ChatGPT Images 2.5: Sketch first, then guide the edit

OpenAI introduced ChatGPT Images 2.5 on September 8, adding a more direct way to shape and revise generated images. Users can draw a rough guide, comment on a specific part of an image, and continue editing while the model aims to leave unrelated details unchanged. The update is rolling out across desktop, mobile, and web.

Start with a sketch, then point to the exact change

The new Sketch feature opens with @Sketch and lets users draw directly inside ChatGPT before using that drawing as a guide for the final image. Comments can also be placed directly on an image to show where a focused edit belongs. OpenAI has added templates for posters, merchandise, flyers, and product photos, while shared prompts let other people reuse an idea with their own details. In early hands-on use, Axios reported better preservation of people and pets alongside more detailed edits; that observation is not a universal benchmark.

Flare favors speed; Sunburst favors control

OpenAI says generation latency can be up to 50% lower than with Images 2.0, though that figure is not a guarantee for every request. For developers, GPT-Image-2.5 Flare is positioned as the faster default for most applications and high-volume work. GPT-Image-2.5 Sunburst targets premium workflows that need tighter editing control and can accept longer generation times.

More realism brings a sharper deepfake concern

OpenAI describes Images 2.5 as producing sharper details, more natural lighting, richer textures, and stronger preservation of subjects from reference photos. Its system card also warns that greater realism can increase deepfake risks involving real people, places, or events. The system continues to use C2PA metadata, invisible watermarking, prompt checks, image checks, and layered safeguards, but those measures do not make misuse impossible.