High-Fidelity Image Model for Product Visuals and Fine Revisions
GPT Image 2.5 Sunburst is the image generation and editing model offered by this platform, positioned with an emphasis on high fidelity and fine control. It is suitable for product background adjustments, event visual revisions, and detail refinements based on existing assets. It can generate new images from text and also perform edits using reference images. Compared with Flare, which is positioned on the platform for rapid generation, Sunburst is better suited to creative tasks with a clear design direction that require repeated instructions about what to preserve and what to modify.
Input parameters and result formats vary by service. Use the public API for this model and follow its guide for generation, task retrieval and editing operations.
Specifications and API Features
Creation method
Text-to-image generation; combined editing with reference images and text
Platform reference image input
A single URL or up to 16 URLs; local images can be uploaded using multipart
Platform image count
n is 1–10; the b64_json return format supports only 1 image
Platform aspect ratio settings
auto or WIDTHxHEIGHT; width and height must be multiples of 16, with the longer side not exceeding 3840 pixels
Platform size limits
Total pixels 655,360–8,294,400, with an aspect ratio no greater than 3:1
Platform result delivery
Image URL or Base64 image data; long-running tasks support asynchronous callbacks
Local editing method
The :official variant supports multipart Alpha PNG masks
The quantities, sizes, and delivery methods above apply to this platform's calling scope and do not represent native model capacity claims; local mask editing uses the corresponding :official variant.
Core Capabilities
Make targeted modifications around the original image
Sunburst focuses on high fidelity and fine control, rather than merely generating images from scratch. When editing, you can specify the background, lighting, or objects that need to change, as well as the subject shape, colors, and composition that should be preserved. It is suitable when the design direction has already been determined and you want to reduce the work of redoing the entire image, while also making it easier to check each modification against requirements.
Turn reference assets into creative constraints
By combining image and text inputs, you can propose new scene concepts based on existing product photos, campaign key visuals, or character assets. Each of multiple reference images should have its purpose explained, such as subject appearance, background mood, or color direction, to avoid having the model guess the relationships between assets. Outputs still need review; reference images do not mean that every detail will be copied precisely.
Keep reversible versions throughout the revision process
Continuous revisions can use a selected result as the input for the next edit, handling the background, lighting, and local objects in sequence rather than redescribing the entire image every round. It is recommended to save each confirmed version and clearly specify what needs to be changed in the current round only. Organizing creation this way makes it easier to identify whether later modifications affect previously approved subject details.
Use Cases
Product image scene replacement
Input a product photo and describe a new studio, holiday display, or lifestyle setting, while also listing preservation requirements such as package outlines, label content, and brand colors. Deliverables can be used for product presentations and marketing creative selection. If the product itself must remain unchanged pixel by pixel, composite the generated background with the original product asset rather than relying solely on the generated result.
Targeted revisions of campaign visuals
Use an existing campaign key visual as a reference, instruct changes to props, the environment, or the compositional focus, and specify brand elements that must not change. This is suitable for exploring different options within an established visual direction, with outputs for designers to select and lay out. When titles, dates, and brand marks are involved, the final text content and placement should still be checked manually item by item.
Fine-tuning after selecting a draft
Submit a selected generated draft to the editing entry point and focus on modifying background distractions, object colors, or lighting relationships. When mask control is needed, you can choose the :official variant to upload the original image and an Alpha PNG mask. The deliverable is an image version that can continue to be revised; during acceptance, focus on comparing the modified areas, edge blending, and content that should be preserved.
How to choose this model
Choosing between Sunburst and Flare
If the task first requires quickly obtaining multiple concept directions, the speed-focused GPT Image 2.5 Flare is more suitable as a candidate; if you already have source material and a clear list of changes and want to refine product appearance or event details, Sunburst can be prioritized. You can also select a Flare draft first, then hand the image to Sunburst for editing. The choice between them should be based on the actual image results, without assuming a fixed degree of quality improvement.
Choosing between the standard ID and :official
gpt-image-2.5-sunburst and gpt-image-2.5-sunburst:official share a positioning of high fidelity and fine control; they do not represent two generations of models. Use the standard ID for regular text generation or revisions based on reference images; choose the :official variant when uploading masks according to a defined workflow. Explicitly specify the full ID in calls, and save the image input, response method, and specific parameters together to facilitate reproducing and comparing revision results.
Getting started
Determine text-to-image or image revision
Provide a prompt when generating; provide both image and revision instructions when editing. Clearly state the original text, requirements for preserving the subject, and the target aspect ratio.
Specify the model and parameter format
Call the image generation or image editing endpoint, explicitly specifying model=gpt-image-2.5-sunburst; use auto or WIDTHxHEIGHT for size, generate one image first, then evaluate it. Configure masks, quality, and file format according to this endpoint guide.
Check images and cost records
Read the URL or Base64 image according to the response format, save task_id asynchronously before querying results; check text, reference details, and alpha channels, and record usage according to the current Pricing rules.
Trial suggestion: detailed revisions to an existing product image
Input and goal
Preserve the product bottle, logo, text, and camera angle, replacing only the background flowers with green leaves while keeping the bottle reflections and light-colored tabletop consistent.
Acceptance and next steps
Make revisions around a selected single image, and zoom in to inspect labels and outlines; do not treat “preserve” instructions as a guarantee that all other areas will remain completely unchanged.
Limitations
High fidelity does not mean the original image pixels remain completely unchanged. When modifying backgrounds or objects, subject textures, labels, people’s appearance, and surrounding lighting and shadows may still change; successive revisions may also affect previously approved details. When product photographs must be preserved exactly as they are, use compositing and save approved versions at each stage.
Text and layout still require review. Accurate wording, character clarity, title placement, and complex information hierarchy in posters cannot be guaranteed by prompts alone. It is suitable to first generate a visual concept, then check brand text, dates, and layout; final deliverables requiring strict typesetting can be completed in design tools.
Mask editing has separate input requirements: use the :official variant and place the original image and mask in the same multipart request. The mask must include an Alpha channel and be the same dimensions as the first original image; URL-based original images and local masks cannot be mixed; natural blending may still occur along editing edges.
Frequently Asked Questions
Can Sunburst generate images directly from text?
Yes. Submit model=gpt-image-2.5-sunburst and prompt to /openai/images/generations to begin text-to-image generation. Describe the subject, environment, lighting, and composition, and clearly state whether text is needed; if an existing image needs to be modified, use the editing endpoint instead.
How can I make reference image edits more closely match the original design?
Submit the image and editing instructions to /openai/images/edits, and explicitly specify the Sunburst ID. The prompt should separately describe the modification target and content to preserve, and multiple reference images should also indicate their respective purposes. Focus on a clear target in each round, and compare with the original image to check subject and brand details.
How should Sunburst masks be prepared?
The mask workflow uses gpt-image-2.5-sunburst:official, with both the original image and mask uploaded via multipart. The mask must be a PNG with an Alpha channel, match the dimensions of the first original image, and not exceed 4MB; transparent areas may be modified, while images containing only black-and-white colors without a transparency channel cannot be used as a substitute.
Can I generate multiple images at once and return Base64?
The platform allows n to be set from 1–10, which is suitable for obtaining multiple candidate images; however, response_format=b64_json supports only 1 image. Use URL responses when multiple results are needed, and make a single-image request when directly processing Base64 image data; do not mix the two settings.
Will multi-round editing automatically remember previous images?
You should use the selected image as the input for the next edit and specify the changes and preservation requirements for the current round in the new instructions; do not rely only on previous requests. Long tasks can include callback_url to first obtain task_id and receive results when completed; save images and instructions from each round for comparison and rollback.
