ChatGPT Images 2.5 turns a doodle into an image instruction
OpenAI's ChatGPT Images 2.5 adds a Sketch window whose doodle becomes part of the generation instruction, plus comments placed on specific image areas. OpenAI claims generation can be up to 50 percent faster, while Axios's early test focuses on likeness preservation and more precise local editing.
Artificial Intelligence··Morning
A doodle becomes a second instruction channel
The most visible addition in ChatGPT Images 2.5 is Sketch. The Verge and Axios report that a user can open a drawing window in the conversation and make a rough shape; that doodle then becomes part of the written instruction for generating or editing an image. Users can also place a comment directly on the area they want changed. The position, approximate shape or target area of an edit therefore no longer has to be described only in words. The interaction is less about drawing skill than showing where an intention belongs.[1], [2]
Early testing focuses on preserving likeness
Axios tried the engine on tasks including producing a logo, converting an existing cat photo into another visual genre and designing a tattoo. It reports that the new engine preserved the appearance of people and pets better than its predecessor and allowed detailed changes to be directed at a narrower area. These are early observations by a publication, not a controlled benchmark. They nevertheless make the product's goal concrete: the release is designed not only to create a first image, but to change an existing one over several turns without losing its subject.[2]
The speed claim comes with a wider set of workflows
OpenAI says Images 2.5 produces more natural lighting and richer textures, follows successive editing instructions more reliably and cuts generation latency by up to 50 percent compared with Images 2.0. The Verge identifies these as company claims and reports that the release is available to ChatGPT, ChatGPT Work and Codex users across desktop, mobile and web. That maximum improvement does not promise the same result for every request; the gain a user sees can vary with image complexity and the number of editing turns required.[1]
