OpenAI is giving ChatGPT’s image generation capabilities another major upgrade with the launch of ChatGPT Images 2.5, a new image model designed to produce sharper visuals while giving users considerably more control over how generated images are edited and refined.

Announced on September 8, OpenAI says Images 2.5 improves image fidelity, editing precision, consistency across multiple editing turns, and generation speed. The update also introduces new creative tools inside ChatGPT, including Sketch, which lets users draw directly in ChatGPT and use their rough drawing as a visual reference for image generation.

OpenAI says people now create more than 3 billion images every week across ChatGPT Images and its GPT-Image models in the API, making image generation a substantial part of how people use its products. Images 2.5 is intended to make that process less dependent on getting a perfect prompt on the first attempt.

Faster generation with better image fidelity

One of the biggest improvements in Images 2.5 is its ability to preserve important details from reference images while transforming them into new settings, compositions, or visual styles.

OpenAI says the model produces more natural lighting and richer textures, while distinctive characteristics of subjects in reference photos are more likely to remain intact. That could be particularly useful for workflows involving people, products, branded assets, and other images where maintaining the identity of the original subject matters.

The company is also claiming a significant speed improvement. Image generation latency has been reduced by up to 50% compared with Images 2.0, allowing users to generate and iterate on images more quickly.

The improvement is not limited to ChatGPT. OpenAI says the same reference-image fidelity improvements are available to developers building image-generation workflows through its API.

More precise image editing

Images 2.5 is also designed to be better at changing one part of an image without unintentionally altering everything around it.

OpenAI says the model can more reliably follow instructions that target a specific element while preserving the rest of the composition. For example, a user can change a product, background, or piece of text while keeping the subject, overall layout, and surrounding visual treatment intact.

That is an important distinction from simply generating another version of an image. Instead of repeatedly asking an image model to recreate an entire scene, users can make increasingly targeted changes to an existing result.

For creative professionals, that could make AI-generated imagery more practical for iterative design work, particularly when a project requires multiple small revisions rather than a completely new image each time.

Multi-turn editing is more consistent

OpenAI is also targeting one of the more frustrating parts of conversational image generation: keeping previous edits intact.

With Images 2.5, OpenAI says changes made earlier in a conversation are more likely to remain consistent as additional edits are applied. Each new instruction is intended to build on the previous version without gradually degrading the image.

That means users can make a sequence of changes instead of trying to describe every requirement in one enormous prompt.

For example, a user could first generate an image, change the clothing, adjust the background, modify the lighting, and then make another small compositional change. The model is designed to retain those earlier decisions as the conversation continues.

OpenAI says this consistency is also important for production workflows where developers need to make targeted changes without regenerating an entire asset from scratch.

ChatGPT can now turn sketches into images

Perhaps the most immediately noticeable addition is Sketch.

The new feature lets users draw directly in ChatGPT and use that drawing as a visual guide for the final image. OpenAI says users can sketch something as simple as a room layout, the rough shape of an outfit, or even a basic doodle, then describe the desired style and other details. ChatGPT can use that rough drawing to generate a more complete image.

The idea is straightforward: sometimes describing a visual concept with words is harder than drawing it.

A rough sketch can communicate positioning, proportions, composition, or the relationship between objects without requiring the user to write a detailed prompt explaining every element.

OpenAI says users can access the feature by typing “@Sketch” in ChatGPT.

That makes Sketch particularly interesting for users who have a specific visual idea but don’t necessarily have the prompt-writing skills needed to communicate it to an image model.

Templates make image creation more structured

OpenAI is also introducing templates for popular creative formats.

Rather than starting from a blank prompt, users can select formats such as posters or merchandise and then provide the information, design elements, and visual style they want incorporated.

The templates are essentially another way of structuring the creative process around the model’s capabilities. Instead of asking users to figure out how to describe a particular format from scratch, ChatGPT can provide a starting point that users can customize.

Images can be shared with their prompts

OpenAI is also making it easier to share the creative process behind an image.

When sharing an image, users can choose to include the prompt that generated it. Someone else can then take that prompt and use their own photos and details to create a personalized version.

This turns individual image creations into reusable starting points rather than static outputs.

Two new image models arrive in the API

For developers, OpenAI is introducing two new API models: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.

GPT-Image-2.5 Flare brings improvements in image quality, editing, and speed to the API and is positioned as the default option for most applications. OpenAI says it produces higher-quality images than GPT-Image-2 while offering 50% lower latency. It is aimed at applications including creator tools, social content, product experiences, visual search, image prototyping, and high-volume generation.

GPT-Image-2.5 Sunburst is aimed at more demanding visual workflows where tighter control over edits is more important than generation speed. OpenAI positions it for use cases such as production-ready campaign creative and polished product imagery.

OpenAI says early customers, including Higgsfield AI, Adobe, Manus, and Runway, are already using the new models.

Adobe, for example, is bringing the GPT-Image-2.5 models into Adobe Firefly, while other companies are using them for areas such as branded assets, presentations, websites, advertising, and creator workflows.

Safety remains part of the image pipeline

OpenAI says Images 2.5 builds on its existing image-generation safeguards, including checks on prompts and generated images designed to prevent harmful outputs.

The company also continues to use C2PA metadata and invisible watermarking to help identify images generated with its tools. OpenAI has published additional information about its safety evaluations and approach in a separate system card.

ChatGPT Images 2.5 is available now

OpenAI says ChatGPT Images 2.5 is rolling out to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web, covering all tiers.

The GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst models are also available through the API, with OpenAI providing separate pricing information for developers.

The broader shift with Images 2.5 is less about simply making AI-generated pictures look better and more about making image generation behave like an editable creative workflow. Faster generation reduces the waiting time between attempts, while precision editing and improved multi-turn consistency make it easier to gradually shape an image without constantly starting over.

And with Sketch, OpenAI is giving users another way to communicate an idea when words aren’t quite enough. For anyone who has ever thought, “I know what I want this to look like, I just can’t explain it,” that may end up being one of the most useful additions in the update.