The Complete GPT Image 2.5/2.6 Prompt Collection: 259 Creator-Grade Prompts, Organized by Task
The Complete GPT Image 2.5/2.6 Prompt Collection: 259 Creator-Grade Prompts
Great images are not produced by the model alone — they are produced by the prompt. OpenAI's GPT Image series renders reliably, but what actually separates a good result from a generic one is prompt quality: lighting, lens, material, composition, and identity consistency all have to be written precisely. The community has now curated a complete set of 259 copy-ready prompts drawn from 52 public creators, released under CC BY 4.0. This guide covers what the collection is, how it is organized, and how to wire it to a per-image-billed image API so the prompts actually become images.
Model availability follows the live catalog
The prompt collection references GPT Image model names from its upstream maintainer (EvoLink). Whether a specific tier is callable on EXAIV depends on the live /models catalog — check there before hard-coding an ID. If your chosen tier is unavailable, substitute another image model from the catalog; the prompts themselves are model-agnostic.
The short version
Copy any prompt, point base_url at https://exaiv.com/api/ai/v1, set model to a live image ID from /models, and call images/generations once to get an image. Swap the size and model freely; it is billed per image (live pricing on /pricing).
What the collection actually is
The collection is published as awesome-gpt-image-2-6-prompts-and-api and maintained by EvoLink. Its point is not theory — it lays out things you can use directly: 250 GPT Image 2.5 prompts plus 9 new GPT Image 2.6 cases, 259 in total. Each entry includes:
- The full prompt text (the complete version, not a summary — copy and run it)
- Creator attribution (X/Twitter handle) and the original post link
- Whether a reference image is required (many editing prompts start from a base image)
- A matching preview or result image
The attribution detail matters. The collection states plainly that brand, character, and celebrity examples are individual creator concepts, not official campaigns or endorsements, and that original prompts and assets belong to their respective owners. So when you reuse or adapt these prompts, keep the "author + source" credit — that is also a hard requirement of CC BY 4.0.
One more thing worth knowing: the prompts are not independently regenerated by the curators. They are transcribed from the creators' published posts, and the model labels follow the creators' own statements. Reference images, local settings, and later edits can all shift the result, so a copied prompt is a strong starting point, not a pixel-perfect guarantee. Treat the collection as a verified library of working ideas, then tune from there.
What CC BY 4.0 means
You may copy, remix, and use it commercially, but you must credit the original creator and note if you modified it. We kept the creator links on the tool page and at the bottom of this post for exactly that reason. Treat it as a usable inspiration library, not a free asset dump.
GPT Image 2.6: Sunburst and Flare
On November 4, 2026, OpenAI released two GPT Image 2.6 variants — Sunburst and Flare. They are not a simple upgrade; they split into two different workflows:
| Variant | Positioning | Best for |
|---|---|---|
| Sunburst | Detail-heavy, edit-preserving | Refinement, repeated local edits, keeping the original structure intact |
| Flare | Faster iteration | Producing multiple versions quickly, trying different styles, speed-first work |
2.6 also added a few capabilities: complexity control (separate knobs for "planning effort" and "rendering quality"), markup-guided edits (specify a region and apply a local change), and SVG / GIF output (vector and animated results, not just bitmaps). One caveat: complexity and rendering quality are independent controls. Raising planning effort can improve detail, but it also raises latency and can introduce changes you did not ask for.
In practice the choice between the two is straightforward. Use Flare when you are exploring — trying five color variants, a different mood, or a quick storyboard — because you will iterate many times and each run should be cheap and fast. Use Sunburst when the image already exists and you only want to nudge it: keep the face, keep the composition, change the jacket. The recurring theme in the 2.6 cases is exactly this split — orbit and stretch animations favor Flare's speed, while the flower-removal and billboard-winter edits favor Sunburst's edit preservation.
The seven-category map
The 259 prompts split into seven creative task categories, plus the nine new 2.6 cases. The distribution looks like this:
| Category | Count | Typical output |
|---|---|---|
| Portraits & Photography | 51 | Cinematic portraits, candid lifestyle, magazine collages |
| Character Sheets | many | Consistent character turnarounds, expression and pose sheets |
| Illustration | many | Storybook, flat, narrative illustration |
| Advertising | many | Product posters, brand visuals |
| Travel Journals | many | Travel illustrations, location mood pieces |
| Image Editing | many | Object removal, sky swap, local condition changes |
| Sequential Art | many | Multi-panel comics, continuous narrative |
| GPT Image 2.6 cases | 9 | Orbit animation, markup edits, GIF characters |
The counts above are approximate where marked "many" because several categories still grow as the curators add cases; the precise per-category totals are shown on the tool page, where you can filter by category directly. The point of the map is to give you a mental index before you open the collection, so you search within a lane instead of scrolling 259 entries blind.
Don't judge by category name alone
Styles vary enormously inside one category. Portraits range from 9:16 cinematic waterfall shots to CCD-textured street snapshots. Check the preview image and the author's notes before deciding a prompt fits — applying an unrelated prompt blindly tends to drift.
Turning a prompt into an image
The collection gives you the recipe; what is missing is the stove. Here is the minimal call to run a prompt through EXAIV. EXAIV exposes an OpenAI-compatible image endpoint; just put a catalog model in the model field, it is billed per image, and the exact credits per image are shown live on /models.
curl https://exaiv.com/api/ai/v1/images/generations \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "<live-image-id-from-/models>",
"prompt": "paste the full prompt text you copied here",
"size": "1024x1024",
"n": 1,
"response_format": "url"
}'
A few points:
- Model: image model availability follows the live /models catalog — check there before writing code. The upstream prompt library references flexible-size tiers and 2K / 4K variants; the 2K tier supports 2048×2048, 2048×1152, 1152×2048, 2048×1365, 1365×2048; the 4K tier is bound by the upstream pixel cap, so 4096×4096 is unavailable.
- Size:
sizemust match the tier the model supports, or the call errors. - Async: images can also run as an async job —
generationsreturns animage.job(statusqueued) first, then pollGET /api/ai/v1/images/jobs/JOB_IDfor the result. Async is steadier for large images or when the queue is busy. - Edits: if a prompt needs a reference image (many editing prompts do), use
POST /api/ai/v1/images/edits, pass the base image with-F "image=@local.png", and describe the change inprompt.
A reference image is not optional
Several editing prompts in the collection (for example "swap the background to a seaside sunset" or "remove the flower from the man's hand") require a base image first; the prompt itself assumes one exists. Running these through generations without a reference produces meaningless output — always check the author's input_requirement.
Prompt techniques worth reusing
After reading 259 prompts, a few patterns recur and are worth pulling out. Note these are fragments; the full text is on the tool page or the source repo:
- Identity consistency: portrait prompts almost always say "use the uploaded image as the primary visual reference and preserve the exact facial identity, proportions, and defining features." Keeping identity while changing outfit, scene, or lighting is the whole trick — write "preserve identity" explicitly.
- Minimal edit instructions: OpenAI's official editing examples are often a single sentence, such as "Remove the flower from man's hand. Do not change anything else." Stating "don't change anything else" beats writing a wall of constraints.
- Lens and medium: "IMAX-level cinematic," "CCD texture," "low angle," "backlit" decide the photographic "gear feel" — they are the switch for output texture.
- Ratio and composition: write the ratio directly (e.g., 9:16 vertical) and the composition (profile, low angle, frame-in-frame); it is far more controllable than a vague "make it pretty."
Where to start if you are new. Don't open the collection at random. Decide the category first — portrait, illustration, or editing — then sort by the result you actually want and copy the closest match. Render it unchanged once to learn the model's baseline for that prompt, then change one variable at a time: size, then lens words, then a single sentence about mood. This keeps failures diagnosable instead of a pile of unrelated bad outputs.
Copy first, then iterate
The most effective path for beginners: pick a prompt in the same category whose preview is close to what you want, render it as-is once, then adjust parameters and wording step by step. Prompts are tuned, not written correctly on the first try.
Attribution and compliance notes
- The prompt collection is CC BY 4.0; reusing or adapting it requires crediting the creators and linking the source.
- Brand, character, and celebrity examples in the collection are individual creator concepts, not official campaigns or endorsements.
- EXAIV provides an OpenAI-compatible image endpoint and billing; it is not affiliated with the prompt creators.
FAQ
Are the prompts model-specific?
Can I use these prompts for commercial projects?
How do I get better results?
Related reading
For the full 259-prompt set with category filtering and one-click copy, see our online tool page; for each model's tiers and per-image price, see /models. The fastest way to learn the style is to open the tool, copy a prompt whose preview matches your goal, and render it once before changing anything.