gpt-image-2:reverse is a publicly callable variant of GPT Image 2, designed for text-to-image generation and visual transformations based on reference images. It is suitable for organizing requirements for titles, subjects, color schemes, and compositions into posters, infographics, product concepts, or character design images, and can also describe desired modifications around existing assets. On this platform, you can create through both generation and editing entry points, receiving image files or asynchronous task results.
Clarify capacity, input/output, and invocation methods before selecting a model.
Understand GPT Image 2's image creation positioning separately from this endpoint's invocation specifications: the quantities, dimensions, and response formats above apply to requests on this platform.
Learn what gpt-image-2:reverse can bring to your work.
GPT Image 2 is well suited for visual creations that include titles, labels, and explanatory text. Prompts can specify text content, subject placement, backgrounds, and information hierarchy at the same time, making text part of the overall composition. Posters and infographics are tasks worth trying first; after generating an image, spelling, numbers, and reading order should still be checked word by word.
When editing, submit the original image together with modification instructions, clearly specifying which elements to retain and which parts to change. For example, preserve the product outline, viewpoint, lighting, and shadows while adjusting only the color or scene. Multiple reference images can separately provide subject, style, and composition references, but their respective purposes should be explained to avoid conflicting requirements from different materials.
Both generation and editing can deliver image results that are convenient for subsequent processing. When comparing options, you can request multiple candidates; when embedding in an application, you can choose image links or Base64 data. Longer tasks can use callbacks to receive final results, separating creative submission, completion waiting, and image saving into clear processing steps.
Start with specific tasks to find where the model can be effective.
Enter the event name, main title, short description, brand color palette, and canvas requirements to create promotional posters or knowledge cards. When there is a lot of information, first define title, subject, and annotation areas to avoid placing all content at the same level. After delivery, focus on checking the text and facts in the image, then refine the layout before official publication.
Use a product photo as editing input, describe the desired background, material atmosphere, and display method, and generate key visuals, lifestyle images, or packaging mockups. Clearly state the shapes and marks that must be preserved in the instructions to make different creative options easier to compare. The results are suitable for concept exploration; formal product displays still require verification of colors, structure, and brand details.
Enter a character description or reference image, and request organized expressions, clothing, equipment, and different views in the image to create character design sheets. You can also generate static images for storyboard preparation based on scene descriptions. Using the same set of appearance descriptions helps reduce deviations, but character consistency across sequential images still needs to be checked and adjusted one by one.
Choose based on task complexity, input materials, and expected results.
When a task includes both titles and layout and also requires adjusting visual content around existing images, you can prioritize trying GPT Image 2. First validate the generated image using real copy and reference assets, then decide whether to incorporate it into the production workflow. gpt-image-2:reverse and :official are publicly callable variants of the same base model and should not be understood as image models from different generations.
If the work focuses on quickly generating candidates, you can further compare GPT Image 2.5 Flare; if high fidelity and fine-grained control matter more, you can compare Sunburst. Existing GPT Image 1.5 workflows are well suited to parallel testing with the same prompts and assets, observing whether text, subject preservation, and composition better meet the goal rather than replacing it based solely on version numbers.
From a small-scale task to formal integration.
Define the goal, required inputs, and output requirements, using real business examples as a starting point.
Open the trial page, confirm the parameters supported by this entry point, then submit a small-scale task to review the results.
Keep the complete model ID, use the request format specified in the documentation, and confirm billing rules on the Pricing page.
Before formal use, understand the scope of output quality and capabilities.
Answers to common questions about using gpt-image-2:reverse.
No. It is a public API ID variant of GPT Image 2, with the same core purpose of image generation and editing. Specify the full ID when calling it to explicitly select this endpoint; do not interpret the :reverse suffix as a new-generation model, or infer image-quality tier differences from it compared with other variants.
Use /openai/images/edits, submit image and prompt, and explicitly set model to gpt-image-2:reverse. Images can be provided as a URL or uploaded as a local file. It is best to separately state what to preserve and what to change, for example, preserve the product and camera angle while only changing the background, then check the subject details after completion.
The editing endpoint can accept up to 16 reference images. Use a URL array for JSON requests and multipart for local files. It is recommended to explain the role of each image: which provides the subject, which provides the style, and which provides the layout. Multi-image input does not mean that different people, products, or compositions will automatically be accurately merged.
Use size=auto when you want the model to determine the aspect ratio based on the creative intent; specify WIDTHxHEIGHT when fixed pixels are needed, subject to size constraints. Results can be returned as URL or b64_json; use URL when comparing multiple options at once, while Base64 responses support only a single image. Check the actual dimensions after downloading.
This endpoint directly handles image generation or editing requests and returns image results; it should not be considered equivalent to the full ChatGPT experience, which includes reasoning, search, or multi-step orchestration. For iterative revisions, have the application save the previous image and submit another edit; longer tasks can receive completed results through callback_url.
Model information · Updated: 2026-10-01. See the API and pricing sections for request parameters and billing rules.
Start with a clear goal and evaluate whether it suits your work based on real results.
Pricing
$0.0105· per generation
You pay$13,584for 142,800 credits · $0.0951 / credits46% OFF
You're billed in credits. Pick a top-up tier to see its per-credit rate — the larger the top-up, the cheaper each credit.
This is only a basic call example. See the full docs for more parameters and advanced usage.
View Docs