GPT-Image-2.6 Flare

Introduction

GPT-Image-2.6 Flare is OpenAI's fast model for generating and editing images in interactive work. It takes up to 32,000 characters of positivePrompt and plans the layout before it renders. A prompt written as a brief, with the subject, where each element sits, how many there are and the exact words in quotes, gives that plan something to follow.

An e-commerce product hero for the website of a running shoe brand. A single low-top running sneaker in off-white knit with a sage green heel tab and a gum rubber sole stands in three-quarter view on a pale concrete plinth, centered in the lower half of the frame. A kraft paper price tag hangs from the laces on a cotton string and reads "$129" in black stamped type. Across the upper third, on a soft sage green backdrop, the headline "LOOPWAY RUNNER" is set in large white sans-serif capitals. Soft diffused daylight from the upper left, a gentle shadow falling to the right of the plinth. Straight-on shot at plinth height, 85mm lens. Photoreal studio product photography, no other text.

This guide covers the request, writing a brief the model can plan from, counts and positions, readable text, and choosing a size.

The request

A call needs only model and positivePrompt, and the size is optional too.

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'openai:gpt-image@2.6-flare',
  positivePrompt: 'An e-commerce product hero for the website of a running shoe brand. A single low-top running sneaker in off-white knit with a sage green heel tab and a gum rubber sole stands in three-quarter view on a pale concrete plinth, centered in the lower half of the frame. A kraft paper price tag hangs from the laces on a cotton string and reads "$129" in black stamped type. Across the upper third, on a soft sage green backdrop, the headline "LOOPWAY RUNNER" is set in large white sans-serif capitals. Soft diffused daylight from the upper left, a gentle shadow falling to the right of the plinth. Straight-on shot at plinth height, 85mm lens. Photoreal studio product photography, no other text.',
  width: 2048,
  height: 1152
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "openai:gpt-image@2.6-flare",
            "positivePrompt": "An e-commerce product hero for the website of a running shoe brand. A single low-top running sneaker in off-white knit with a sage green heel tab and a gum rubber sole stands in three-quarter view on a pale concrete plinth, centered in the lower half of the frame. A kraft paper price tag hangs from the laces on a cotton string and reads \"$129\" in black stamped type. Across the upper third, on a soft sage green backdrop, the headline \"LOOPWAY RUNNER\" is set in large white sans-serif capitals. Soft diffused daylight from the upper left, a gentle shadow falling to the right of the plinth. Straight-on shot at plinth height, 85mm lens. Photoreal studio product photography, no other text.",
            "width": 2048,
            "height": 1152
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "imageInference",
      "taskUUID": "6e1c9a4f-2b87-4d3e-9f05-8a7d3c2b1e64",
      "model": "openai:gpt-image@2.6-flare",
      "positivePrompt": "An e-commerce product hero for the website of a running shoe brand. A single low-top running sneaker in off-white knit with a sage green heel tab and a gum rubber sole stands in three-quarter view on a pale concrete plinth, centered in the lower half of the frame. A kraft paper price tag hangs from the laces on a cotton string and reads \"$129\" in black stamped type. Across the upper third, on a soft sage green backdrop, the headline \"LOOPWAY RUNNER\" is set in large white sans-serif capitals. Soft diffused daylight from the upper left, a gentle shadow falling to the right of the plinth. Straight-on shot at plinth height, 85mm lens. Photoreal studio product photography, no other text.",
      "width": 2048,
      "height": 1152
    }
  ]'

runware run openai:gpt-image@2.6-flare \
  positivePrompt="An e-commerce product hero for the website of a running shoe brand. A single low-top running sneaker in off-white knit with a sage green heel tab and a gum rubber sole stands in three-quarter view on a pale concrete plinth, centered in the lower half of the frame. A kraft paper price tag hangs from the laces on a cotton string and reads \"\$129\" in black stamped type. Across the upper third, on a soft sage green backdrop, the headline \"LOOPWAY RUNNER\" is set in large white sans-serif capitals. Soft diffused daylight from the upper left, a gentle shadow falling to the right of the plinth. Straight-on shot at plinth height, 85mm lens. Photoreal studio product photography, no other text." \
  width=2048 \
  height=1152

{
  "taskType": "imageInference",
  "taskUUID": "6e1c9a4f-2b87-4d3e-9f05-8a7d3c2b1e64",
  "model": "openai:gpt-image@2.6-flare",
  "positivePrompt": "An e-commerce product hero for the website of a running shoe brand. A single low-top running sneaker in off-white knit with a sage green heel tab and a gum rubber sole stands in three-quarter view on a pale concrete plinth, centered in the lower half of the frame. A kraft paper price tag hangs from the laces on a cotton string and reads \"$129\" in black stamped type. Across the upper third, on a soft sage green backdrop, the headline \"LOOPWAY RUNNER\" is set in large white sans-serif capitals. Soft diffused daylight from the upper left, a gentle shadow falling to the right of the plinth. Straight-on shot at plinth height, 85mm lens. Photoreal studio product photography, no other text.",
  "width": 2048,
  "height": 1152
}

[
  {
    "taskType": "imageInference",
    "taskUUID": "6e1c9a4f-2b87-4d3e-9f05-8a7d3c2b1e64",
    "imageUUID": "b3f8d2a6-5c19-4e70-a4b1-7d6e9f0c2a58",
    "imageURL": "https://im.runware.ai/image/os/a14d18/ws/2/ii/b3f8d2a6-5c19-4e70-a4b1-7d6e9f0c2a58.jpg"
  }
]

The rest of the surface, with the full parameter list in the model reference:

• width and height set the frame inside an envelope of four rules, and leaving both out hands the choice to the model. Both cases are covered at the end of this guide.
• settings.quality sets the rendering effort and settings.thinkingLevel how much the model plans before it renders, high by default, both covered in quality and thinking level.
• inputs.referenceImages takes up to 16 images and turns the request into an edit, covered in editing images. For edits that need precise control, GPT-Image-2.6 Sunburst has a guide to detailed editing.
• inputs.markupImage carries marks drawn over the first reference to show where the edit goes, covered in markup editing.
• settings.background set to transparent returns the subject on an alpha channel, covered in transparent backgrounds.
• outputFormat also takes SVG and GIF on this model, covered in SVG output and animated GIFs.
• settings.moderation is auto or low, where low applies less restrictive filtering.

Writing a brief the model can plan from

Anything the prompt leaves out is a decision the model makes for you. Name a dashboard and stop, and the model picks the panels and every label on them. Write the brief a designer would need, and the model plans from the same information.

A product screenshot of a SaaS analytics dashboard.

A product screenshot of a SaaS analytics dashboard for a subscription business, shown flat and full screen with no device frame, as if captured from a real product. A narrow dark navy sidebar on the left holds the logo "Tallyfield" at the top and five menu items below it, "Overview", "Revenue", "Customers", "Churn" and "Settings", with "Overview" highlighted. Across the top of the main area sit four KPI cards in one row, each with a small gray label above a large figure: "MRR" and "$48,290", "Active customers" and "1,284", "Churn rate" and "2.1%", "ARPU" and "$37.61". Below the cards, a line chart titled "Revenue, last 12 months" fills the left two thirds, rising steadily from left to right, with the months "Jan" to "Dec" along the x-axis. To the right of the chart, a table titled "Top plans" lists three rows: "Pro" with "$21,400", "Team" with "$18,050" and "Starter" with "$8,840". Clean light interface, white cards on a pale gray background, one indigo accent color, crisp legible UI text, no other text.

In the first render the model invented everything: the product name InsightPro, the metrics and every figure. In the second, every label and figure is a string from the prompt, and each panel sits where the prompt put it.

The brief walks the screen one region at a time. Each clause says where its region sits, how many things it holds and what they read, so a missing card or a wrong label traces back to the one clause that described it.

Open with the deliverable. "A product screenshot of a SaaS analytics dashboard" says what the image is for, and "shown flat and full screen with no device frame" rules out a photo of a laptop on a desk. Describe an interface as a product that already exists, in terms of its layout and its labels, and leave concept-art language out of it.

Rendering readable text

Words to render go in positivePrompt with the rest of the scene. Wrap every string in quotation marks so the model renders those exact characters, and say where each string goes and what it is, such as the header or a row of the schedule.

A class schedule poster for a boutique fitness studio, made to post on social media. Flat graphic design on a deep charcoal background with one bright coral accent color. At the top, the studio name "KYNETIK STUDIO", spelled K-Y-N-E-T-I-K, in heavy white condensed sans-serif capitals, with "SATURDAY CLASSES" in coral directly below it. Below the header, five evenly spaced rows, each with the time on the left in coral and the class name on the right in white: "7:00 AM" and "Spin 45", "8:15 AM" and "Power Yoga", "9:30 AM" and "HIIT Circuit", "11:00 AM" and "Pilates Core", "12:15 PM" and "Mobility Flow". At the bottom, a smaller line reads "Book in the app · 24 spots per class". Crisp accurate typography, a clear size hierarchy between the header, the times and the class names, generous spacing, no photographs, no other text.

Write times and prices exactly as they should print. Quote "7:00 AM" when the poster should say 7:00 AM, since "seven" or "7am" leaves the format to the model. Pairing each time with its class in one clause, "7:00 AM" and "Spin 45", tells the model which two strings share a row.

Two more clauses pay off on a text-heavy design. Spell an invented name letter by letter, as the prompt does with KYNETIK, because a made-up word has no spelling for the model to recall. Then end with "no other text", so the empty space on the poster stays empty instead of filling with labels nobody wrote.

Choosing a size

width and height are optional. Leave both out and the model picks the frame, which suits a first draft before the image has a slot to fill. Set them, and the pair has to fit one envelope with four rules:

• Each side is between 16 and 3840 pixels.
• Each side is a multiple of 16.
• The aspect ratio is between 1:3 and 3:1.
• The area is between 655,360 and 8,294,400 pixels.

The top of the area band is exactly 3840 × 2160, so a 4K frame fits at full size and the largest square is 2880 × 2880. The common delivery shapes all have a size that passes the four rules:

The same listing prompt at 3:1 for a property portal banner, at 2:3 for a social poster, and with no size sent:

A listing photograph for a real estate agency website. A two-story white clapboard farmhouse with a black front door, a covered front porch with two wooden rocking chairs, and a freshly mowed green lawn with a stone path leading to the steps, under a clear late afternoon sky. A bold red ribbon banner runs diagonally across the top-left corner of the photo, laid over the image as a graphic, and reads "JUST LISTED" in white capitals. A white rounded label laid over the bottom-right corner reads "$685,000" in large dark type above a smaller line, "4 bd · 3 ba · 2,450 sq ft". Eye-level shot from the front walk, 24mm lens, the whole house in frame. Photoreal real estate photography, bright natural color, no other text.

A listing photograph for a real estate agency website. A two-story white clapboard farmhouse with a black front door, a covered front porch with two wooden rocking chairs, and a freshly mowed green lawn with a stone path leading to the steps, under a clear late afternoon sky. A bold red ribbon banner runs diagonally across the top-left corner of the photo, laid over the image as a graphic, and reads "JUST LISTED" in white capitals. A white rounded label laid over the bottom-right corner reads "$685,000" in large dark type above a smaller line, "4 bd · 3 ba · 2,450 sq ft". Eye-level shot from the front walk, 24mm lens, the whole house in frame. Photoreal real estate photography, bright natural color, no other text.

A listing photograph for a real estate agency website. A two-story white clapboard farmhouse with a black front door, a covered front porch with two wooden rocking chairs, and a freshly mowed green lawn with a stone path leading to the steps, under a clear late afternoon sky. A bold red ribbon banner runs diagonally across the top-left corner of the photo, laid over the image as a graphic, and reads "JUST LISTED" in white capitals. A white rounded label laid over the bottom-right corner reads "$685,000" in large dark type above a smaller line, "4 bd · 3 ba · 2,450 sq ft". Eye-level shot from the front walk, 24mm lens, the whole house in frame. Photoreal real estate photography, bright natural color, no other text.

Each frame is laid out again for its shape rather than cropped from a larger one. The ribbon and the price label keep their corners at every size, because the prompt anchors them to the corners of the frame rather than to the house.

Pick the size before you write the framing clause. Positions written against the corners and edges of the frame carry over to a new shape. "The house on the left third" was written for a wide frame and has to be rewritten for a tall one.

Tips

1. Open with the deliverable. "A product screenshot" or "a class schedule poster" in the first clause tells the model what the image is for before any detail arrives.

2. Brief the layout region by region. One clause per region, each with its position, its count and its exact text.

3. Write counts as numbers. "Four KPI cards" is four. "Some metrics" is however many the model settles on.

4. Quote every string as it should print. Times and prices included: "7:00 AM", not "7am".

5. Spell invented names letter by letter. A made-up brand has no spelling the model can recall.

6. End text-heavy prompts with "no other text". It keeps empty space empty.

7. Give in-scene text a surface and overlay text a corner. A tag or a sign for text inside the photo, a named corner or edge for a graphic laid over it.

8. Choose the size first. The model lays out the frame it is given, so the framing clause and the aspect ratio are one decision.