GPT Image 2 Prompting Guide: 12 Techniques From OpenAI's Cookbook

GPT Image 2 launched on April 21, 2026 as OpenAI's state-of-the-art image generation model — the recommended default for any new image workflow, with stronger generation quality, near-perfect multilingual text rendering, identity-sensitive editing, and flexible sizing up to a 4K hero. Alongside the model, OpenAI shipped an official prompting guide in the OpenAI Cookbook documenting the patterns that show up repeatedly in production use.

This article condenses that guide into twelve technique-driven examples you can copy directly. Every pattern below is grounded in OpenAI's public documentation: the prompt anatomy, the four supported prompt formats, the rules for in-image text, the constraints language for surgical edits, the multi-image referencing convention, and the iterative-refinement approach. Use it as a working reference, not a one-shot read.

OpenAI's cookbook recommends a consistent structural order for every prompt: background and scene first, subject second, key visual details third, and constraints last. The model performs noticeably better when the prompt follows this order than when the same information is reshuffled into a free-form sentence, because the early tokens set the rendering "mode" — ad, UI mock, infographic, photograph — and later tokens then flesh out the detail.

The official guide also recommends explicitly stating the intended use of the asset ("a marketing ad for…", "a UI mockup for…", "an educational diagram for…"). That single line changes the level of polish, the typographic choices, and the composition the model picks by default.

GPT Image 2 was trained to accept multiple prompt formats. The cookbook explicitly notes that minimal prompts, descriptive paragraphs, JSON-like structures, instruction-style prompts, and tag-based prompts can all work well, "as long as the intent and constraints are clear." For production systems the recommendation is to prioritize a skimmable template over clever syntax — pick one format and stay with it across your codebase.

GPT Image 2 has an explicit photorealistic rendering mode, and the most reliable way to engage it is to use the word "photorealistic" directly in the prompt. The cookbook also flags adjacent phrases — "real photograph," "taken on a real camera," "professional photography," "iPhone photo" — as effective alternatives.

Detailed camera specs (specific lens model, sensor size, ISO) are interpreted loosely; treat them as look-and-feel cues rather than physical simulation. The bigger lever for honest realism is naming the imperfections: pores, fine lines, asymmetry, fabric wear, available light. Anti-cues like "no glamorization" and "no heavy retouching" push the model away from the generic AI portrait look.

The cookbook is direct about composition: specify framing and viewpoint (close-up, wide, top-down), perspective and angle (eye-level, low-angle), and lighting and mood (soft diffuse, golden hour, high-contrast). When layout matters, name placement explicitly — "logo top-right," "subject centered with negative space on left." For wide, low-light, rain, or neon scenes, add scale, atmosphere, and color descriptors so the model does not trade mood for surface realism.

The example below combines all three composition layers in a single prompt: a defined framing (medium-wide), a defined viewpoint (low-angle), and a defined lighting style (golden hour key + soft fill).

In-image text is GPT Image 2's biggest single capability gap from earlier models. The cookbook is specific about how to invoke it reliably: put literal copy in quotes or ALL CAPS, and specify typography details (font style, size, color, placement) as constraints. For tricky words — brand names, uncommon spellings, mixed-case product names — spell them letter-by-letter inside the prompt to boost character accuracy.

The single most useful add-on phrase is "EXACT, verbatim, no extra characters" applied to the literal copy. Without it, the model occasionally paraphrases or appends decorative additions; with it, the rendered text matches the prompt exactly. Use medium or high quality for small text, dense information panels, and multi-font layouts.

GPT Image 2's text rendering pipeline lays glyphs as vector shapes before rasterizing them into the scene, which is why English, Japanese, Korean, Arabic, Chinese, Turkish, and Hebrew all render correctly on the first attempt in most cases. The prompting rules are identical to Latin text: quote the copy, name the typeface family, call out placement, and invoke "EXACT, verbatim" if the script is sensitive (kanji stroke order, Arabic ligatures).

For mixed-script layouts, list each text element separately with its script, position, and typeface. The model handles the script-pairing and kerning automatically as long as the elements are unambiguous.

The cookbook calls this section out specifically: for people in scenes, describe scale, body framing, gaze, and object interactions. Generic phrases like "a person doing X" tend to drift on body proportion and limb articulation. Concrete phrases — "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" — pin down the geometry.

This is the difference between "two friends laughing" rendering as a stiff promo shot versus rendering as a believable in-the-moment photograph.

Constraints are the most underused part of GPT Image 2 prompting. The cookbook recommends stating exclusions and invariants explicitly: "no watermark," "no extra text," "no logos/trademarks," "preserve identity/geometry/layout/brand elements." For brand-mark and packaging work, the recommended template is to require an "original, non-infringing" mark, which steers the model away from regurgitating existing trademarked logos and toward a clean, original silhouette.

The other lever for logo work is asking explicitly for clean vector-like shapes, a strong silhouette, balanced negative space, and scalability across sizes — a one-line creative brief in prompt form.

Style transfer is one of the cleanest GPT Image 2 edit patterns. You pass a reference image and a text instruction; the model preserves the visual language of the reference (palette, brushwork, line weight, texture, film grain) while replacing the subject. The cookbook is explicit about the prompting recipe: describe what must stay consistent (style cues) and what must change (new content), and add hard constraints — background, framing, "no extra elements" — to prevent peripheral drift.

The key phrasing is "use the same style from the input image and generate…" rather than "make this look like…", which the model sometimes interprets as a loose creative riff.

When you pass multiple input images to the /v1/images/edits endpoint, the cookbook's recommendation is to reference each input by index and description ("Image 1: product photo… Image 2: style reference… Image 3: target scene…") and describe how they interact ("apply Image 2's style to Image 1," "place the dog from Image 2 next to the woman in Image 1"). Be explicit about which elements move where.

This indexing convention is what unlocks the "insert this object/person into that scene" workflow without re-generating the whole frame. It also lets you combine three or four references in one call — for example, a person in image 1 dressed in garments from images 2, 3, and 4 — with the model treating each input as a distinct asset rather than blending them into one composite reference.

GPT Image 2 supports targeted edits without masks, but the prompt has to be tight to avoid drift. The cookbook's phrase-of-record is: "change only X" + "keep everything else the same" + repeat the preserve list on each iteration. For genuinely surgical edits — interior design swaps, object removals, label tweaks — also explicitly say not to alter saturation, contrast, layout, arrows, labels, camera angle, or surrounding objects.

The restated preserve list is what differentiates a clean first-pass edit from a near-miss that requires three retries. Repetition is intentional: the model treats both the change instruction and the preserve list as constraints, and listing the preserve elements twice raises the weight on each one.

The cookbook explicitly warns against overloaded prompts: "Long prompts can work well, but debugging is easier when you start with a clean base prompt and refine with small, single-change follow-ups." The recommended pattern is to ship a clean first prompt at medium quality, then iterate with phrases like "make the lighting warmer," "remove the extra tree," "restore the original background." Use references like "same style as before" or "the subject" to leverage context — but re-specify any critical detail if it begins to drift.

This is the inverse of the design-tool reflex, where each pass adds more constraints. With GPT Image 2, each refinement pass should remove the noise from the prior pass and change at most one or two things. Multi-region simultaneous edits (three or more independent changes in a single call) are explicitly flagged as needing 2–3 iterations for a clean result.

A capability the cookbook calls out specifically: GPT Image 2 pairs strong reasoning with world knowledge, so you can prompt with situational cues — a date, a place, a notable event — and let the model infer the visual context without spelling it out. The canonical example in the official guide is "Bethel, New York on August 16, 1969," which the model correctly infers as Woodstock and renders with period-accurate clothing, staging, and environment.

This works for any well-documented event, place, or cultural moment within the model's knowledge cutoff (December 2025). For subjects that emerged after that date, the inference fails silently — the model will produce a plausible image that is factually wrong. For post-cutoff subjects, switch to providing a reference image instead of relying on world knowledge.

For productivity assets — slides, workflow diagrams, charts, infographics — the cookbook recommends writing the prompt like an artifact spec rather than an illustration request. Name the exact deliverable ("one pitch-deck slide titled 'Market Opportunity'"), define the canvas and hierarchy, supply the actual text and numbers, and describe the visual language. Include practical constraints: readable typography, polished spacing, no decorative clutter, no generic stock-photo treatment.

The model will not invent figures — it renders the values you supply as written. For data-heavy infographics, this means the burden of accuracy stays with you (or with a thinking-mode web search), not with the model. Use a landscape size for deck-style outputs and quality="high" when the image contains small text, legends, axes, or footnotes.

GPT Image 2 exposes three quality levels — low, medium, and high — and accepts any size that meets four constraints: maximum edge length less than 3840px, both edges multiples of 16, long-to-short edge ratio not greater than 3:1, and total pixels between 655,360 and 8,294,400. Above 2560×1440 (2K), the cookbook flags results as experimental. Square images are generally fastest to generate.

The cookbook's top-line recommendation is to start at quality="low" for latency-sensitive or high-volume use cases and only escalate when fidelity is genuinely the bottleneck. Low is good enough for a wider range of work than most teams expect.

If you take one thing from this guide, take the template. It follows the cookbook's recommended structural order and works as a copy-paste starting point for almost every use case in this article:

Intended use → Scene / background → Subject (with scale, pose, gaze) → Key visual details (materials, textures) → Composition (framing, viewpoint, focal length) → Lighting (direction, quality) → Literal in-image text in quotes (EXACT, verbatim) → Constraints (preserve list / no watermark / no extra text).

Start at quality="medium" and a 1024×1024 square, run two generations to calibrate the prompt, then move to high quality and a non-square aspect ratio for the final asset. For refinements, edit the existing image with a natural-language instruction rather than regenerating from scratch — the latter is the single biggest source of brand drift in production work.