How to Use GPT Image 2.6: Text, Edits, GIF Stickers and SVG
How to use GPT Image 2.6 by OpenAI: which version to pick, text in images, edits from a reference, animated GIF stickers and SVG vectors, plus common mistakes.
To use GPT Image, the same OpenAI engine that draws images in ChatGPT, open it in the IIshki studio: write a prompt, choose an aspect ratio and a resolution up to 4K, and press Generate. The newest series is GPT Image 2.6, available in two versions: Sunburst for precise edits from a reference and Flare for posters and layouts. On top of regular images, 2.6 can output an animated GIF sticker or an SVG vector. You don't need a ChatGPT subscription or a VPN. Below: which version to choose, how to write prompts with text, how to edit your own photo, and the mistakes people run into most often.
What is GPT Image and which versions are there
GPT Image is OpenAI's image generation model. Its main strength is following the prompt closely and rendering readable lettering, which is why people use it for posters, covers, infographics, product cards and interface mockups. It does two jobs: drawing from a text prompt and editing an existing image by instruction.
The studio currently has five cards across three generations:
- GPT Image 2.6 Sunburst The most precise version. It keeps the face, product and composition intact when an image is edited several times in a row, and it handles small text carefully.
- GPT Image 2.6 Flare The fast version of 2.6, focused on layout work such as posters, banners and ad mockups.
- GPT Image 2.5 Sunburst and GPT Image 2.5 Flare The previous series. Same strengths in text and edits, but no GIF or SVG.
- GPT Image 2 The earlier generation. Useful for reproducing an old result with the same engine; for new work, 2.5 or 2.6 is the better start.
Every version shares the same basic form: up to four references (JPG, PNG or WebP up to 30 MB), Auto plus eight aspect ratios (1:1, 3:2, 2:3, 16:9, 9:16, 4:3, 3:4, 21:9), and 1K, 2K or 4K resolution. At 4K a wide 16:9 frame comes out at 3840×2160 and a square at 2880×2880.
Which version to pick for the job
| Task | Model or version | Why |
|---|---|---|
| Editing your photo, portraits, a series of edits on one image | GPT Image 2.6 Sunburst | Faces and products don't drift from edit to edit |
| Poster, banner or product card with text | GPT Image 2.6 Flare | Stronger at layout and typography |
| Animated sticker or emoji | GPT Image 2.6, GIF format | 4 to 16 frames with adjustable speed |
| Icon or simple illustration as a vector | GPT Image 2.6, SVG format | Scales without losing quality |
| Lots of drafts in a row | GPT Image 2.5 Flare | A lighter tier for quick variations |
| Reproducing an older image with the same engine | GPT Image 2 | The previous generation stays in the catalog |
| Photorealism and complex edits with up to 8 references | Nano Banana Pro | Better at the "shot on a camera" look |
| Cutouts on a transparent background, up to 10 references | Seedream 5.5 | Has a transparent background setting |
For a closer look at how GPT Image stacks up against Google's model, see the GPT Image 2 vs Nano Banana Pro comparison; for the full list of image models, see the roundup of the best AI image generators.
How to use GPT Image 2.6: step by step
1. Open the model and choose a file format
Go to GPT Image 2.6 Sunburst or Flare. The File format setting defaults to PNG. JPEG gives a lighter file without transparency, WebP suits websites, and GIF and SVG switch the model into separate modes covered below.
2. Set the frame and resolution
Under Aspect ratio, pick the format for your platform: 9:16 for stories and Reels, 16:9 for a video thumbnail, 3:4 or 1:1 for a marketplace listing. With Auto, the model picks the proportions itself. Keep the resolution at 1K while you tune the prompt and move to 2K or 4K for the final: the cost grows with image area and is shown on the button.
3. Write the prompt from broad to specific
GPT Image responds best to a prompt built from five parts: setting → who or what is in the frame → details (materials, light, angle) → what the image is for → what to keep or leave out. Long prompts are fine.
A Scandinavian-style kitchen in the morning, a ceramic mug of coffee and a croissant on a wooden counter, soft window light from the left, shot from above at 45 degrees. A cover image for a food blog, no extra objects and no lettering
Concrete details beat adjectives. "85 mm lens, side light from a window" changes the picture; "masterpiece, ultra detailed" just makes the prompt longer.
4. Add text if you need it
Put the exact words in quotes and say where the text sits, in what typeface and at what size. This is where GPT Image shines, especially Flare.
A jazz festival poster: a large headline "JAZZ BY THE SEA" in the top third, a small line "July 12–14" below it, a saxophonist's silhouette against the sunset at the bottom, a strict grid, a warm two-color palette
One or two short lines come out more reliably than a paragraph. Check small text letter by letter before you print.
5. Upload references to edit
The Image slot takes up to four files. A simple pattern works for edits: what to change → into what → what to leave untouched.
Replace the blazer with a chunky navy knit sweater and make the background a brick wall. Keep the face, hair, pose and lighting exactly as in the reference
With several files, refer to them in upload order and give each one a role: "the product from image 1, the background from image 2, the style of image 3". Make one change per edit and run the next edit from the new result.
6. Press Generate
The finished image appears in the studio and in My generations; if you signed in with Telegram, it also arrives as a file in the bot. Not happy with it? Change one detail of the prompt instead of rewriting it, so you can tell what actually made the difference.
How to make a GIF sticker in GPT Image 2.6
The GIF format turns the model into a short animation generator: it draws a series of frames and loops them into a sticker. Three new settings appear:
- Frame size From 288×288 to 768×768, plus a vertical 288×512 and a horizontal 512×288.
- Frames 4, 6, 9, 12 or 16. More frames mean smoother motion and a more expensive run. At 768×768 you can use up to 12 frames.
- Frame duration From 20 ms to 2 seconds. Shorter means faster animation.
Describe a single simple action in the prompt, not a story: "a kitten waving its paw", "a cup of coffee with steam rising", "a pulsing heart". Several actions in one sticker look jerky. The loop length is frame count times frame duration, so 9 frames at 200 ms give an animation of almost two seconds. GIF also works with a reference, for example to turn your own character into a sticker.
When you need SVG
SVG is a vector file: you can stretch it to banner size or shrink it to an icon without losing sharpness, and open it in any vector editor. It fits icons, rough logo concepts, flat illustrations and interface elements.
The mode has two limits. SVG only works without references: the model draws the vector from scratch and can't convert a photo. And photography as a genre doesn't suit vectors, so ask for flat graphics, clean shapes and a limited palette.
A flat icon for a delivery app: an orange truck inside a circle, thick outline, three colors, no shadows or gradients
Common mistakes and how to fix them
| Mistake | Cause | What to do |
|---|---|---|
| Lettering with typos or warped glyphs | The text is too long or not in quotes | Keep one or two short lines in quotes, specify position and typeface |
| The face changed after an edit | The prompt didn't say what to keep, or the wrong version was used | List what must stay unchanged and edit in 2.6 Sunburst |
| The edit rewrote the whole image | Too many changes in one edit | Change one thing at a time, the next one from the new result |
| SVG won't start | A reference is uploaded in the slot | Remove the photo from the slot or switch to PNG |
| The GIF looks jerky | Several actions in the prompt or too few frames | One action, 9–16 frames, a frame duration of 100–200 ms |
| You wanted a transparent background and got white | The GPT Image form has no transparent background | Make the cutout in Seedream 5.5 or choose SVG |
| The image looks plastic, like a render | No concrete details about light, optics and texture | Add a light source, lens and materials; for photorealism try Nano Banana Pro |
| The crop isn't what you need | Auto is on and the model chose the proportions | Set the aspect ratio manually for your platform |
Ready-made GPT Image prompts
Marketplace product card (Flare, 3:4):
A product card: white wireless earbuds in the center on a light gray background, three short benefits on the right "40 hours of playback", "Noise cancelling", "Fast charging" in a clean sans-serif, a soft shadow under the product
Editing a product shot from a reference (Sunburst):
The same bottle as in the reference, placed on a marble countertop, morning light from a window on the right. Don't change the label, its lettering or the bottle shape
Post cover (any version, 1:1):
A post cover with a large headline "5 BEGINNER MISTAKES", flat geometric shapes, two colors, plenty of empty space around the text
More examples for each version are on the model pages: GPT Image 2.6 Flare and GPT Image 2.5 Sunburst.
Don't want to fiddle with prompts?
Ready-made scenes: upload your photo and get the result in one click.
Models in this article
Frequently asked questions
How do I use GPT Image without a ChatGPT account?
Open GPT Image 2.6 Sunburst or Flare in the IIshki studio, write a prompt, pick an aspect ratio and a resolution, and press Generate. To edit your own photo, attach up to four images. You don't need a ChatGPT subscription: sign in with email or a social account.
Can I use GPT Image for free?
Picking a model, format and settings in the studio costs nothing; tokens are spent only on the generation itself. The cost of each run is shown on the Generate button before you press it and depends on the version, resolution and file format.
How is GPT Image 2.6 different from 2.5 and 2?
All three generations render text in images and edit pictures from a reference at up to 4K. Version 2.6 adds a file format choice: besides PNG, JPEG and WebP you can get an animated GIF sticker or an SVG vector. In 2.6, Sunburst holds faces and products more steadily across a series of edits, while Flare is stronger at poster and product card layouts.
Flare or Sunburst: which version should I pick?
Pick Sunburst for edits from a reference, portraits and images with small text, since it is more precise. Pick Flare for posters, banners, product cards and whenever you want to try many variants quickly. Within one generation both versions cost the same.
How do I get a transparent background in GPT Image?
The GPT Image form has no transparent background switch. If you need a cutout, such as a sticker or a product for a layout, use Seedream 5.5: it has a transparent background setting. For a clean vector without a background, the SVG format in GPT Image 2.6 works too.
Every model in this roundup is available in the IIshki studio at reduced club prices. Club members get tokens for generations every month.
Join the clubMore articles
- AI Photo Prompts: 30 Ready-to-Use Prompts for an AI PhotoshootAI photo prompts that work: how a strong photo prompt is built, plus 30 ready examples for photoshoots, portraits, avatars, photo editing and products.
- Nano Banana Pro reviews: what it actually doesNano Banana Pro reviews: what holds up — photorealism, text in images, face consistency, edits from a photo — and where the model still slips. Try it online.
- AI Greeting Card and Video Greeting: New Year and Birthday GuideHow to make an AI greeting card and a video greeting for New Year or a birthday: a card with readable text, a talking photo, a festive clip, prompts.