OpenAI launches ChatGPT Images 2.5, its new visual AI

OpenAI has launched ChatGPT Images 2.5, a new model for creating and editing images with more detail, better instruction following and faster generation. The company says its image tools now produce more than 3 billion images every week across ChatGPT and its API models.

The update is rolling out to ChatGPT, ChatGPT Work and Codex on desktop, mobile and the web. It is available to users on all plans.

More faithful images and more precise edits

Images 2.5 is especially improved when you work from a reference photo. It can better preserve the appearance of a person, product or place while changing the setting, visual style or composition.

In practice, you can use a product photo to create multiple advertising images without losing the object’s key features. It is also more useful for transforming a personal photo without changing the face or distinctive details too much.

The model also tries to modify only what you ask for. You can change the background, replace a product or correct text without completely redoing the rest of the image. This matters especially for brand designs, where the composition, colors and visual identity need to remain consistent.

Another improvement appears in longer conversations. If you make several edits in a row, Images 2.5 does a better job of preserving earlier changes and reduces image degradation with each new instruction.

Generation is faster too: OpenAI says latency has been reduced by up to 50% compared with Images 2.0. That lets you try more versions and make corrections without waiting as long between attempts.

New tools for creating in ChatGPT

The update is not limited to the model. ChatGPT adds several features to give you more control over the creative process:

• Sketch lets you draw directly in ChatGPT and use the sketch as a guide. You can draw a room layout, the silhouette of a garment or a quick idea and ask it to turn that into a complete image. To activate it, type @Sketch.
• Templates offers templates for common formats, such as posters and promotional products. Instead of starting with a blank canvas, you choose a base and add the text, style and elements you need.
• Image comments let you point to specific areas and request more focused changes.
• Share prompts lets you publish the instruction used to create an image as well. Other people can try the same idea with their own photos and details.

The model also handles complex visual instructions better. It can work with more structured designs, transparent backgrounds, real-world information and specific artistic styles. That can be useful for creating presentations, interface concepts, campaigns and image series that need to follow the same visual direction.

Two new models for developers

OpenAI has also added two models to its API, the service that lets developers integrate image generation into other applications:

• GPT-Image-2.5 Flare is the default option for most projects. It offers the quality, editing and speed improvements of Images 2.5, with latency 50% lower than GPT-Image-2, according to OpenAI. It is designed for social media content, product experiences, visual search, prototypes and large-scale generation.
• GPT-Image-2.5 Sunburst is aimed at work that needs more precise control and longer generation times, such as production-ready advertising campaigns or polished product images.

For a company, the difference is not just the ability to generate an attractive image. It is being able to change one element without rebuilding the entire asset. For example, a marketing team can update a product package or the text in an ad while keeping the rest of the design intact.

What changes for you

If you use ChatGPT to create images, you will have more control without needing to learn design software. A rough sketch can serve as a map, a template can speed up the first draft and comments can replace long, difficult-to-explain instructions.

The biggest improvement appears when you need several rounds of changes. Creating an image once is simple; keeping the same person, product or composition after ten edits is much harder. Images 2.5 is designed to handle that problem more effectively.

OpenAI is keeping its safety controls for reviewing instructions and images, along with C2PA metadata and invisible watermarks that help identify images created with its tools. The next thing to watch will be how the model performs in real projects, where consistency across many images often matters more than the quality of a single image.