Evaluating Reference Preservation in GPT Image 2.5 Multi-Turn Editing

A product label can remain in roughly the same place while one character changes. For repeated image edits, “most pixels stayed put” is not the same acceptance criterion as “the product is still correct.”

For GPT Image 2.5 Sunburst, the decision is whether to admit it to a reference-preserving editing workflow—not whether a leaderboard score makes it production-ready. The CometAPI model page, captured September 14, 2026, identifies Sunburst as part of the GPT Image 2.5 family, lists gpt-image-2.5-sunburst, and positions it for precision editing and consistency across revisions. That positioning makes preservation worth testing; it does not demonstrate that your assets will survive the revision chain.

What the reported results can tell you

The page reports Arena scores of 1,421 for text-to-image and 1,520 for image editing for Sunburst. These are human-preference results, sensitive to the evaluation population, task mix, model version and testing period. They do not directly measure preservation of a reference image.

Its account of a Tosea test is closer to this decision: a five-page, 16:9 presentation workflow, with 41 successful API calls across three models and a three-round editing experiment. The described drift metric counts pixels that change beyond a perceptual threshold outside the requested modification. The 41 calls are not 41 Sunburst trials, nor an API reliability result.

There is also a discrepancy that matters before using those numbers as an adoption threshold:

Sunburst drift as presented on the page Round 1 Round 2 Round 3
Benchmark table 4.1% 3.1% 2.8%
Cumulative narrative 4.1% 6.8% 9.9%

The narrative gives GPT Image 2 cumulative drift as 5.6%, 8.1% and 11.4%. However, the page’s table and cumulative account do not align. Its footnote refers to cumulative drift; it does not establish that the displayed table values are verified per-round increments. Do not add them together or silently choose one series as the correct result.

These are reports from the model page, not independently checked Arena or Tosea evaluations. Taken together, they support shortlisting Sunburst for a controlled test—not a verified preservation advantage, a universal quality ranking or a statistical-significance claim. Even an unambiguous pixel-drift result would not by itself establish correct identity, lettering or product geometry.

A proposed acceptance test for Sunburst

Use representative assets from the workflow you actually intend to support. The following is a test plan, not a completed experiment.

  1. Define what must survive. For a packshot, record the product silhouette, cap geometry, label wording and logo placement. For a portrait, record identity-defining details. For a layout, record text and alignment. Mark intended-edit regions and protected regions for evaluation; this is not a claim about API mask support.
  2. Write a realistic revision chain before generating. For example: replace the background, change a heading, then add a decorative element. Specify what may change at each turn, including any permitted effect on shadows or surrounding layout. Use the previous output for the next edit if that is how the planned application will work.
  3. Keep the comparison interpretable. Record the exact model/version information available, prompts, dimensions, quality selection and supplied references. Hold relevant settings consistent within a comparison. If you change the quality configuration, evaluate that configuration separately rather than assuming the earlier result carries over.
  4. Score the edit and preservation separately. Did the requested change happen? Did every protected detail survive? A low drift score must not earn a pass when the requested edit was ignored. Compare each output with both the previous turn and the original reference, including the final image after the full chain.
  5. Pair pixel checks with visual inspection. Inspect label characters, facial details, edges and composition at useful magnification. A changed digit should be able to fail the asset even if it occupies very few pixels. Record failures by asset category rather than hiding them in one average.

Set acceptance criteria before seeing the outputs. An illustrative rule for packaging is: every requested change must be correct, protected text must remain exact, and product geometry must pass visual review; choose any pixel-change tolerance separately for the permitted background changes.

Adopt Sunburst only for the asset categories, revision sequences and settings that meet those criteria. If it fails, retain review or narrow the supported workflow instead of treating the reported editing score as an override. The useful result is a bounded decision about your reference assets, not another number standing in for preservation.