OpenAI has released an updated version of its image generator, ChatGPT Images 2.5. The new model puts less emphasis on generating polished images and more on precisely following edit instructions for existing photos. It also does a better job preserving key details in a shot, including a person’s face, when the style or scene changes.
The core improvement is instruction accuracy. When editing a photo, ChatGPT Images 2.5 now changes only the specific element a user asks for—the background, an item of clothing, or a caption—and leaves everything else untouched. That matters most in multi-step edits, where earlier versions sometimes altered a person’s face midway through or undid changes from a previous step.
Speed has improved too. OpenAI says the new model completes requests up to 50% faster than its predecessor. It also handles complex prompts better, including detailed compositions, text within images, transparent backgrounds, and specific visual styles.
After generating an image, users can now share the prompt alongside the picture itself, letting others try to recreate the same result with their own photo. ChatGPT Images 2.5 also introduces a new Sketch mode: users can draw a rough layout by hand directly in the ChatGPT interface when a composition is difficult to describe in words. OpenAI added ready-made templates for common tasks as well, such as posters and marketing materials.
Follow Mediasat on Google News: the latest in TV, telecom and tech
ChatGPT Images 2.5 is available across ChatGPT, ChatGPT Work, and Codex—on desktop and mobile apps, on the web, and via the API. OpenAI is offering it in two versions: the simpler, faster GPT-Image-2.5 Flare for most everyday tasks, and the slower GPT-Image-2.5 Sunburst for maximum control and editing precision.
Other major companies have recently updated their own image generators. In July, Meta launched Muse Image, a model that checks and corrects its own outputs. That same month, Google introduced Nano Banana 2 Lite, which generates an image in a matter of seconds.
Every image created or edited with ChatGPT Images 2.5 gets an invisible watermark and C2PA metadata. The marking signals that the image has been processed with AI.
