GPT Image 2.5 comes in two versions on Lunart: Flare for generating new images quickly, with reliable text in the image, and Sunburst for editing an existing image while keeping faces, layout and lettering intact. Both have the same settings and the same price, so the choice depends only on the job.
Flare or Sunburst: which one to pick
| GPT Image 2.5 Flare | GPT Image 2.5 Sunburst | |
|---|---|---|
| Best for | New images, drafts, posters with text | Edits to an image you already like |
| Strength | Fast, accurate text rendering | Keeps faces, layout and text unchanged |
| Reference images | Up to 16 | Up to 16 |
| Resolution | 1K / 2K / 4K / custom | 1K / 2K / 4K / custom |
| Transparent background | Yes | Yes |
| Outpainting | Yes | Yes |
| Price | Same | Same |
A simple rule: start with Flare when you are creating something new. Switch to Sunburst when you give the model an image and ask it to change one thing, such as swapping a background, changing a product color or fixing a headline, and you want everything else to stay put.
Generate your first image
- Sign in and click New canvas.
- In the composer on the right, switch the mode to Image.
- Open Select model and choose GPT Image 2.5 Flare.
- Write your prompt, then set the aspect ratio, Resolution and Quality.
- Check the credit total shown in the composer and click Generate.
The image lands on the infinite canvas, where you can compare versions side by side and keep editing.
What it costs: quality and resolution
The price per image is the quality level plus a resolution add-on. Both versions use the same table.
| Quality | 1K | 2K | 4K |
|---|---|---|---|
| Low | 3 | 13 | 28 |
| Medium | 5 | 15 | 30 |
| Auto | 12 | 22 | 37 |
| High | 12 | 22 | 37 |
| X-High | 22 | 32 | 47 |
| Max | 50 | 60 | 75 |
The cheapest workflow is to explore at Low or Medium, 1K, then rerun the winning prompt at High or above in 2K or 4K. If a generation fails, the credits are returned automatically. Current pack prices are on the pricing page.
Writing prompts that work
GPT Image 2.5 follows long, specific prompts well. Instead of a list of keywords, write it the way you would brief a designer:
- One sentence per element: subject, setting, lighting, style and where the image will be used.
- Put on-image text in quotes:
headline "Fall Collection", subline "Free shipping over $50". Short headlines, prices and taglines come out accurately, which is where Flare is strongest. - Say what to leave out: "plain background, no extra props or text".
- For edits, name what must not change: "replace only the background with a beach at sunset; keep the person's face, pose and clothing exactly the same". This is the kind of request Sunburst is built for.
If you are stuck, the prompt optimization button expands a rough idea into a fuller prompt, and Reverse prompt writes a prompt from an existing image. Each costs 1 credit.
Transparent background PNGs
For logos, stickers, icons and product cutouts:
- Choose GPT Image 2.5 Flare or Sunburst.
- Turn on Transparent background in the composer.
- Describe a single subject, for example "a ceramic coffee mug, centered, studio lighting".
The result is a PNG with no background, ready for a slide, a Shopify product page or a design file. For an image you have already generated, use Remove background instead; see how to remove a background with AI.
Custom sizes and 4K
Besides the preset aspect ratios, GPT Image 2.5 accepts a Custom size: width and height in multiples of 16, no edge longer than 3840 px, and an aspect ratio no wider than 3:1. Enter the exact pixel size your destination needs, such as an Amazon listing image or a website hero banner, and skip cropping later. The composer shows the price for that size before you generate.
Reference images for consistency
You can attach up to 16 reference images per generation, uploaded from your device or picked from the canvas. Type @ in the prompt to say what each one is for:
- Style match: "use the colors and layout of @Reference 1 for a new product poster".
- Product consistency: "place the product in @Reference 1 into the kitchen scene in @Reference 2; do not change the product".
- Combining elements: one reference each for the person, the outfit and the location.
The more references you add, the more clearly each one's role should be stated.
Keep editing on the canvas
Select any image on the canvas to continue:
- Outpaint extends the scene beyond its edges, for example turning a portrait image into a wide banner, either to a target ratio or by an expansion factor.
- Omni Edit lets you draw arrows or boxes on the image and write a note for each spot; only those areas change.
- Upscale increases resolution.
- Split layers separates the image into editable layers you can export as a PSD; see how to split an image into PSD layers.
Comparing GPT Image 2.5 with the Nano Banana models? Read GPT Image 2.5 vs Nano Banana. To try it now, open Lunart, start a new canvas and pick GPT Image 2.5 Flare.