GPT Image 2.6 is useful in production when you stop treating every revision as a new generation. A safer pattern is a two-stage pipeline: use Flare to explore a composition, then use Sunburst to edit the selected asset while explicitly protecting everything that has already been approved.
This solves a common product problem: a user asks to change the background, but the model also modifies the logo, product geometry, or character identity. The fix is not merely a longer prompt. It is a workflow with a preservation contract, input validation, and an acceptance gate.
The requests below are minimal examples based on the public Ace Data Cloud API documentation. They use placeholder credentials and were not executed for this article.
1. Generate a candidate with Flare
Use the generation endpoint for exploration. Keep the request explicit so you can reproduce the asset later:
curl https://api.acedata.cloud/openai/images/generations \
-H "Authorization: Bearer <YOUR_API_KEY>" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-image-2.6-flare",
"prompt": "Editorial product photo of a matte red travel mug on a light gray desk, three-quarter view, soft window light, no text, no extra objects",
"size": "1536x1024",
"quality": "medium",
"output_format": "png",
"response_format": "url"
}'
Flare fits this stage because the public launch guide positions it for faster iteration and previews. That is a workflow recommendation, not a latency guarantee: actual completion time can vary. Save the returned asset URL together with the prompt, model, size, and a stable internal asset_id.
At this point, let a human or an application rule select one candidate. Do not edit “the latest result”; edit a specific immutable input. That small distinction prevents a retry from silently changing the source image.
2. Turn approval into a preservation contract
Before editing, represent the requested change and the protected features separately:
{
"change": "Replace the desk with pale oak and add a subtle plant shadow",
"preserve": [
"mug shape and handle geometry",
"matte red color",
"camera angle and crop",
"lighting direction",
"no text or logo"
]
}
Then compile those fields into a constrained edit instruction. This is easier to audit than accepting one unstructured sentence from a UI.
function buildEditPrompt(change, preserve) {
if (!change?.trim()) throw new Error("change is required");
if (!Array.isArray(preserve) || preserve.length === 0) {
throw new Error("at least one preservation rule is required");
}
return [
`Change only this: ${change.trim()}.`,
`Preserve exactly: ${preserve.join("; ")}.`,
"Do not introduce text, logos, or additional products."
].join(" ");
}
This validation is intentionally conservative. It may reject a legitimate free-form request, but it makes accidental broad edits less likely. For creative tools, expose an “unconstrained edit” mode separately rather than weakening the default contract.
3. Edit the selected asset with Sunburst
The edits endpoint uses multipart/form-data. Download the selected asset into controlled storage, verify it is an accepted image type, and submit that exact file:
curl https://api.acedata.cloud/openai/images/edits \
-H "Authorization: Bearer <YOUR_API_KEY>" \
-F "model=gpt-image-2.6-sunburst" \
-F "image=@selected-candidate.png" \
-F "prompt=Change only this: replace the desk with pale oak and add a subtle plant shadow. Preserve exactly: mug shape and handle geometry; matte red color; camera angle and crop; lighting direction; no text or logo." \
-F "size=1536x1024" \
-F "quality=high" \
-F "output_format=png" \
-F "response_format=url"
Sunburst is the better fit when detail and consistency matter more than rapid exploration. The API currently accepts both GPT Image 2.6 Flare and Sunburst for generation and editing, so this split is an application-level policy, not an endpoint restriction.
If the editable region is known, add a PNG mask. The edit documentation requires an alpha channel, a maximum size of 4 MB, and dimensions matching the first input image. Transparent pixels mark areas that may change; opaque pixels protect the rest. A black-and-white RGB image without transparency is not a valid substitute.
4. Validate before spending a request
For a fixed production canvas, reject invalid dimensions locally:
function validateCanvas(width, height) {
const pixels = width * height;
const ratio = Math.max(width / height, height / width);
if (width % 16 || height % 16) throw new Error("dimensions must be multiples of 16");
if (Math.max(width, height) > 3840) throw new Error("long edge exceeds 3840");
if (pixels < 655360 || pixels > 8294400) throw new Error("pixel budget is invalid");
if (ratio > 3) throw new Error("aspect ratio exceeds 3:1");
}
Also check MIME type, decoded dimensions, file size, and whether the source can be fetched without forwarding user credentials. Store remote results in your own object storage if you need durable retention; a returned URL should not be your only database record.
5. Add an acceptance gate
An HTTP 200 means the API produced an image, not that the edit is acceptable. Record the result as needs_review, then compare it with the preservation contract:
- Is the product silhouette unchanged?
- Did the protected color drift?
- Are crop and camera angle stable?
- Did unwanted text or extra objects appear?
- Was the requested background change actually applied?
Automatic similarity or OCR checks can flag obvious failures, but a human reviewer is appropriate for brand-critical assets. If review fails, create a new edit attempt linked to the same source; never overwrite the approved original.
For slow operations, the documented API also supports asynchronous processing and callbacks. Persist task_id, trace_id, source asset ID, request parameters, and status, and make callback handling idempotent. Retry transport failures with backoff, but do not automatically retry validation errors.
Trade-offs
The two-stage design costs an extra request and adds storage and review state. In return, it separates exploration from controlled revision, makes failures traceable, and prevents a successful edit from quietly replacing an approved asset. That is usually the right trade for product imagery, recurring characters, and brand templates.
You can inspect GPT Image 2.6 Flare, GPT Image 2.6 Sunburst, or try the workflow in Ace Data Cloud Studio. The full request fields are in the generation documentation and editing documentation.
Disclosure and sources
This article was prepared for an Ace Data Cloud promotional task and is commercially related to Ace Data Cloud. It is an independently reorganized technical guide based on Ace Data Cloud's public GPT Image 2.6 launch article, generation documentation, and editing documentation. The author does not claim to have personally executed the example requests. AI assistance was used in drafting and editing; the technical claims and links were checked against the public sources available at publication time.
Top comments (0)