How to Use GPT Image 2.6: Features, Settings & Prompts

A useful first image is a small brief: one subject, one setting, one composition. Asking for a product photograph, a poster, a new background, and an animated version in the same request makes it hard to tell which instruction failed.

This guide starts with a portrait, moves through a focused reference edit, and finishes with a designed poster. You will learn which settings to choose, how to write a brief, and what to inspect before spending more credits.

Open the GPT Image 2.6 generator to follow along. It supports text-to-image and reference editing, with Flare and Sunburst options. Sign in, choose a mode, and check the credit total before generating.

What GPT Image 2.6 does

GPT Image 2.6 creates images from a written brief and edits uploaded references. Useful outputs include portraits, product photographs, posters, website banners, transparent product assets, simple vector concepts, and short animated loops.

Flare is a practical starting point for quick iteration. Sunburst is an option to compare when fine details and edit precision matter. For either version, inspect the result: faces can drift, lettering can miss a character, and a requested edit can affect surrounding details.

Which features can I use in the generator?

The exercises below choose Flare and PNG. Start there, then change one control at a time. A separate planning-complexity setting and a dedicated markup brush are not part of this workspace.

Choose text-to-image or reference editing

Use text-to-image when the subject can be invented. A fictional portrait, an imaginary storefront, or a conceptual editorial scene does not need a reference to be useful.

Use image-to-image when identity matters. If the brief says “this person” or “this product,” supply its photograph. Describe the face, clothing, or product shape as details to preserve. Describing them from memory leaves the model to invent a replacement.

With several references, do not write “combine these.” Decide which image supplies the product, which supplies the palette, and which supplies the composition. Conflicting product views or lighting can make the result harder to judge.

Author-created workflow diagram. It is not a screenshot of the generator.

Start with these settings

My suggested first pass is medium quality, 1K resolution, one image. Choose the aspect ratio for the destination before submitting. This is an economical review setup, not an official quality benchmark.

The workspace offers auto, 1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16, 21:9, 27:16, 16:27, 9:8, 8:9, 4:5, and 5:4. A square product tile and a wide website banner need different framing; changing only the crop after generation can remove the shoulders, product details, or copy area.

Resolution and quality answer different questions. Resolution selects an output tier. Quality changes rendering effort and cost in this workspace. Neither repairs an incorrect product or an ambiguous brief.

Quality is not a separate planning-complexity switch. This workspace exposes quality and resolution; typing complexity: low into the prompt does not create another control. Put your effort into a clear visible change before raising quality.

The labels 1K, 2K, and 4K are resolution tiers, not a promise of one exact width and height across all aspect ratios. Inspect the downloaded file's actual dimensions when a placement has a strict pixel requirement.

Step 1: Create your first image

Open the GPT Image 2.6 workspace. Check the displayed model, sign in when prompted, and choose text-to-image. Choose Flare, PNG, medium quality, 1K, one image, and 2:3 for this exercise.

Copy this brief:

TASK: Create a professional editorial portrait of a fictional adult woman.
SUBJECT: About 32 years old, short dark curly hair, medium-brown skin,
a plain navy blazer over an ivory crew-neck top, relaxed expression.
SCENE: A plain warm-gray studio background.
COMPOSITION: Head and shoulders, both shoulders visible, comfortable headroom.
LIGHT: Soft light from the left, gentle shadow on the right.
TEXT: No text, lettering, labels, or logos.
CHECK: Natural face proportions, clear eyes, believable skin texture.

Portrait brief illustration · ImageGen.

Check the total credits before clicking Generate. Inspect the face, eyes, hair edges, shoulders, and background. A pleasing thumbnail can hide an uneven iris or an unnatural edge around the hair.

If the framing is wrong, change the relevant line before raising resolution. For example, replace “head and shoulders” with “a waist-up portrait with both arms inside the frame.” This gives the next request a visible target.

Save the accepted draft with its prompt. The next edit needs the actual image file; describing a similar person does not supply the same reference.

Step 2: Edit a reference without rewriting the whole scene

Switch to image-to-image and upload the accepted image. The workspace accepts JPEG, PNG, or WebP references, with up to 16 images. Keep your first edit to one reference and one requested change.

Use the upload order to name references: @Image1, @Image2, and so on. A reference token does not fetch a missing file. If your prompt names @Image2 but you supplied only one image, add the missing reference or correct the token.

TASK: Edit @Image1.
CHANGE: Change only the navy blazer to muted terracotta orange.
PRESERVE: The person's face, skin tone, hairstyle, expression, pose,
shirt, camera angle, background, light direction, and framing.
TEXT: No text, lettering, labels, or logos.
CHECK: The blazer color changes; the person remains recognizable.

A preservation instruction is a requirement to inspect. Compare the source and edit at the same size. If the eyes, jawline, expression, or hairstyle change, the edit has failed the identity requirement even if the new jacket color looks good.

For a real person's portrait, use a reference you have permission to edit. Keep the original alongside the edit so the person can review what changed.

Assign roles when you add a second image

Use a second reference when it explains something the first cannot:

TASK: Create a portrait using two references.
SUBJECT: Use the person in @Image1 as the identity reference.
SCENE: Use @Image2 only for the neutral backdrop and soft light palette.
COMPOSITION: Head and shoulders, comfortable headroom, eyes toward camera.
PRESERVE: The face, skin tone, hairstyle, expression, and clothing from @Image1.
EXCLUDE: Do not copy people, objects, or lettering from @Image2.

Upload files in the order you describe. If the second picture contains another face, explicitly exclude it or choose a cleaner backdrop reference. Extra images should resolve ambiguity rather than compete to define the subject.

Step 3: Make an image with readable text

Switch back to text-to-image so the previous references are not used. Choose Flare, PNG, and 2:3. This exercise creates a new poster; it does not reuse the portrait from Step 2.

Treat text as copy with a location. Give the model the exact spelling, a hierarchy, and a short instruction about everything else.

TASK: Create a vertical editorial poster.
SUBJECT: A cobalt paper arch and a terracotta sphere on a cream plinth
in the lower half.
COMPOSITION: Wide margins; a clean headline area in the upper half.
STYLE: Minimal editorial photography, warm paper, dark charcoal type.
TEXT: Headline exactly "MAKE SOMETHING" on two lines.
TEXT: Small supporting line exactly "CREATIVE WORKSHOP" below the headline.
EXCLUDE: No other words, labels, logos, or decorative letterforms.
CHECK: Both text strings readable, paper sculpture visible, no overlap.

Layout illustration · ImageGen background with separately typeset text.

Choose 2:3 for this poster. Read the output letter by letter, including small text. Check line breaks and punctuation independently from whether the composition looks balanced.

For a spelling error, try an edit that names the incorrect line and supplies the exact replacement. If exact copy keeps drifting, generate a clean visual with empty space and add the final type in your design file. The PNG remains a flattened image; a convincing poster is not an editable layout document.

The poster prompt collection offers more layout ideas. Its examples were made with GPT Image 2, so use the structures as inspiration rather than evidence of 2.6 performance.

Use advanced outputs only when the deliverable needs them

The interface also lets you choose Sunburst, transparent backgrounds, exact dimensions, SVG, and animated GIF. Finish the first PNG exercise before changing several controls together.

For SVG, try a small text-only brief: A simple flat vector icon of one leaf, clean silhouette, two solid colors, no text. SVG generation does not turn an uploaded product photo into a faithful vector trace, because references are not accepted for that format.

For a first GIF experiment, enter 512x512 in Custom size (pixels), then choose 8 frames and 100 milliseconds per frame, then describe a simple movement: A simple illustrated leaf drifting gently from side to side in a seamless loop. Keep the background still. No text. These are valid example settings, not a measured animation result. The controls accept 2–64 frames and 20–2000 milliseconds per frame. Inspect the transition back to the first frame and whether the subject deforms.

GIF and SVG use 1K only. Raster exact sizes also have pixel-budget limits for the chosen resolution: a technically valid width is not enough if width multiplied by height exceeds that tier. Follow the validation message beside the size fields rather than guessing that every 4K-shaped rectangle is accepted. Selecting a format or variant does not change the per-image credit schedule; quality, resolution, and count do.

How much does GPT Image 2.6 cost?

This site's per-image credit schedule, checked November 5, is:

Multiply the per-image price by the requested count. Medium/1K with four images is 8 × 4 = 32 credits; high/4K with four is 108 × 4 = 432. These are this site's generation credits. The total shown in the generator is the amount to check before submission.

Failed jobs return credits; missing outputs receive a proportional credit refund. A completed image that fails your aesthetic brief is different from a failed job; do not assume it triggers an automatic refund. If you need help, keep the job reference and the settings with your report.

Check the result in four passes

1. Identity: Is it the intended subject? Compare the face, clothing, or product outline with the reference, depending on your brief.
2. Layout: Are the margins, crop, copy area, and subject position usable for the actual placement?
3. Details: Read every word; inspect faces, fingers, repeated objects, edges, and contact shadows.
4. File: Check format, real dimensions, and the image at its intended display size.

Do not count a larger file as a better result until these checks pass. For continuity across several frames, the camera shot guide explains how to change framing while keeping story details fixed. Its image examples are from the earlier GPT Image 2 workflow.

Common problems and what to change

A practical stopping rule is to change method when repeated attempts keep failing the same requirement. For example, after two unsuccessful corrections to exact lettering, finish the lettering in a design tool. That is a suggested workflow limit, not a measured model threshold.

FAQ

Where can I use GPT Image 2.6 online?

Use the GPT Image 2.6 generator. Sign in, select text-to-image or image-to-image, choose Flare or Sunburst, and generate with your account credits. It runs in the browser; no desktop installation is required.

Is GPT Image 2.6 free?

Generation uses credits. A medium-quality 1K image costs 8 credits; one low-quality 1K image costs 2. Check your account balance and any current offer in the workspace. Do not assume unlimited free generations. Selecting more images increases the total.

Which should I choose: Flare or Sunburst?

Start with Flare while refining the brief. Compare Sunburst on the same input if small details or a precise reference edit matter most. Keep quality, resolution, and output count unchanged while comparing so you can judge the difference.

Is high quality the same as high complexity?

No. Quality is a rendering control. Complexity concerns planning before rendering; this workspace does not expose it separately.

Does writing “transparent background” create a transparent PNG?

Choose PNG and set Background to Transparent in the controls. JPEG cannot carry transparency. A prompt alone is not proof of an alpha channel: inspect the downloaded file over a contrasting background. The site's transparent background workflow is another option for a cutout task.

Can I write prompts in Chinese?

You can enter an English or Chinese brief. State the subject, composition, and change explicitly, and quote any text you want inside the image. If the image needs another language, give that exact wording rather than asking the model to invent a translation.

How do I make a GIF with GPT Image 2.6?

Choose GIF and 1K, fill in the custom pixel size, then set the frame count and milliseconds per frame. Start with a simple movement and inspect the whole loop. The advanced-output section above includes a valid 512×512 example.

Where should I start for a real project?

Pick one deliverable, one destination ratio, and one acceptance check. Start in the generator, and use the GPT Image 2.6 ecommerce use cases to turn an accepted image into a small campaign set.

Settings and credit examples checked November 5, 2026. Illustrations explain the briefs and were made with ImageGen; they are not GPT Image 2.6 test results. The displayed credit total is the amount to check before each request.