GPT Image 2.5

How GPT Image 2.5 reads a prompt

GPT Image 2.5 comes in two variants. OpenAI's guide calls Flare the small model, built for speed, and the launch post puts its latency 50% below GPT Image 2. Sunburst is the base model, built for quality, and it takes longer to generate. Both generate from text, edit images and handle transparent backgrounds, and OpenAI says both improve on GPT Image 2 at precise editing and at keeping the subject of a reference photo recognizable.

OpenAI's prompting guide says the format of a prompt matters less than its clarity. A short sentence, a paragraph, labeled sections and a JSON-like list can all express the same request. Pick the one that is easiest to read and update. The guide does not rely on special syntax or style keywords; it asks you to describe the result.

A structure that works

For anything beyond a simple image, OpenAI suggests organizing the prompt as scene, subject, details and constraints, with labeled sections:

Use: A product photo for a launch email, portrait format.
Scene: A sunlit kitchen counter, white tile, a plant out of focus.
Subject: A matte black coffee grinder, centered, three-quarter view.
Details: Soft window light from the left, real metal texture, 50mm look.
Text: The label reads "Burr One", exactly once. No other text.
Constraints: No logos, no watermark, no props in front of the grinder.

Describe what is visible: materials, lighting, colors and the medium. Ask for "photorealistic" or "a real photograph" when that is the goal. OpenAI treats camera terms such as "35mm" or "50mm lens" as cues for the look, not as an exact simulation. For people, describe framing, gaze and what the hands are doing ("full body visible, feet included", "looking down at the open book").

Exact text and layouts

Put required wording in quotes, say where it goes and what the type looks like, and say how many times it should appear. Ask for no extra text. Spell unusual names letter by letter if the first result gets them wrong. For slides, charts and diagrams, write the prompt like a spec: name the deliverable, give the real numbers and labels, and describe the hierarchy. OpenAI's guide suggests the high quality setting for small text, legends and footnotes, and a check of every label and fact in the output.

Edits and references

Separate the change from the constraints. Say "change only X", then list what must stay: identity, geometry, layout, lighting, labels. For a local edit, also name the nearby details that must not move, such as the camera angle or the objects around the edited area.

When you attach several images, give each one a number and a role: subject, style, clothing or background. Then explain how they combine and what moves where. Endless takes up to 16 reference images for this model.

For a series of edits, feed the latest result back in and ask for one change at a time. OpenAI says Images 2.5 keeps earlier changes more reliably across turns, but also warns that repeated edits can still alter details you meant to keep, so restate the critical constraints when a result drifts. If a region must stay pixel-identical, composite the approved edit into the original instead of relying on the prompt alone.

Settings that matter

Endless exposes these settings for GPT Image 2.5:

• Mode: Flare (the default) or Sunburst. OpenAI's advice is to start with Flare when speed matters and Sunburst when quality matters most, then test whether Flare also passes on the same prompt before switching back.
• Quality: low, medium or high (the default). Use low for quick drafts. Raise it for small text and dense layouts. A higher setting does not guarantee a better result on every prompt.
• Resolution: 720p, 1080p (the default) or 4k, combined with the aspect ratio to set the pixel size. OpenAI marks outputs above 2560 × 1440 as experimental.
• Aspect ratio: auto or one of ten fixed ratios from 21:9 to 9:16. Auto keeps the input image's shape when you edit.
• Background: auto, transparent or opaque. For a cutout, choose transparent and also ask for an isolated subject in the prompt. A drawn checkerboard is not transparency.
• Reference images: up to 16.

When you compare Flare and Sunburst, change only the mode and keep the prompt, references, quality and size the same.

Common failure modes and fixes

• Text is misspelled or placed wrong. OpenAI lists text placement as a remaining limit. Quote the text, state its position, raise the quality, and spell rare words out.
• An edit changes too much. List what must stay the same and add "do not change anything else".
• The wrong reference controls the result. Number the images and give each one job.
• A character drifts across a series. Repeat the character's defining details in every prompt and attach the established image as the reference.
• A cutout has a halo or a fake background. Set the background to transparent, ask for clean edges with no backdrop, shadow or checkerboard, and check the edges around hair and glass.
• A diagram has wrong facts. Give the facts and labels in the prompt and verify them in the output.