GPT Image 2.5: spot edits, sketches and less lost details

OpenAI released ChatGPT Images 2.5 on September 8. The main thing in the update is more manageable editing: according to the company, the model better retains the original details and previously made changes. For a designer, this is potentially more important than another spectacular result from the first query.

## What changed

OpenAI claims more natural lighting, detailed textures and better preservation of object recognition from references. The model performs local edits more accurately, follows a complex visual task more confidently and works better with a transparent background. According to OpenAI, the waiting time for generation has decreased to 50% compared to Images 2.0; this is not a guaranteed win on every request.

A separate emphasis is sequential editing. After a few refinements, the image should better retain the details already agreed upon. This is the developer’s statement, not the result of our comparative test.

Source: анонс OpenAI.

## Sketch and comments instead of a long prompt

In ChatGPT appeared Sketch: you can draw a sketch and use it as a reference for generation. He is called through @Sketch. Comments appeared directly on the images for address edits, templates of popular formats - for example, posters and merch - and the ability to share the picture along with the request.

For art direction, the form of communication itself is interesting: the position of the subject is easier to show than to describe with a paragraph. But the sketch should be perceived as a task, not a promise of accurate reproduction of the grid.

Source: объявление в OpenAI Developer Community.

## What is useful for the designer

Consider a possible working scenario: the customer agreed on the subject composition, but asked to change the background and make the light softer. The criterion for success here is not the beauty of the new version in itself, but the preservation of what has already been approved.

To check for the update, we would take one source and give three consecutive tasks:

1. Replace the background with light gray, without changing the subject and its position.
2. To soften the shadow, preserving the result of the first edit.
3. Add space to the right under the title without distorting the proportions of the product.

This is an editorial example that we did not run. After each step, it is worth comparing the silhouette, small inscriptions, color and position of the object with the source. If it is necessary to restore the packaging for the sake of a new shadow, there is no time saving yet.

For a series of promotional materials, it is useful to evaluate not the best frame, but the percentage of options that can be brought to delivery without reassembly. For presentations, check to see if there is room for the text. For product cards – how accurately the real features of the product are preserved.

## How to write requests: recommendations from the guide

OpenAI advises you to describe the result, composition, and limitations; each reference should be assigned a role. Text for the image - indicate in quotation marks together with the requirements for placement. When editing, separate changes and saved details, moving along one edit.

A short editorial example for the cover, not tested on the model:

Image 1 is only a reference palette and texture. Image 2 is the main subject. Collect the horizontal cover: the item on the right, the title "New Collection" on the left. Don't move the reference.

Next request:

Change the background color to dark blue. Save the subject, composition and title.

The request does not guarantee the immutability of the pixels. For absolutely accurate preservation of the site, the guide recommends pasting the approved edit into the original.

## Transparency and 4K: nuances for layouts

In the API, transparency is set by the background="transparent" parameter, the format is PNG or WebP; the alpha channel must be checked. Images larger than 2560×1440 pixels in total area are marked as experimental, including 4K. API settings are not replaced by words in the prompt.

Source: официальный гайд GPT Image 2.5.

## Flare and Sunburst: two models in the API

According to the OpenAI guide, Flare is optimized for speed at the GPT Image 2 level, Sunburst is optimized for higher quality. Compare them on the same queries, references and settings.

Source: выбор модели в гайде.

The rollout of Images 2.5 began in ChatGPT, ChatGPT Work, and Codex across all tariffs, web, mobile, and desktop applications. This does not mean the same limits or the absence of regional restrictions.

Source: доступность и версии API в анонсе OpenAI.

## What to check before surrender

We would still leave the final text, logo, size and color check to the designer. This material did not check the quality of the Cyrillic alphabet and the accuracy of compliance with the brandbook.

The main question for Images 2.5 is how many agreed decisions will survive the next edit. It is on such a test that it will be clear whether the update helps to complete the work faster, and not just get new pictures faster.