It carries real faces from reference photos into new scenes, and edits a single detail without disturbing the rest of the picture.
ChatGPT Images 2.5, a newly built image generation model, went out on September 8 to everyone using ChatGPT, ChatGPT Work or Codex, on desktop, mobile and the web. It replaces Images 2.0.
The company says it is better at working from reference photos — moving familiar subjects into new settings, styles and compositions, with lighting and textures more natural and distinctive features more likely to carry through. Images that take in real-world information come out with more accurate content.
People already make more than 3 billion images a week across ChatGPT Images and the GPT-Image models in the API.
OpenAI says it continues to attach C2PA metadata and invisible watermarking to help identify images made with its tools, and that prompts and outputs are checked to head off harmful ones.
Editing is the other half. Asked to change one element, OpenAI says, the model leaves the rest of the picture as it was even where subjects and backgrounds are complex — a product, a background or a line of copy swapped while the subject and the composition around it stay put.
Over a longer session, earlier changes are more likely to hold, and image quality less likely to slip as further edits pile up.
Three additions arrive inside ChatGPT: Sketch, for drawing in the app as a reference; comments placed on a picture to steer a focused edit; and the option to attach the prompt behind an image when sharing it.
Generation is faster too: OpenAI puts latency down by up to 50% against Images 2.0.
Developers get two models through the API. Flare is the default — higher-quality images than GPT-Image-2 at half the latency; Manus, a launch partner quoted in the announcement, puts the gain at two to four times the speed. Sunburst trades the other way: longer per image, more exactness on fine detail.