OpenAI Launches ChatGPT Images 2.5 with Precise Editing and Sketch Reference

The biggest upgrade over Images 2.0 is the model's understanding of edits. Previously, AI drawing models were better at generating an image from scratch, but when asked to adjust a detail they often mistakenly changed other elements—for example, altering a character's outfit could change facial features, and adjusting the background could shift the overall style. OpenAI said Images 2.5 has been optimized for precise editing, allowing it to execute specified modifications while preserving the original subject, composition and visual style. Users can adjust a single element such as a product, background or text without regenerating the whole image.

In hands-on tests, Images 2.5 preserved the facial features of a reference person when generating a high-definition creative work, while changing photography style, clothing and environment. It can now display generation progress numerically, and some netizens even played the Snake game while images were generating. Tests involving a selfie of Elon Musk with The Godfather on set showed that the model maintained facial structure, expression and pose. In a product design test, the model generated a Coke can made entirely of soft plush material.

A complex visual logic test prompted for an orange cat sitting in an office chair holding an iPad, with the same cat holding the same iPad shown on the screen repeatedly to create recursion. The output correctly understood the nested relationship and produced a multi-level recursive effect. The model also showed improved understanding of complex artistic styles, including images closer to game concept art in a Black Myth: Wukong aesthetic and a serial-picture-book illustration with black line drawing, light color washes and aged paper texture, resembling classic Chinese comic style. Its handling of Chinese-language content was also strong, as with the previous generation.

The improvements can be summarized in three areas: enhanced comprehension of complex prompts with multiple styles and constraints, stronger preservation of reference images during repeated edits, and an editing flow that supports continuous adjustments on the same image rather than regenerating from scratch. However, noise issues remain a weak point, and some official showcase images still contain minor flaws, according to netizens.

New features around the creation process include Sketch, which lets users draw rough layouts, clothing outlines or creative sketches directly in ChatGPT and use them as visual references for the AI to generate a complete image. This addresses the difficulty of describing spatial relationships in words. Sketch also works for animated content such as GIFs. Template presets help users start from common design scenarios like posters and product images, then add text, style and design requirements. Users can also leave comments directly on an image to specify where to modify, and share the prompt used to generate an image so others can create their own versions based on the same idea.

For developers, OpenAI launched two API image models. GPT Image 2.5 Flare serves as the default model for most applications, offering higher quality, editing capability and speed; compared with GPT Image 2, it reduces latency by 50%, making it suitable for social content, e-commerce visuals, visual search and large-scale image generation. GPT Image 2.5 Sunburst targets more demanding creative workflows that require finer control, for use cases such as advertising and premium product design. ChatGPT Images 2.5 is now available to ChatGPT, ChatGPT Work and Codex users across desktop, mobile and web, while API users can call both new models.