GPT Image for Design Work: Layouts, Text and Edits

The myth: an image model is an illustrator, so a design brief like "make me a poster" will come back as a finished, correctly spelled, pixel-placed layout. If it doesn't, the model is bad.

The correction: GPT Image is better treated as a designer who follows a written spec. It's strong at following detailed instructions, rendering text and editing an existing picture. But OpenAI's own docs are blunt that it can still struggle with precise text placement and with exact positioning in layout-sensitive work. You get the best results when you write the brief like a spec, check the output like a proofreader, and fix problems with edits instead of rerolling from scratch.

This post covers how to do that for layouts, text and edits. Everything about the model below comes from OpenAI's documentation, and I've linked each page at the end.

Write the prompt as a spec, not a mood

OpenAI's prompting guide recommends writing design prompts like artifact specs. Name the deliverable (slide, diagram, poster), describe the canvas and hierarchy, supply the real text or data, then describe the visual style.

Here's a pattern that works for a social promo:

Instagram promo graphic, 4:5 portrait. Top third: headline. Middle: product photo of a ceramic mug on a wooden table. Bottom strip: a call-to-action button. Flat pastel palette, lots of whitespace, bold sans-serif type.

Notice what's in it: the format, where each element sits, and the style. What's missing is vague adjectives like "stunning" or "professional". Those add nothing.

If you're unsure which shape to ask for, the aspect ratio guide covers the common social sizes. For general prompt ordering, see AI Image Prompt Structure: Subject, Setting, Light, Lens.

Text: quote it, constrain it, check it

OpenAI's guidance for text is simple and practical:

• Put the literal words in quotes (or capitals).
• Say that this is the only text allowed.
• Specify the typography: style, colour, size, placement.
• For unusual words or names, spell them out letter by letter.
• Use medium or high quality for small text.

Their example wording looks like this:

Include ONLY this text (verbatim, no extra characters): "Fresh and clean". Typography: bold sans-serif, high contrast, centered.

That "ONLY this text" line matters. Without it, models like to invent a tagline or a fake URL in the corner.

The docs also say that even with improved rendering, the model "can still struggle with precise text placement and clarity." So expect to proofread every output. Brand names, prices and phone numbers are where a single wrong character costs you. For a deeper look at text across models, see How to Get Readable Text Inside AI Images.

A practical rule for text-heavy designs

Keep the text short. A headline and a button label usually come out clean. A paragraph, a legal footer and a table of prices is asking for trouble. If you need dense copy, generate the artwork with a clear empty area and add the real text in your design tool.

Layouts: ask for structure, then verify it

For dense layouts (slides, infographics, charts with legends and axes), OpenAI suggests the high quality setting. A slide prompt in their guide follows this shape:

Create a pitch-deck slide titled "Market Opportunity" with a TAM/SAM/SOM diagram in muted colors, specific numbers, and a bar chart showing growth.

Two cautions.

First, the docs list composition control as a known limit: the model may have difficulty placing elements precisely in structured or layout-sensitive compositions. If your grid needs to line up to the pixel, treat the output as a draft or background.

Second, any chart with real numbers needs checking. The transparent-assets cookbook says to verify that labels, values and proportions match the source data before using generated charts. A chart that looks right can still be wrong.

Dimensions matter too. OpenAI lists 1024x1024, 1536x1024 and 1024x1536 as the common sizes. For custom sizes on the newer model, both edges must be multiples of 16, the ratio can't exceed 3:1, and the longest edge must stay under 3840 pixels. It also says to treat anything above 2560x1440 as experimental. (Details may change, so check the image generation guide if you're working at the edges.)

Edits: say what changes and what stays

Editing is where GPT Image earns its place in a design workflow. The core principle in OpenAI's guide is to state exclusions and invariants explicitly, and to repeat the "keep" list every time you iterate, because drift builds up over rounds.

The pattern is "change only X, keep everything else the same":

Change only the background color to deep green. Keep the product, its label text, the shadow and the framing exactly the same.

For clothing swaps, the guide's example lists everything to preserve: face, skin tone, pose, expression, hairstyle and proportions. Be that specific.

Combining several images

With multiple inputs, refer to each by index and description, and say how they interact:

Image 1: our storefront photo. Image 2: the logo on a white background. Place the logo from Image 2 on the shop sign in Image 1, matching the lighting.

Masks are guidance, not scissors

If you use a mask to limit an edit, the docs say masking with GPT Image is entirely prompt-based. The model uses the mask as guidance but may not follow its exact shape. So don't rely on a mask to protect a logo or a face. Say in words what must not change, and check the edges.

One more detail from the prompting guide: input fidelity is listed as disabled on the newest model and available on gpt-image-1.5 and earlier. If an edit keeps altering things you wanted untouched, that setting and the model version are worth looking at.

Transparent assets for stickers, icons and overlays

OpenAI has a cookbook workflow for transparent PNG assets. The key points:

1. Prompt for the isolated subject only. Don't describe a backdrop.
2. Request transparency explicitly: background="transparent" with PNG or WebP output (not JPEG).
3. Use high quality for fine edges and any text.
4. Open the file and confirm it has a real alpha channel, not a white box.

The caveat is worth underlining: if your prompt describes a background, scene or colour, the model may draw that instead of giving you transparency. The guide also flags transparent backgrounds as a preview feature, so don't build a client pipeline on them without testing.

A workflow that saves rerolls

1. Draft cheap. Test the composition at low quality before you spend on a finished one.
2. Lock the layout first. Get the placement of the big blocks right before worrying about the words.
3. Add the text with the quoted, "only this text" pattern.
4. Fix with one-change edits. One edit per round, with the preserve list repeated.
5. Proofread at 100%. Zoom in on every character.
6. Finish in a design tool. Final text, alignment and brand colours are often faster there.

Where it falls short

• Run-to-run variation. The same prompt can give a clean headline once and a typo the next time.
• Consistency. OpenAI's docs note it can struggle to keep recurring characters or brand elements the same across generations. For characters, see Keep a Consistent AI Character Across Images and Video.
• Speed. Complex prompts can take up to two minutes, according to the docs.
• Exact placement. Pixel-perfect grids aren't realistic.

Trying it in Mediafy Studio

GPT Image is one of the image models you can pick in the Mediafy Studio. Choose Image, select the model, write your spec, and check the credit cost on the Generate button before you run it. If a short brief is all you have, the Enhance button can rewrite it into a fuller prompt. And if you want to compare how another model handles the same layout, you can switch models and keep the same prompt. Results will vary, so judge them on your own designs.

Want to test a layout prompt yourself? Sign in to Mediafy Studio.

Sources

• OpenAI, "Image generation" (API guide): https://developers.openai.com/api/docs/guides/image-generation
• OpenAI Cookbook, "GPT Image Generation Models Prompting Guide": https://developers.openai.com/cookbook/examples/multimodal/image-gen-models-prompting-guide
• OpenAI Cookbook, "Generate Transparent Image Assets for Campaigns and Presentations": https://developers.openai.com/cookbook/examples/multimodal/transparent-image-assets-for-campaigns-and-presentations