ChatGPT Image Prompts: How to Write Ones That Actually Work

The best ChatGPT image prompts describe the scene first, then the subject, then the key details, then any constraints, and they put any text you want in the image inside quotes. For edits, say exactly what to change and repeat what must stay the same every time; that one habit fixes most “it ruined my picture” complaints.

Below is a practical breakdown based on OpenAI’s own image generation prompting guide in the OpenAI Cookbook, plus launch coverage of the current model. Everything here reflects the situation as of November 2026.

What model you are prompting

ChatGPT’s image generation got a major update on April 21, 2026, when OpenAI launched ChatGPT Images 2.0, according to Axios. The underlying model is available to developers as gpt-image-2, and OpenAI’s cookbook recommends it as the default model.

Launch coverage focused on much better text rendering and on handling more complex requests, including infographics, slides, maps and multilingual text (VentureBeat). Axios reports that all users get the standard model, while a thinking mode that reasons before it draws is limited to paid plans. Reports differ on extra Pro-only features, so treat anything beyond that as unconfirmed.

Why this matters for prompting: this model follows long, detailed instructions well. You get more out of being specific than out of tricks or “magic words.”

The basic prompt structure

OpenAI’s guide recommends a consistent order:

1. Scene or background. Where are we? “A rainy Tokyo side street at night.”
2. Subject. Who or what is the focus? “An elderly street-food vendor behind a steaming cart.”
3. Key details. Materials, textures, lighting, framing, mood. “Neon reflections on wet pavement, warm lantern light on his face, eye-level medium shot.”
4. Constraints. What must or must not appear. “No text, no watermark, no logos.”

Two extra tips from the guide make a noticeable difference:

• State the intended use. Saying “for a magazine ad,” “as a UI mockup” or “as a classroom infographic” tells the model how polished and what kind of layout you expect.
• Break long prompts into labeled lines. For complex requests, short segments like Scene:, Subject:, Style: and Text: work better than one dense paragraph.

The format itself is flexible. The guide says minimal prompts, paragraphs, JSON-like structures and tag-based prompts all work; pick whatever you can read and reuse easily.

Getting the look right

Be concrete about style and materials

Name the medium (photo, watercolor, 3D render, pencil sketch), the materials and the textures. Add quality cues such as “film grain” or “macro detail” only when you actually want that effect.

Photorealism

OpenAI’s guide says to use the word “photorealistic” directly. Phrases such as “real photograph” or “iPhone photo” also help. You can mention camera and lens details (“shot on 35mm film, 50mm lens”), but the guide notes these are interpreted loosely. Use them to set the general look and framing, not to simulate a specific camera exactly.

Composition and people

Spell out framing, viewpoint and lighting. You can also place elements: “logo in the top-right corner.” For people, describe scale, how much of the body is in frame, where they are looking and what they are holding. “Full body visible, feet included” prevents a lot of awkward crops.

Text inside images

Text rendering is where Images 2.0 improved most, but your prompt still matters. From OpenAI’s guide:

• Put the exact words in quotes or ALL CAPS.
• Describe font style, size, color and placement.
• Spell tricky words letter by letter, especially brand names.
• Ask for the text to appear once, verbatim, with no extra characters.

Example:

A minimalist highway billboard at dusk. Large bold sans-serif text reads “FRESH AND CLEAN” in white, centered. The text appears exactly once, verbatim, no extra characters, perfectly legible.

Editing an image without wrecking it

Most frustration with ChatGPT images comes from edits that change more than you asked. OpenAI’s guide gives a clear recipe:

1. Name the one change. “Change only the jacket color to dark green.”
2. Lock everything else. Add “keep everything else the same.”
3. List what must be preserved. For example: “Do not change her face, facial features, skin tone, pose, background, lighting or camera angle.”
4. Repeat the preserve list every round. The guide specifically recommends this to reduce drift over multiple edits.
5. Make one change per follow-up. Small, single-change requests (“make the lighting warmer,” “restore the original background”) are more reliable than bundling five fixes.

For surgical edits on diagrams or designs, also protect color saturation, contrast, layout, arrows and labels. If the model drifts after several turns, “same style as before” can help, but the guide advises re-stating the critical details outright.

Working with several images

If you upload more than one image, refer to each one by number and a short description (“Image 1: the woman on the bench; Image 2: the brown dog”), then explain how they should combine: “Place the dog from Image 2 sitting beside the woman in Image 1, matching the lighting.”

Keeping a character consistent

For a series, such as a children’s book, the guide’s approach is to create one “anchor” image of the character first, then use that image as the input for each new scene while repeating the character’s key features.

Copy-ready prompt templates

Replace the bracketed parts. Keep the constraint sentence at the end; it is the part people most often forget.

Things that work in the API but not as simply in ChatGPT

OpenAI’s guide also covers settings that developers set as API parameters, such as exact output sizes, a quality level (low, medium or high), the number of images and transparent backgrounds. It recommends medium or high quality for small text, dense diagrams and close-up portraits. In the ChatGPT app you generally describe what you want in words instead, for example “portrait orientation” or “on a transparent background,” and how closely those requests are honored can vary.

If you need guaranteed dimensions or batch output, the API is the official route; for everyday prompting, the structure above is what moves the needle.

Quick checklist

Before you hit enter, check that your prompt has:

• A scene, a subject and specific details, in that order
• The intended use (“for a poster,” “for a slide”)
• Any text in quotes, with placement and style
• Explicit exclusions (“no watermark, no extra text”)
• For edits: one change plus a preserve list

Features and plan limits for ChatGPT images change frequently, so if something here behaves differently in your account, check OpenAI’s current documentation.