (magazine spread, quote poster, infographic)

Scene: editorial magazine spread, white background
Subject: a typography-forward quote poster
Details: large serif headline "The new API is a brain, not a brush.", small caption "42 | AI TRENDS TODAY Q3 2026", sidebar titled "AT A GLANCE" with bullet points for "Core Strengths" and "Cost Matrix"
Use case: editorial print poster
Constraints: every word must be legible, professional magazine layout, no extra text

Why it works: this is where GPT Image 2 shines. Put literal text in quotes. Specify font style descriptively (serif, sans-serif, monospace, modern, vintage). State the hierarchy (headline, caption, sidebar). At 98.5% typographic accuracy, it's the safest of the four primary models for dense text layouts.

5. Multi-image composite (product on model, character in scene, style transfer)

Scene: an outdoor park setting, soft afternoon light
Subject: a woman from image 1, wearing the dress from image 2
Details: walking on a path, natural pose, soft smile
Use case: ecommerce fashion shoot
Constraints: preserve the woman's face exactly, preserve the dress design exactly, natural lighting, no watermark

Why it works: pass the reference images alongside the prompt. Specify which elements come from which reference ("woman from image 1, dress from image 2"). State the preserve list explicitly. GPT Image 2's edit endpoint preserves identity well when you constrain it.

What GPT Image 2 is actually best at

The model's headline capabilities, from OpenAI's official guide:

High-fidelity photorealism with natural lighting, accurate materials, and rich color
Reliable text rendering with crisp lettering, consistent layout, strong contrast
Robust identity preservation for character consistency across edits and multi-step workflows
Complex structured visuals including infographics, diagrams, multi-panel compositions
Precise style control with minimal prompting
Strong real-world knowledge — accurate depictions of objects, environments, and scenarios

If your use case hits any of these, GPT Image 2 is the right default. If your use case is high-volume social or fast iteration, look at Nano Banana 2 or Grok Imagine first.

The 10 official generation use cases (from OpenAI cookbook)

For reference — these are the use cases OpenAI's own cookbook documents as working with GPT Image 2:

Infographics
Translation in images
Photorealistic images
World knowledge (depicting real-world objects/environments)
Logo generation (concepts, not exact brand marks)
Ads generation
Story-to-comic strip
UI mockups
Scientific / educational visuals
Slides, diagrams, charts, productivity images

The 9 official edit use cases (from OpenAI cookbook)

Style transfer
Virtual clothing try-on
Drawing → image (rendering a sketch)
Product mockups (clean background + label integrity)
Marketing creatives with real text in-image
Lighting and weather transformation
Object removal
Insert the person into a scene
Multi-image referencing and compositing

Size and resolution rules

GPT Image 2 has hard constraints. Get these wrong and the API rejects the call or produces degraded output.

Max edge length: < 3840px
Both edges: multiples of 16
Aspect ratio: long edge / short edge ≤ 3:1
Total pixel range: 655,360 to 8,294,400
Reliability threshold: above 2560×1440 (3,686,400 total pixels), treat output as experimental
Aspect ratio presets: 1024×1024, 1024×1536, 1536×1024, or pass any custom size that fits the constraints

Practical sizes:

1024×1024 — general-purpose default
1024×1536 — portrait
1536×1024 — landscape
2560×1440 — widescreen (recommended upper reliability boundary)
3824×2144 — experimental 4K (max-edge rule rounds down)

For 4K output, use the 4K beta, but expect more variable results.

Quality parameter tradeoffs

quality is the single most important parameter for cost and latency:

quality="low" — fastest, cheapest, sufficient for many use cases including iteration and high-volume generation
quality="medium" — middle ground, good for most production assets
quality="high" — slowest, most expensive, best for hero assets and dense text

The recommended workflow: start at low for iteration, switch to high for the final. You can lock the final at the same prompt — no need to retune.

Common mistakes that silently break results

Using "DALL-E" or "ChatGPT" as model IDs — those are retired. Use gpt-image-2 (or chatgpt-image-latest for auto-updates).
Skipping the word "photorealistic" — leaving it off produces a more "AI" look for portraits. The word engages the model's photoreal mode.
Exceeding the 2560×1440 reliability boundary — going above 2K produces more variable results. Stay at or below 2K for production.
Using decorative quality words for functional tasks — "Beautiful" doesn't make a label readable. Use "sharp label text," "clean kerning," "readable from a distance" for text.
Skipping aspect ratio — a good square image may fail as a vertical ad or video thumbnail. Always specify.
Treating logos like generic text — GPT Image 2 can design logo concepts, but exact brand marks should be composited from approved brand assets, not generated.
Skipping the preserve list in edits — without "keep everything else the same," the model may also change lighting, color, or layout.
Vague scale and framing language — "a person" produces generic results. "A 30-year-old woman, full body visible, feet included, child-sized relative to table" produces precise results.
Long single-paragraph prompts — break complex requests into labeled segments. The model parses structured prompts better than walls of text.

Quick start: a 3-section prompt that works for most GPT Image 2 use cases

Scene: [where this happens, with lighting/mood]
Subject: [who/what is in the image, with specific details]
Constraints: [style cue like "photorealistic" or "3D render", output constraints like "no watermark, no extra text"]

Example: "Scene: a sunlit modern office, golden hour light. Subject: a 35-year-old woman, dark hair pulled back, cream blazer, laughing at the camera, holding a coffee cup. Constraints: photorealistic, 85mm portrait lens, no watermark, no extra text, 1024×1024."

Three sections. Scene, subject, constraints. For UI/text-heavy work, add "Use case: [context]" as a 4th section.

The summary

GPT Image 2's prompt engineering isn't about clever keywords. It's about:

Order: background/scene → subject → key details → constraints
Use case: state what the image is for
Constraints: explicit "no" list, especially for text and branding
Quality parameter: low for iteration, high for finals
Size discipline: stay at or below 2560×1440 for production reliability
Edit precision: "change only X" + "keep everything else the same" + repeat

Once you internalize those, the prompt patterns in this guide are just shortcuts. The underlying fundamentals do the work.

Want to see how this stacks up against the other primary models? Compare GPT Image 2 vs Nano Banana 2 for editing and text-rendering differences, or browse the full prompt library for ready-to-use templates across all 4 primary models.