GPT Image 2.5 Sunburst: AI Image Editor Online

These limits are the ones VdoBloom's GPT Image 2.5 service enforces, live-checked against the provider on 12 September 2026. The provider's own documentation lists extra ratios for the editing model (5:4, 4:5, 2:1, 1:2, 3:1, 1:3, 9:21); the live API rejects all of them, so VdoBloom does not offer them.

The charge follows the resolution you select. One thing to know: the four 1K-only ratios (27:16, 16:27, 9:8, 8:9) are rendered at 1K even if you select 2K or 4K, and you are charged the 1K price for them, not the tier you picked. The dollar figures use the Lite rate of $0.05 a credit.

Sunburst and Flare share every limit and every price; the difference is tuning, so trying the same edit on both costs the same as trying it twice on one. Against GPT Image 2, Sunburst gains six reference slots, the 4K square and four extra ratios, and loses 4:5 and 5:4. Against Nano Banana 2.1 and Qwen Image 2.1, it costs the same at 1K and more at 2K and above. VdoBloom does not publish a quality ranking; run your own edit on two.

Price does not decide it, because it is identical. Pick Sunburst when the edit has to hold up to close inspection: matching light across several references, fine texture, small lettering, a restored face. Pick Flare when you are iterating quickly and turnaround matters more. VdoBloom's tools follow the same split: Room Restyle and Die-cut PNG run on Sunburst only, and tools such as Relight, Restore & Colorize, Makeup Looks, Translate Image Text, Age Progression and Car Color & Wrap offer Sunburst with Flare as the alternative. A practical routine is to draft the instruction on either variant at 1K, then run the final at 2K or 4K on Sunburst.

In Image Edit you choose everything. In the image tools VdoBloom writes the prompt, runs the model at 2K, and shows the credit total before you generate.

Lead with the one change, then fence everything else: 'Relight the scene as late golden hour from the left. Keep the people, their expressions, clothing, the furniture and the camera angle exactly as they are.' With many references, number them and assign roles ('Images 1 to 4 are the same product from four angles; image 5 is the shelf'). Sixteen is a ceiling; give the model only the photos that carry information it needs.

Put text that must render in quotes and spell it exactly, including capitals. Choose the ratio before you write, and avoid the four 1K-only ratios when the deliverable is 2K or 4K. If you plan to use the result as a cut-out, describe an isolated subject on a plain background; the provider documents transparent output as needing 2K or 4K and a prompt of that kind.

GPT Image 2.5 introduced four ratios that GPT Image 2 does not have. The provider renders them at 1K only, and VdoBloom clamps the resolution instead of refusing the job. They suit specific crops rather than everyday posts.

Sixteen references is the headline change from GPT Image 2, and it is useful when a job really needs many sources: a product seen from several angles, a set of brand colour swatches, a style reference and a background plate together, or the frames of a sprite sheet. The price does not change with the number of photos, so the cost is attention, not credits. Number every image in the prompt and tell the model what each one is for. Group similar references ('images 1 to 4 are the same sneaker from four sides') so the model treats them as one subject rather than four. If the result mixes up details, remove the references that add least and run again at 1K before paying for 2K or 4K.

27:16, 16:27, 9:8 and 8:9 are 1K-only, and the charge follows the resolution you select, so select 1K with those ratios. There is no 4:5 or 5:4 on GPT Image 2.5, even though GPT Image 2 has both; for a 4:5 feed crop use GPT Image 2 or Nano Banana 2.1. Transparent background is not a switch in Image Edit; VdoBloom uses it inside the Die-cut PNG tool. Sunburst is the slower of the two variants by design.

The STRICT label reflects OpenAI's provider rules: edits involving real people, suggestive content or violence are declined more often than on models labelled HOT. Reports of soft or noisy results usually trace back to an ambiguous instruction or a ratio that was clamped to 1K, so check both before judging the model. Failed jobs are refunded automatically. It is a proprietary OpenAI model, used through VdoBloom's hosted route.