GPT Image 2.5 transparent assets: Flare and Sunburst workflow

A transparent product cutout or illustration needs usable edges, not just a file named .png. This guide uses the platform IDs gpt-image-2.5-flare and gpt-image-2.5-sunburst; gpt-image2.5 is not the ID used by these integrations. It covers asset preparation and compositing checks rather than repeating the full generation and editing references.

Choose the workflow before choosing a quality level

For a new illustration, describe the subject and leave images out. For a reference-guided edit, supply your authorized source image in input.images and explicitly describe what must remain unchanged. Reference input is guidance, not a guarantee of exact product reproduction.

Both variants use the platform POST /v1/tasks envelope. Their shared repository schema declares background, output_format, quality, size, resolution and aspect_ratio. A shared schema does not establish equal output quality, speed or price. Start with the variant available to your account and evaluate the deliverable before comparing alternatives.

Decision	Suggested starting point	Check before delivery
Transparent asset	background: "transparent", output_format: "png"	File actually has non-opaque alpha
Opaque photograph	Explicit opaque background if required	No unwanted cutout edges
Square composition	resolution: "1k", aspect_ratio: "1:1", omit size	Returned dimensions and crop
Quality comparison	Keep prompt, size and references fixed	Edge quality and accepted-asset cost

PNG can store transparency, but selecting PNG alone does not ask the model for a transparent background. JPEG cannot preserve an alpha channel. Use PNG here even though a schema may accept other combinations.

Construct one transparent illustration task

This is illustrative request JSON, not an executed generation. Confirm the live schema through GET /v1/media-models before submitting; generation incurs usage.

{
  "model": "gpt-image-2.5-flare",
  "input": {
    "prompt": "One isolated blue ceramic bird figurine, three-quarter view, complete silhouette with space around every edge. Transparent background. No lettering, no scenery, no ground plane and no cast shadow outside the figurine.",
    "background": "transparent",
    "output_format": "png",
    "quality": "medium",
    "resolution": "1k",
    "aspect_ratio": "1:1"
  }
}

For Sunburst, use the exact gpt-image-2.5-sunburst ID after checking access and the current contract. For an edit, add an images array containing a reachable authorized reference URL and replace the illustration prompt with a preservation brief. Do not add native multipart fields or an undeclared mask parameter to this task body.

The Flare editing guide and Sunburst editing guide contain the complete input references. Save the returned task ID and poll the original task; an accepted task is not the final PNG.

Verify the delivered alpha and edge quality

Open the returned asset in an editor that displays transparency, then composite it over white, black and a saturated background. Check three separate things:

1. Alpha: pixels outside the subject are transparent rather than a painted checkerboard. A PNG file with all alpha values opaque is not a cutout.
2. Edges: no light halo, clipped silhouette, missing fine detail or unwanted shadow outside the subject. Inspect at the actual delivery scale as well as close up.
3. Identity: reference-guided edits preserve the required shape, colors and text. If exact lettering matters, use approved artwork in a later deterministic layout step.

A successful task status establishes completion, not these acceptance criteria. Save the original output before resizing or format conversion, which can flatten transparency. Test the final exported asset in the destination application too.

Compare quality settings without losing budget control

Use one approved brief and a fixed size for a small budgeted comparison. Change one setting at a time and record task ID, variant, input parameters, actual dimensions, review result and charge. Do not assume max or a larger resolution is always the economical choice for a small UI asset.

When supplying size, follow the size and pricing precedence rules; do not mix sizing controls and infer the billing tier from resolution alone. Use batch budgeting to include rejected attempts in cost per accepted asset.

Sources and limits

Checked 2026-11-03: OpenAI Flare model, Sunburst model, and image generation guide. These establish model identity and native concepts; platform fields follow the platform contract. No quality benchmark or price ranking is claimed here.