GPT Image 2.5 Sunburst:official is an image generation and editing endpoint focused on high fidelity and fine-grained control. It can create new images from text and use reference images to adjust backgrounds, objects, and visual details, making it suitable for product assets, campaign key visuals, and iterative revisions to existing designs. Compared with Flare, which emphasizes speed, Sunburst is better suited for tasks that require attention to modification requirements and preserved details.
Clarify capacity, input/output, and invocation methods before selecting a model.
Sunburst is positioned for high fidelity and fine-grained control; the quantity, format, and size rules above correspond to this platform's actual invocation endpoint.
Learn what gpt-image-2.5-sunburst:official can bring to your work.
When editing, you can submit images and text requirements together, separately explaining “what needs to change” and “what needs to be preserved.” When changing a product background, modifying props, or adjusting lighting, clearly specify the subject’s shape, color, labels, and composition requirements so revisions build on the existing visual rather than redesigning the entire image each time.
When you need to focus changes on a specific area, upload the original image together with an Alpha PNG mask: transparent areas may be edited, while non-transparent areas indicate what should be preserved. The prompt should still describe the complete target image and clearly specify local changes. Masks are suitable for replacing background elements or correcting objects, but edge blending should still be checked after completion.
For consecutive edits, the image selected in the previous round can be used as the reference for the next round, handling backgrounds, lighting, and object details in sequence. Save confirmed versions after each round to prevent later changes from affecting previously approved parts. Generated results can be integrated into applications via image URL or Base64, and completion results for longer tasks can also be received through callbacks.
Start with specific tasks to find where the model can be effective.
Provide product photos, target environment descriptions, and packaging details that need to be preserved to create studio, holiday, or lifestyle visuals. Background areas can be constrained with masks, then check product outlines, labels, and shadows. Deliverables are suitable for creative review; if product pixels must remain completely unchanged, retain the original product and use compositing instead.
Provide the selected key visual and brand references, specify the new scene, props, negative space, and copy placement, and create revisions based on the existing concept. This is suitable for tasks where the design direction has already been determined and only several visual elements need to change. Before delivery, check brand colors, text content, and layout item by item, and avoid treating generated drafts directly as final print files.
Use the selected concept image as a reference, list changes for objects, backgrounds, and local details, and complete targeted revisions in stages. You can first use Flare to explore visual directions, then use Sunburst for detailed revisions. Focus each round on a clear objective, output candidate images and retained versions, and make it easier for designers to compare and choose.
Choose based on task complexity, input materials, and expected results.
When you need to quickly explore composition, style, and creative directions, prioritize the speed-focused GPT Image 2.5 Flare; when you already have reference images and the requested changes focus on products, brand elements, or local details, Sunburst is more suitable. The trade-off between the two should depend on the task stage, rather than treating Sunburst as necessarily the better option for every generation task.
gpt-image-2.5-sunburst:official and the suffix-free Sunburst share the same core creative positioning; the suffix does not indicate a new image version. This endpoint is billed based on actual Token usage for text input, reference image input, and image output, while the suffix-free endpoint is billed per successfully generated image. Explicitly specify the full ID when calling it, and save the parameters used to make reproduction and comparison easier.
From a small-scale task to production integration.
Define the objective, required inputs, and output requirements, and start with real business examples.
Open the trial page, confirm the parameters supported by this endpoint, then submit a small-scale task to review the results.
Keep the full model ID, use the request format specified in the documentation, and confirm the billing rules on the Pricing page.
Before formal use, understand the output quality and capability boundaries.
Answers to common questions about using gpt-image-2.5-sunburst:official.
Yes. Use /openai/images/generations for text-to-image generation, and /openai/images/edits to modify existing images. The former submits a description and the full model ID, while the latter also requires the original image. Reference-image editing can be used to change backgrounds, modify props, and adjust details, rather than merely regenerating a similar image.
Use the product photo as a reference, and clearly list the outlines, colors, labels, and composition to preserve, while specifying the background or objects that need to be changed. Local edits can be combined with masks, but pixel-perfect preservation cannot be guaranteed; if the original product must remain exactly identical, it is recommended to generate the environment and then composite it with the original product.
This is determined by the Alpha channel: transparent pixels with an Alpha of 0 indicate areas that may be edited, while non-transparent pixels indicate areas that should be preserved; it is not determined by black-and-white colors. The mask must be a PNG no larger than 4MB and have the same dimensions as the first original image; the original image and mask must be uploaded together via multipart.
The editing endpoint supports up to 16 reference images, submitted using a URL array or multiple multipart image fields. The generation count n supports 1–10; if b64_json is selected as the return format, only 1 image can be generated. When using multiple image inputs, explain the role of each reference image in the prompt to reduce ambiguity about the target.
No. This endpoint is billed based on the actual Token usage of text input, reference-image input, and image output, and pricing before a request is an estimate. Size, quality, reference images, and generation quantity all affect usage. When evaluating a complete revision task, multiple rounds of revisions should also be included in the cost, rather than considering only the initial generation.
Model information · Updated: 2026-10-01. See the API and pricing sections for call parameters and billing rules.
Start with a clear goal and assess from actual results whether it suits your work.
This is only a basic call example. See the full docs for more parameters and advanced usage.
View Docs