动察 Beating AI Flash: OpenAI has officially released ChatGPT Images 2.5, with major upgrades in generation speed, character fidelity, and image editing. Compared to Images 2.0, the new version can reduce image generation latency by up to 50%. When using photos as references, it also better preserves the likeness and details of subjects such as people and pets.

The most practical change this time is that editing has finally become more "obedient." Users can now directly circle the area they want to modify on the image and write a single instruction on how to change it—such as swapping out just one petal or altering the color of a single object—without repeatedly describing the exact location in text. During multi-round edits, the model is also less likely to alter parts that weren't requested to be changed.

ChatGPT now also supports direct sketching to control composition, and has added common templates for posters and merchandise. Images 2.5 has begun rolling out to ChatGPT, ChatGPT Work, and Codex users.

The API simultaneously launches GPT-Image-2.5 Flare and Sunburst. In the latest Arena evaluation, Sunburst ranks first and Flare second, with both models sweeping the top two spots across the text-to-image, single-image editing, and multi-image editing leaderboards.