GPT Image and ChatGPT Images Prompt Guide

To write a controllable GPT Image prompt, state the intended use, describe subject and scene, set must-include and must-avoid constraints, quote any literal text that must appear, then choose the surface that matches the job. Treat the first render as a draft you will compare against that brief.

As of the official docs checked for this guide, gpt-image-2 is OpenAI's recommended GPT Image model for new generation and editing work in the API. In ChatGPT, the product experience is branded ChatGPT Images (Images 2.0). Those names are related, not interchangeable: product UI controls are not the same as Image API parameters or Responses API tool options.

OpenAI's image generation guide still lists limitations for GPT Image models, including imperfect text placement or clarity, consistency challenges for recurring characters or brands, and difficulty with precise layout. Do not treat strong prompt adherence as a guarantee.

This page is OpenAI-workflow specific. For portable image craft (subject, style, lighting, composition, exclusions), see the AI Image Prompt Guide. For iteration discipline beyond images, see How to Write Better AI Prompts.

Pick the surface that matches the job. Never assume ChatGPT Images, the Image API, and the Responses API image tool share the same knobs.

• ChatGPT Images (Images in ChatGPT): create or edit in chat or the Images hub; upload a reference; use the selection editor or describe edits in conversation; choose aspect ratio with the picker or in the prompt. Help Center notes that ChatGPT Images can follow instructions to make a background transparent. Selection highlights are not always precise, and edits may extend beyond the selected area. “Images with thinking” is a ChatGPT plan feature label (Plus, Pro, and Business per Help); treat it as a product mode name, not a reason to request hidden chain-of-thought.
• Image API (/v1/images/generations, /v1/images/edits): choose a GPT Image model such as gpt-image-2 directly; best when you need a single generate or edit from one prompt with explicit size, quality, format, and moderation options. See the gpt-image-2 model card.
• Responses API + image_generation tool (image generation tool): choose a supported mainline text model; the tool uses a GPT Image model under the hood. Prefer this for conversational multi-turn editing, File IDs, optional action (auto, generate, or edit), and streaming partials. GPT Image model IDs are not valid as the Responses model field. The tool may expose a revised_prompt; compare the image to your written request, not to a hidden rewrite alone.

Developers may need to complete Organization Verification in the developer console before using GPT Image models, including gpt-image-2. Older GPT Image IDs such as gpt-image-1.5, gpt-image-1-mini, and chatgpt-image-latest are on a documented deprecation path toward gpt-image-2; re-check the deprecations page on the day you ship. Historical DALL·E access in ChatGPT remains available via the DALL·E GPT per Help; it is not the primary teaching target here.

This guide uses SCQIVE: Subject & scene, Constraints, Quality & canvas, Inputs, Verification, and Edit intent. It is a practical checklist created for GetBetterPrompts for OpenAI image workflows, not an industry standard. Related ideas appear in OpenAI's GPT Image prompting cookbook (structure, constraints, iteration) and in API output controls such as size, quality, and format.

Intended use: [ad / UI mock / poster / photo / icon]
Subject & scene: […]
Constraints: [must include / must avoid]
Canvas: [aspect or size] | Quality: [low/medium/high/auto or ChatGPT default]
Inputs: [Image 1 = …; Image 2 = …; mask = …]
Exact text (if any): "[…]"; typography/placement: […]
Edit intent (if revising): Change only [X]. Keep [identity, pose, lighting, layout, …]
Verification: Compare to request; note mismatches; next single change: […]

Skip empty parts. Use Verification for visible checks against your brief. Do not ask the model to reveal private chain-of-thought or hidden reasoning as proof that an image is correct.

Before OpenAI-specific controls, cover the same basics every image model needs: subject, setting, style or medium, lighting mood, framing, and exclusions. Name one primary subject, one setting, one visual medium, and what must not appear. Camera language (lens, framing) is a look cue, not a physics simulator.

Keep this section short on purpose. For deeper composition, lighting catalogs, style stacking, and color direction, use the AI Image Prompt Guide. The rest of this page focuses on ChatGPT Images versus API workflows, references, masks, multi-turn edits, and output controls.

OpenAI's cookbook recommends a consistent structure: scene, then subject, then details, then constraints, and stating the intended use. Labeled segments are easier to debug than one long run-on paragraph.

Weak:

Make a cool coffee shop.

Improved:

Create a photorealistic lifestyle photo for a cafe homepage hero. Scene: rainy afternoon, foggy street window, warm interior lamps vs cool blue-gray exterior. Main subject: empty window-side table with one ceramic cup and a closed paperback. Composition: wide landscape, subject left third, negative space right for headline overlay. Style/lighting: candid 35mm look, soft window light, natural imperfections. Constraints: no people, no logos, no readable brand names, no watermark, no extra text.

Why this is more controllable: intended use, subject, composition, lighting, and exclusions replace an open aesthetic vibe.

For photorealism, include “photorealistic” and ask for real texture when you want it. Treat exact camera numbers as look guidance. Generate at the publish aspect or size when you can, rather than cropping a mismatched frame later.

Put required words in quotes, specify typography and placement, spell tricky brand names carefully, and ban extra labels. For dense text on the API, raise quality to medium or high. Official Limitations still say text placement and clarity can fail, so verify lettering visually.

Weak:

Poster that says summer sale.

Improved:

Create a print-ready portrait poster for a retail promo. Exact text: "SUMMER SALE" in bold white sans-serif, centered in the upper third. Secondary text: "Up to 40% off" smaller, directly under the headline. Constraints: no other text anywhere; no logos; high contrast on a deep teal background; leave lower third mostly empty for a product photo later. Quality: medium or high if using the API.

Why this is more controllable: quoted strings, placement, hierarchy, and a ban on invented captions make failures diagnosable. Still expect occasional text errors and regenerate or edit when needed.

State what to change and what must stay identical. The cookbook pattern is: change only X; keep identity, pose, lighting, colors, background, and composition. One change per follow-up keeps failures diagnosable. Residual drift can still happen; do not promise face locking or near-zero identity drift.

Weak:

Make it better.

Improved:

Change only the cup: replace it with a matte black tumbler. Keep everything else exactly the same: cafe interior, window rain, book, camera angle, color grade, and empty right-side negative space. Do not add people, logos, or text.

Why this is more controllable: a single change plus an explicit preserve list makes mismatches easy to spot and correct.

In ChatGPT Images, you can select a region or describe the edit in conversation. Help Center warns that highlights are not always precise and edits may bleed outside the selection, so name the region in text as well.

Index each input and say what to borrow versus what to replace. OpenAI's image generation guide shows compositing with multiple reference images (the documented example uses four). Say “Image 1…”, “Image 2…” and separate identity (keep this label or product shape) from style borrow (match material or lighting only).

Weak:

Use these images.

Improved:

Image 1: product bottle (keep label text and bottle shape). Image 2: marble countertop style reference (borrow material and lighting only). Create a square product photo of the Image 1 bottle standing on a surface matching Image 2’s material. Constraints: do not redesign the label; no extra products; no watermark.

Why this is more controllable: indexed roles prevent the model from inventing which photo is the product and which is only a texture cue.

For gpt-image-2, omit input_fidelity; the API processes inputs at high fidelity and can cost more input tokens. Re-check the current guide if you use older GPT Image models that still expose that parameter.

On the Image API and supported tool flows, provide a mask to indicate the region to replace, plus a prompt that describes the new content. Official mask requirements: the image and mask must share the same format and size, stay under 50MB, and the mask must include an alpha channel.

If you supply multiple images, the mask applies to the first image. Masking with GPT Image is prompt-guided; do not expect perfect surgical inpainting every time.

Improved mask-oriented prompt:

Using the provided mask on Image 1, change only the lounge pool water to include one pink flamingo float. Keep architecture, furniture, lighting, and camera angle identical. Do not alter unmasked areas.

Why this is more controllable: the mask narrows the region while the preserve list and “change only” line constrain bleed. Still verify edges visually.

In ChatGPT Images, refine with natural follow-ups: one variable per turn. On the Responses API, continue with previous_response_id or by referencing a prior image_generation_call id, and optionally set action to edit when an image is already in context. Forcing edit without an image in context errors; leave action at auto when the model should decide.

Turn 1: use the improved cafe hero prompt above.

Turn 2:

Change only the lighting: warmer interior lamps, keep framing, props, and empty right third identical. No new objects.

Turn 3 (if needed):

Remove the faint reflection on the window glass only. Keep everything else the same.

Why this is more controllable: one variable per turn avoids overloaded mega-prompts and conflicting style stacks.

On the API with gpt-image-2, quality options are low, medium, high, and auto. Start with low for drafts when cost or latency matters; move up for dense text or final assets. Default output format is png; jpeg and webp are also available with compression options.

Flexible sizes are constrained, not unlimited. Per the current image generation guide: maximum edge length ≤ 3840px; both edges multiples of 16px; long-edge to short-edge ratio ≤ 3:1; total pixels between 655,360 and 8,294,400. Outputs above about 2560×1440 (3,686,400 pixels) are treated as experimental.

Practical examples include 1024×1024, 1536×1024, 1024×1536, and larger supported sizes listed in the docs. Re-check the live guide on the day you ship; do not assume “any resolution.”

Transparency is surface-specific. ChatGPT Images Help states the product can follow instructions to make a background transparent. Separately, gpt-image-2 does not support background: "transparent" in the API or Responses tool; those requests fail.

Do not copy a ChatGPT transparency prompt into the gpt-image-2 API and expect the same parameter behavior. If you need API transparency, re-check whether another GPT Image model still supports it that day, or composite offline.

ChatGPT Images (product) transparent-background example:

Create a single ceramic mug on a transparent background, centered, soft contact shadow only if needed for grounding. No table, no room, no extra objects, no text, no watermark.

Why this is more controllable on ChatGPT: it matches a product behavior documented in Help, with a single subject and hard exclusions. It is not an API background parameter recipe for gpt-image-2.

Moderation on the API supports auto (default) and low; blocked requests need prompt or input changes, not blind retries. Pricing depends on tokens, quality, size, and input images. Check current API pricing and the image generation calculator rather than memorizing dollar amounts. ChatGPT product pricing and limits follow your ChatGPT plan, not Image API line items.

• Weak composition: add framing, negative space, and intended use; regenerate at the publish aspect.
• Unwanted objects: add explicit exclusions; use “change only / remove X”; one object pass at a time.
• Incorrect text: quote strings; raise quality; ban extra text; accept official text Limitations and iterate.
• Identity or reference drift: index refs; preserve face, outfit, or label; one change per turn; re-supply the reference.
• Overloaded prompts: return to SCQIVE; split conflicting styles across turns.
• Style conflicts: pick one medium; remove competing adjectives.
• Wrong region edited: name the region; tighten selection or mask; say “do not alter surrounding objects.”
• Capability mismatch: expecting ChatGPT picker knobs in the Image API, expecting gpt-image-2 transparency, or using a GPT Image ID as the Responses model. Match task → surface → current model card.

Do not ask for private chain-of-thought as verification. Compare the image to the written request, then change one constraint.

• Intended use is stated (ad, UI, photo, poster, icon…)
• Subject, setting, and framing are explicit
• Constraints cover must-include and must-avoid
• Exact in-image text is quoted (or “no text” is stated)
• Aspect or size matches the publish crop
• Surface matches the job (ChatGPT vs Image API vs Responses)
• Model capability matches the job (especially transparency, masks, multi-turn)
• References and masks are indexed with clear roles
• For edits: change-only + preserve list ready
• Quality tier chosen when using the API (start low for drafts)
• Time-sensitive limits re-checked on the official page the same day
• Verification plan: what you will visually compare, and the next single change
• No request for hidden chain-of-thought

Controllable GPT Image work is a clear production brief, the correct OpenAI surface, honest limits, and visible iteration. Keep ChatGPT Images UI, Image API parameters, and the Responses image tool separate. Use SCQIVE to stay organized, link out to the AI Image Prompt Guide for deep visual craft, and verify every important render against your written request.

When you want help tightening an image brief before you generate, use the free image prompt tool.