> For the complete documentation index, see [llms.txt](https://docs.pletor.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.pletor.ai/models/image-models/gpt-image-2.5.md).

# GPT Image 2.5

OpenAI's latest image model. Available in two modes and different quality tiers. Direct alternative to Nano Banana.

## Overview

OpenAI's latest image model, and the one to reach for when edits have to be exact.

GPT Image 2.5 keeps everything [GPT Image 2](/models/image-models/gpt-image-2.md) did well (text, layout, brand accuracy) and improves where it felt short with faster generations.

Available in two variants and different quality options. Same model, different speed-to-precision tradeoff.

| Variant           | Flare                                                                   | Sunburst                                                                    |
| ----------------- | ----------------------------------------------------------------------- | --------------------------------------------------------------------------- |
| Best for          | Default for most work: PDP visuals, social creative, high-volume drafts | Premium campaign creative, polished product imagery, multi-step edit chains |
| Speed             | Fast. Up to 50% faster than GPT Image 2                                 | Slower, trades time for precision                                           |
| Editing control   | Strong                                                                  | Tightest available                                                          |
| Quality reference | Above GPT Image 2 High                                                  | Above Flare on fine detail retention across edits                           |

Start with Flare. Move to Sunburst when a Flare edit drifts or when the asset is a final.

***

## How to use effectively

GPT Image 2.5 rewards the same structured prompts as GPT Image 2, plus one new habit: when editing, say what to keep.

**Key principles**

* **State the invariant.** In edit prompts, name what must not move: "Keep the product, its position, the lighting and the background unchanged." The model respects it.
* **One change per edit.** Chain small edits instead of stacking three changes in one prompt. 2.5 holds quality across turns, so use that.
* **Pick the variant per step.** Flare for drafts, iterations and the bulk of the chain. Sunburst for the final pass or for the one edit that keeps drifting.
* **Be explicit with copy.** Exact text in quotes, with placement (header, label, CTA).
* **Specify layout when structure matters.** Grids, columns, hierarchy, spacing.
* **Ask for transparency explicitly.** "Transparent background, no shadow, no ground plane" when you need a clean cut-out.
* **Use references for products and people.** Fidelity to the reference is where 2.5 pulls ahead; feed it a clean one.

**Effective prompting patterns**

<details>

<summary>For single-element edits:</summary>

`Change only [element] to [new value]. Keep the [subject], its position, scale, lighting, shadows and the background exactly as in the reference. Do not alter any text.`

</details>

<details>

<summary>For copy swaps and localization:</summary>

`Replace the headline "[old copy]" with "[new copy]" and the CTA "[old CTA]" with "[new CTA]". Preserve font, weight, color, alignment and position. Everything else in the image stays identical.`

</details>

<details>

<summary>For product placement across scenes:</summary>

`Place the exact product from the reference image into [scene]. Preserve its packaging, label typography, colors, proportions and the "[wordmark]" as shown. Lighting: [description]. Camera: [lens, angle]. Aesthetic: [reference].`

</details>

<details>

<summary>For variant sets:</summary>

`Create [N] variations of the reference image. Vary only [scene / colorway / backdrop]. Keep the product, its position, the composition and all text elements identical across every output.`

</details>

<details>

<summary>For transparent-background assets:</summary>

`[Subject or object], isolated on a fully transparent background. No shadow, no ground plane, no reflections. Clean edges suitable for compositing.`

</details>

<details>

<summary>For static ads with text:</summary>

`A [format] ad for [brand]. Headline: "[exact headline copy]". CTA button: "[exact CTA]". Optional fine print: "[exact text]". Visual: [scene/subject]. Brand colors: [hex or name]. Typography should feel [adjective]. Layout: [hierarchy description].`

</details>

<details>

<summary>For editorial photos:</summary>

`A photorealistic [shot type] of [subject], [action or expression], set in [environment]. Lighting: [description]. Camera: [lens, angle, framing]. Atmosphere: [reference style, color treatment, era]. Mood: [adjectives]. Composition emphasizes [key textures, details].`

</details>

**Example prompts**

* **Background swap**: "Change only the background to a sunlit Mediterranean terrace with terracotta tiles. Keep the sneaker, its angle, scale, shadow and the 'Run light' headline exactly as in the reference."
* **Localization**: "Replace the headline 'Soft hands, every day.' with 'Des mains douces, chaque jour.' and the CTA 'Shop now' with 'Acheter'. Same font, weight, pink color, position. Nothing else changes."
* **Colorway variants**: "Create 4 variations of the reference tote bag shot. Vary only the leather color: cognac, forest green, black, cream. Keep the bag's position, the concrete background, the lighting and the 'Crafted in Paris' embossing identical."
* **Transparent cut-out**: "The hand cream tube from the reference, isolated on a fully transparent background. No shadow, no surface, clean edges. Preserve the 'Hold my hand!' callout and pink cap exactly."
* **Product in scene**: "Place the exact serum bottle from the reference on a marble bathroom shelf beside eucalyptus. Preserve label typography, amber glass color and the dropper cap. Soft morning window light, 50mm, shallow depth of field."
* **Edit chain, final pass (Sunburst)**: "Add a small 'New' badge in the top-right corner, white text on a black circle. Keep every other pixel of the image unchanged."
