OpenAI Image Editor: Edit Any Photo With GPT Image 2.6
A step-by-step walkthrough of GPT Image 2.6 Edit on each::labs: uploading, writing the instruction, Sunburst vs Flare, size, transparency, complexity, multi-image edits and 10 prompts you can paste in.
By Eftal Yurtseven · Nov 20, 2026 · 9 min read

Generating a picture from nothing is the easy part now. Fixing the picture you already have is where people give up: the product shot with a coffee cup in the corner, the portrait lit like a parking garage, the banner that needs one word changed and nothing else. That's why people search for an openai image editor and not another generator. They don't want a new image. They want this one, better.
OpenAI's answer is GPT Image 2.6 Edit, and on each::labs you can run it in a browser or through one API call. This guide walks the whole job in order: upload, instruction, model choice, the settings that actually matter, and ten edit prompts you can paste straight in. No theory about diffusion. Just the steps that get you from "almost" to "done".
What the OpenAI Image Editor Actually Is on each::labs
OpenAI released GPT Image 2.6 on November 4, 2026, in two flavors, Sunburst and Flare, both built for generation and editing. Its API changelog lists adjustable complexity as one of the new controls, and that dial shows up on each::labs too. Each flavor gets its own edit page here: Sunburst Edit and Flare Edit. If you landed on the older OpenAI Image Edit page first, the 2.6 pages are the current generation of the same idea. The launch itself (GIFs, SVGs and markup edits) has its own write-up in GPT Image 2.6: GIFs, SVGs and Edits You Can Draw On, so this post sticks to editing a photo.
The mental model is simple. You give it one or more source images and a sentence describing the change. It returns an edited image that keeps the subject, structure and composition you didn't ask it to touch. That last part is the whole product. An editor that "improves" everything you didn't mention isn't an editor, it's a reroll.
Two inputs are required: prompt and image_urls. Everything else has a sensible default, which means your first edit can be two fields and a click. The rest of this guide is about when to stop accepting the defaults.

Step 1: Upload the Photo You Want to Fix
Open the edit page, drop your image into the image field, and you've filled half the request. Under the hood it lands in image_urls, a list that takes at least one image and up to 16. For a plain edit, one is all you need.
Pick your source carefully, because the model keeps what it sees. A sharp, well-exposed photo gives it real detail to preserve. A blurry screenshot of a screenshot gives it mush to preserve faithfully. If you have a higher resolution original, use that one. And crop before you upload if half the frame is irrelevant: the model reads the whole image, and stray clutter at the edges is stray clutter it might try to honor.
One habit worth building early: keep the untouched original next to every result. Edits stack, and it's easy to drift three rounds deep and forget what the photo looked like when it was still right.

Step 2: Write the Edit Like a Note to a Retoucher
The prompt field takes up to 32,000 characters. Ignore that number. The best edit instructions are short, specific and slightly bossy, the way you'd brief a human retoucher who hates guessing.
Good instructions do three things. They name the change ("replace the grey sky with a clear late afternoon sky"). They say what must stay put ("keep the building, the people and the reflections in the windows exactly as they are"). And they describe the result in visual terms, not adjectives: "warm side light from the left" beats "make it nicer" every time.
Change one thing per round when precision matters. Pile five edits into one sentence and, when one goes wrong, you won't know which instruction caused it. Do them in sequence, feed each result back in as the new source, and you keep control. For a deeper set of prompt patterns across generation and editing, our GPT Image 2.6 prompts post goes further than this one needs to.
Step 3: Pick Sunburst or Flare Before You Type a Word
This is the choice that changes your day most, so make it on purpose. OpenAI describes Sunburst as built "for detailed visual work" and Flare as built "for fast iteration" in interactive editing workflows. Both edit pages on each::labs share the exact same parameters, so switching is free of rework: same prompt, same settings, different engine.
Here's the opinionated version. Use Flare Edit while you're still figuring out what you want. Background options, rough relights, five versions of a product angle: speed matters more than the last few percent of fidelity there, and fast feedback keeps you deciding instead of waiting. Switch to Sunburst Edit once the direction is locked and the image is going somewhere people will look closely: a hero banner, a packshot, a face.
Draft on Flare, finish on Sunburst. It's not a rule OpenAI wrote, but it's the workflow both model descriptions point toward.

Step 4: Set Size, Background and Format
The image_size default is 1024x1024. Change it whenever your output has a destination. The presets cover square (1024x1024, 2048x2048), landscape (1536x1024, 2048x1152, 3840x2160) and portrait (1024x1536, 2160x3840). There's also auto, which lets the model choose, and custom for your own WIDTHxHEIGHT. Custom sizes need both edges in multiples of 16, a longest edge of 3840px at most, an aspect ratio between 1:3 and 3:1, and a total between 655,360 and 8,294,400 pixels. Anything above 2560x1440 is marked experimental, so don't make 4K your first attempt at a tricky edit.
The background setting is the one most people miss. Set it to transparent and you get a cutout ready for a slide, a storefront or a layered design. The catch is the format: transparency only works with output_format set to png or webp. PNG is the default, so you're fine unless you switched. If you don't need transparency, opaque forces a solid background and auto lets the model decide.
Speaking of format, jpeg is faster than PNG, and output_compression (0 to 100) lets you shrink JPEG and WebP files for the web. It's ignored for PNG.
Step 5: Turn the Complexity and Quality Dials
GPT Image 2.6 has two dials that sound alike and aren't. complexity controls how much the model plans before it renders: low, medium or high, with high used when you leave it empty. quality controls the rendered output: low for fast drafts, medium or high for final assets, xhigh or max for maximum detail with extra latency, plus auto as the default.
Think of complexity as how hard it thinks and quality as how finely it draws. A simple "remove the logo from the mug" doesn't need deep planning, so drop complexity to low and save the wait. A multi-image composite with text, lighting and three people does need it, so leave complexity high. Quality follows the destination: low while drafting, high or above for the file you'll ship.
Finally, num_images returns 1 to 10 variations in one request. On Flare, asking for four at once is the fastest way to see whether your instruction is clear. If all four agree, the prompt works. If they scatter, rewrite before you scale up.

Step 6: Feed It More Than One Photo
The quiet superpower here is that image_urls takes up to 16 images. That turns the editor from a retouching tool into a compositing tool: put a product into a lifestyle scene, dress a model in a garment from a flat lay, or build a group shot from separate portraits.
The trick is to reference images by their order in your prompt. "Place the bottle from the first image on the marble counter in the second image" is unambiguous. "Combine these" is a coin toss. Tell it which image supplies the subject, which supplies the setting, and which, if any, supplies the style or light. When identity matters, say so outright: "keep the face from the third image unchanged".
Multi-image edits are where Sunburst and high complexity earn their keep. There's more to reconcile, so give the model room to plan.
10 Copy-Ready OpenAI Photo Editor Prompts
These are prompts to try, written for the edit pages as they stand. Swap the bracketed details for your own and keep the "keep unchanged" clauses: they do more work than they look like.
1. Change one thing
Change the color of the jacket to deep forest green. Keep the person, pose, face, background and lighting exactly the same.
2. Remove an object
Remove the coffee cup on the left side of the table and fill the space with the same wooden surface and natural shadows. Change nothing else.
3. Swap the background
Replace the office background with a sunlit Mediterranean terrace, soft focus. Keep the person, clothing and framing unchanged, and match the light on the face to the new scene.
4. Relight the scene
Relight this portrait with warm golden hour light coming from the left, gentle shadow on the right side of the face, and a soft rim light on the hair. Keep the expression and features identical.
5. Add text
Add the words "SUMMER SALE" in bold white sans-serif letters across the top third of the image, centered, with a subtle drop shadow. Do not alter the product or background.
6. Product cleanup
Clean up this product photo: remove dust and fingerprints from the glass bottle, straighten the label, even out the white background, and add a soft contact shadow under the base.
7. Transparent cutout
Isolate the sneaker on a fully transparent background. Keep the laces, stitching and sole texture sharp, and remove every trace of the original floor.
Pair this one with background set to transparent and PNG or WebP output.
8. Combine images
Place the perfume bottle from the first image on the stone ledge in the second image. Match the scene's late afternoon light and add a natural reflection on the ledge. Keep the bottle's shape and label exactly as in the first image.
9. Restyle without losing the subject
Turn this photo into a watercolor illustration with loose brush strokes and visible paper texture. Keep the composition, the dog's markings and its pose recognizable.
10. Seasonal swap
Turn this summer street scene into a snowy winter evening: snow on the rooftops and pavement, warm light glowing from the shop windows, light snowfall. Keep every building and the street layout unchanged.
Run the Same Edit From Code
Once a prompt works in the browser, moving it into a pipeline is one request. You post the model slug and the same inputs you set on the page to the each::labs prediction endpoint, then fetch the result by its prediction ID (or pass a webhook_url and let it come to you).
POST https://api.eachlabs.ai/v1/prediction
{
"model": "gpt-image-v2-6-sunburst-edit",
"input": {
"prompt": "Remove the coffee cup on the left and fill the space with the same wooden surface. Change nothing else.",
"image_urls": ["https://example.com/table.jpg"],
"image_size": "1536x1024",
"quality": "high",
"complexity": "low",
"output_format": "png",
"num_images": 1
}
}Swap the slug to gpt-image-v2-6-flare-edit for the fast variant. Nothing else in the body changes. If an edit is part of a bigger job (generate a base with GPT Image 2.6 Sunburst text-to-image, edit it, then pass it on), chain the steps in each::labs flows instead of wiring it by hand. And if you want to compare it with other image models, the call shape stays the same.
Frequently Asked Questions
Is there an OpenAI image editor I can use without writing code?
Yes. The Sunburst Edit and Flare Edit pages on each::labs run in the browser: upload a photo, type the instruction, adjust the settings and generate. The API is there for when one edit turns into a thousand.
Should I use Sunburst or Flare for photo editing?
Start where the work is. OpenAI pitches Flare at fast iteration and Sunburst at detailed visual work, so explore on Flare and render the keeper on Sunburst. The inputs are identical, so moving a prompt between them takes seconds.
How many images can GPT Image 2.6 Edit combine?
Up to 16 source images in one request. Refer to them by order in your prompt ("the first image", "the second image") so the model knows which one supplies the subject and which supplies the scene.
Can it give me a transparent background?
Set background to transparent and keep output_format on PNG or WebP. JPEG can't carry transparency, so the request needs one of those two formats.



