GPT Image 2.5 Flare and Sunburst offer the same broad generation and editing controls, while their published positioning differs: Flare targets everyday work and Sunburst spends more time on detailed work. Choose using your delivery requirement and a matched evaluation, rather than assuming the variant name is a quality setting.
On Raywake, generation and editing are separate operations for each variant. The model choice, output size, quality level and number of references all belong in the decision.
Flare and Sunburst at a glance
The provider's model overview describes Flare as the default option and Sunburst as the precision-focused alternative. Those are vendor descriptions, not Raywake benchmark results.
| Decision | What to check |
|---|---|
| Generate a new image | Select a text-to-image operation |
| Modify an existing asset | Select an edit operation and attach the source |
| Keep changes within an area | Check the edit operation's mask input |
| Export a transparent object | Select transparent background and a format that retains alpha |
| Deliver a larger image | Set explicit dimensions within the endpoint limits |
| Increase processing effort | Choose the quality setting deliberately |
The reviewed quality choices are auto, low, medium, high, xhigh and max. They are separate from the Flare/Sunburst choice. Comparing Flare at one quality with Sunburst at another confounds two changes.
Use the same acceptance test for both variants
Suppose you are editing a leather wallet for an ecommerce detail page. Your acceptance criteria might include the original stitch count, seam position, leather color and clasp shape. The background may change, but those identity details should not.
Run a matched brief only after settling the source image, crop, quality level and output dimensions. Check the image at its actual delivery size as well as in a close crop. An enlarged preview can make a difference seem decisive even when customers will only see a small thumbnail.
A practical comparison record has five fields: settings, quoted amount, settled amount, time until the result is ready, and accepted or rejected. If one variant needs fewer corrections for your specific assets, that is useful evidence. It is more useful than borrowing a general ranking from someone else's prompts.
What do references and masks control?
The reviewed edit input accepts up to 16 reference images and an optional mask. Reference order and a clear instruction help establish which image supplies the product, composition or material. A mask identifies the region you intend to change; it is still necessary to inspect the rest of the result for unwanted edits.
For an existing product shot, begin with one authoritative product image. Add a second reference only when it answers a specific visual question. An oversized reference set can introduce competing backgrounds, colors or angles without making the brief clearer.
Source: GPT Image 2.5 Flare editing documentation. Raywake's Flare Edit model page shows the current available controls.
Transparency and output dimensions
The background control offers auto, transparent and opaque. Use PNG or an appropriate WebP output for a cutout, then inspect the alpha channel on both a light and dark background. JPEG cannot carry transparency.
For text-to-image, the reviewed size rules include a maximum edge of 3840 pixels and a maximum aspect ratio of 3:1. Concrete dimensions must satisfy the model's other pixel constraints as well. “4K” in a brief is not a substitute for specifying the required width and height.
If your final layout is 1200 by 630 pixels, decide whether you need a larger master for cropping. Generate for the actual design need, then resize for web delivery. A larger generation is wasteful when the extra detail will never be used.
What should the cost comparison include?
Use Raywake quotes for the exact inputs. A text-only generation and an edit with several references are different workloads, even if both return one image. Quality and dimensions can also change the charge.
Do not publish a provider's historical sample price as a guaranteed Raywake price. The image and video cost guide explains quotes, reservations and settled charges. For a purchasing decision, compare cost per approved asset and count deliberate retries.
Common questions
Is Sunburst automatically the right final-render model?
No. Its published positioning makes it a candidate for detail-sensitive work. Your own acceptance criteria determine whether the difference matters enough for your workflow.
Can I switch variants without reviewing settings?
Check the current schema and quote again. Similar controls do not make a previously approved price reusable for a different model.
Is this a measured speed comparison?
No. This guide compares documented controls and a selection process. It does not claim a measured latency advantage from Raywake tests.
Explore Flare and Sunburst on RaywakeReviewed on 19 October 2026. The cover is an editorial illustration generated with Nano Banana 2; it is not a test result from the models discussed.