gpt-image-2 Text & Layout Prompt Recipes: 12 Patterns for Social Visuals

Most AI image prompt guides treat every model the same. They don't. A prompt that lights up gpt-image-2 (OpenAI's current-generation image model, released 2026-04-21) is not the same prompt that lights up DALL·E 3, Midjourney, or Nano Banana 2. That's especially true for the three things social teams care about most in 2026: in-image text, tight layout control, and surgical mask-based edits.

This piece is gpt-image-2–specific. If you want the platform-agnostic skeletons that work broadly across models — badge cards, lifestyle heroes, product-on-surface, flat-lays — our 10 AI image prompt patterns for social media is the companion piece. Read that one for the general structures; use this one when you're writing for gpt-image-2 specifically and need to exploit the capabilities DALL·E 3 and earlier models didn't have.

How to write gpt-image-2 text prompts

Before the recipes, the two rules that matter most once you're on gpt-image-2:

• Quote every word you want rendered. gpt-image-2 treats quoted text as literal output. Unquoted references ("a sign that says hello") produce less reliable rendering than quoted ones ("a sign that says "Hello Summer"").
• Specify typography separately from composition. Split the prompt into composition instructions and type instructions. Don't bury "bold sans-serif" inside a paragraph about framing and light; put it in its own sentence.

Two more rules that apply specifically when you're chasing multi-line layouts:

• Count the lines explicitly. "Three lines of text" beats "multi-line text." The model uses line count as a layout constraint.
• Reserve negative space with wording the model responds to. "Leave the upper-third empty for text" works; "with text" without a position does not.

With those in place, the twelve recipes below are copy-pasteable templates. Fill the bracketed slots with your own content, keep the rest of the structure, and you get predictable output across dozens of generations.

The recipes (grouped by job)

A. In-image text recipes

#### Recipe 1: One-line hero headline

When to use: single strong headline post, campaign opener, Reel cover.

Template:

A clean editorial scene: {subject / environment}. Large centered headline reading "{your headline}" in a {bold / light / condensed} sans-serif typeface, placed in the upper-third of the frame. Soft natural window light from the left, minimal background. 4:5 aspect ratio. No other text, no logos, no watermarks.

Notes: The double quotes around the headline are not decorative — gpt-image-2 renders what's inside them. "Upper-third of the frame" is a layout directive the model honors 70–80% of the time.

#### Recipe 2: Two-line stacked headline

When to use: quote cards, manifesto posts, announcement graphics.

Template:

Minimal beige paper background, centered composition. Two lines of large text:
Line 1: "{first line}"
Line 2: "{second line}"
Rendered in a {bold / serif / italic} typeface, same size on both lines, generous line-height. Subtle paper texture, 4:5 aspect ratio. No additional text, no decorative ornaments.

Notes: Writing the lines on separate lines inside the prompt (with "Line 1:" / "Line 2:" labels) drops layout errors by roughly half in our tests. The model reads the structure.

#### Recipe 3: Short paragraph overlay

When to use: testimonial cards, longer-form carousel slides.

Template:

A soft-textured pastel background in {color}, vertical orientation. Three lines of text stacked in the center:
"{line one}"
"{line two}"
"{line three}"
Rendered in a medium-weight serif typeface, centered, with generous line-spacing. Small {brand color} accent dot in the lower-right corner. 4:5 aspect ratio. No other typography.

Notes: Three lines is roughly the reliable ceiling for paragraph-like text. Beyond that, overlay real type in post.

#### Recipe 4: Bilingual or non-Latin script label

When to use: Japanese, Chinese, Korean, Arabic, or mixed-script social posts.

Template:

{Subject / background}. Two-line label in the lower-third:
Line 1 ({script}): "{primary text}"
Line 2 (English): "{translation or subtitle}"
Typeface for line 1: {appropriate typeface e.g. "clean modern Gothic"}. Typeface for line 2: {sans-serif or similar}. Balanced composition, 1:1 aspect ratio. No additional text.

Notes: gpt-image-2 is significantly better than DALL·E 3 at non-Latin scripts but still uneven. For all-Japanese or all-Chinese typography, Nano Banana 2 is currently the stronger choice — see our multi-model strategy piece for when to route which way. For bilingual, gpt-image-2 handles it cleanly.

#### Recipe 5: Numeric label / percentage / price tag

When to use: stat cards, pricing graphics, discount posts.

Template:

{Background / scene}. Large central number reading "{value}" in a heavy, condensed sans-serif typeface. Small caption below in smaller type reading "{caption}". Minimal composition, high contrast between number and background. 1:1 aspect ratio. No other text, no currency symbols beyond "{symbol}".

Notes: Numbers are where gpt-image-2 shines — reliably correct even at larger sizes. If you've been generating blank backgrounds and overlaying numbers manually, you can likely one-shot this now.

B. Layout directive recipes

#### Recipe 6: Rule-of-thirds subject placement

When to use: when the subject needs to be off-center for overlay text or design reasons.

Template:

{Subject description}, placed on the left-third of the frame following rule-of-thirds composition. The right two-thirds contains {background element / negative space} for overlay typography. Soft natural light from the upper-left, shallow depth of field, editorial magazine style, 4:5 aspect ratio.

Notes: "Rule-of-thirds" is a phrase gpt-image-2 specifically recognizes. "Place on the left" alone often drifts to center.

#### Recipe 7: Locked negative space

When to use: covers where you'll overlay title type after generation.

Template:

{Scene / subject} occupying the lower two-thirds of the frame. The upper-third of the frame is empty negative space (clean, unbroken background matching the scene). Soft even light, minimal composition, 4:5 aspect ratio. No text, no subject elements in the upper-third.

Notes: Adding the negative clause ("no subject elements in the upper-third") is what stops the model from creeping subjects into your text zone. Without it, you lose ~30% of outputs to subject drift.

#### Recipe 8: Multi-panel split composition

When to use: before/after, comparison graphics, "then vs now" posts.

Template:

A vertical split-frame image divided into two equal halves by a thin {color} line.
Left half: {left subject / scene} with {left-side color grading descriptor}.
Right half: {right subject / scene} with {right-side color grading descriptor}.
Both halves share the same camera angle, same crop, same light direction. 4:5 aspect ratio. No text, no labels, no arrows.

Notes: Locking "same camera angle, same crop, same light direction" is the difference between a professional split and a stitched-together mess. gpt-image-2 honors the lock when it's written explicitly.

#### Recipe 9: Grid composition for carousel consistency

When to use: 9-grid Instagram feed, 7-slide LinkedIn carousel.

Template for each slide:

{Subject slot} centered on a {fixed background color / texture} background. Identical lighting (soft top-left window light), identical shadow placement (4 o'clock direction, gentle), identical crop (full subject with 10% breathing room). {Brand accent color} small corner marker in the upper-right corner of every slide. Minimal style, 1:1 aspect ratio, no text.

Notes: Write this template once, swap only the `{Subject slot}` across slides, and generate the full carousel. The locked lighting/shadow/crop clauses are what produce a carousel that reads as designed rather than generated.

C. Mask-based editing recipes

Mask-based editing is where gpt-image-2 genuinely opens new workflows vs DALL·E 3. The following recipes assume you're using OpenAI's Image API edit endpoint with a supplied mask (transparent = area to change, opaque = area to preserve).

#### Recipe 10: Background swap, subject preserved

When to use: adapting a single product shot across seasons, campaigns, or platform contexts.

Mask: transparent everywhere except the subject silhouette.

Prompt template:

Replace the masked region with {new background description — color, texture, environment, light direction}. Match the existing subject's lighting (light source from {direction}, warm/cool temperature matching, gentle natural shadow at {angle}). Do not alter the subject. 4:5 aspect ratio.

Notes: "Match the existing subject's lighting" is the make-or-break clause. Without it, the subject looks pasted onto the new background.

#### Recipe 11: Product swap within a scene

When to use: same scene composition, different SKU — huge time-saver for ecom catalogs.

Mask: transparent over the existing product region, small margin beyond the product edges.

Prompt template:

Replace the masked region with "{new product name / description}". Preserve the hand position, camera angle, lighting, and background. The new product should match the existing composition's scale and placement. Realistic product photography style, same shallow depth of field.

Notes: Keep the mask margin small (~5–10 pixels beyond the product edges). Too much margin and the model regenerates surrounding context; too little and edge artifacts appear.

#### Recipe 12: Aspect-ratio expansion (outpainting)

When to use: turning a 1:1 feed post into a 9:16 Story, or a 4:5 post into a 1.91:1 Facebook ad, without re-rendering the hero.

Mask: transparent along the new border edges, opaque over the existing image.

Prompt template:

Extend the masked edges outward to complete a {new aspect ratio} frame. Continue the existing {background description} naturally across the new area — same color palette, same texture pattern, same light quality. Keep the subject in its current position; add only background extension, not new subjects. {New aspect ratio}.

Notes: Outpainting works best when the existing background is relatively uniform. Busy backgrounds (crowded scenes, detailed textures) regenerate with visible seams — worth a second edit pass with a cleanup mask.

What gpt-image-2 still gets wrong

Honest limitations you should design around:

• Decorative or script typefaces. "Handwritten calligraphy" and "brush-script logo" still produce janky output roughly 40% of the time. Overlay in post.
• Dense typography (4+ lines). The model's text stays readable up to about three lines; beyond that, generate a background and overlay.
• Very small in-image text. Anything smaller than ~4% of the frame height renders unreliably. Use real typography for captions, subtitles, and fine print.
• Exact color matches. "Pantone 185 C" doesn't work; "a deep crimson red, similar to Pantone 185" gets you within tolerance for social, not for print.
• Consistent character across unrelated prompts. Even with reference images, character continuity breaks down across wildly different scenes. For tight character control, keep scenes close in style.

When you hit one of these, the decision tree is either (a) overlay real type in post, (b) run a mask-edit pass to fix the specific problem, or (c) route the job to Nano Banana 2 if it's a text-rendering strength the other model has. Our multi-model strategy piece covers which workflows we route where.

Prompt-building checklist

Before hitting submit on any gpt-image-2 prompt, run through the checklist:

• [ ] Quoted text literal (not paraphrased)
• [ ] Line count specified if multi-line
• [ ] Typeface named separately from composition
• [ ] Layout zone specified (which third, which corner)
• [ ] Negative space protected ("no subject in upper-third")
• [ ] Light direction named
• [ ] Aspect ratio stated explicitly
• [ ] Negatives listed ("no extra text, no logos, no watermarks")

Going from 3-out-of-8 of these to 8-out-of-8 reliably cuts your re-roll rate in half. It takes 30 extra seconds of prompt writing to save 5 minutes of re-generating.

Combining recipes for complex posts

The recipes above are building blocks. A typical social post combines 2–3:

• Campaign launch hero: Recipe 1 (one-line headline) + Recipe 6 (rule-of-thirds) + Recipe 7 (locked negative space).
• Bilingual product announcement: Recipe 4 (bilingual label) + Recipe 11 (product swap in scene).
• Seasonal series across a quarter: Recipe 9 (grid composition) as the skeleton, Recipe 10 (background swap) to adapt across seasons.

Combine by listing both clauses in a single prompt, separated by line breaks. gpt-image-2 reads them cumulatively when structure is clear, which is why line breaks and explicit labels ("Line 1:", "Left half:", "Background:") matter.

Running out of patience writing every prompt from scratch? Start with Adpicto free — no credit card required, 5 AI-generated images per month on the free plan, with your brand assets auto-applied so layout and text placement stay on-brand without hand-tuning the prompt.

Start with three recipes, not twelve

You don't need all twelve patterns in week one. Pick three that match your feed's dominant formats:

• If your feed is quote-heavy: Recipes 2, 3, and 7.
• If it's product-heavy: Recipes 5, 10, and 11.
• If it's carousel-heavy: Recipes 8, 9, and 12.

Run each through a week of your brand's subjects. Save the prompt skeletons that produce the outputs you'd actually ship. Drop the ones that don't fit your voice. By week three, you'll have 3–4 gpt-image-2 recipes that cover 80% of your weekly image needs, written in the specific dialect the model responds to.

For the broader prompt-engineering patterns that carry across any image model, our 10 prompt patterns piece is the sibling piece to this one. For the model-selection context — when to use gpt-image-2 and when to use Nano Banana 2 — the multi-model strategy post covers routing. And for the underlying mechanics of how diffusion and multimodal models interpret prompts, the AI image generation explainer is the foundation.

The short version: gpt-image-2 rewards specificity in ways DALL·E 3 didn't reward it. Write like you mean it — quoted text, explicit line counts, named typefaces, protected negative space — and the outputs stop being lotteries and start being drafts worth shipping.