OpenAI on the 9th unveiled ChatGPT Image 2.5, the latest image generation model that more faithfully preserves the characteristics of the original and improves image quality and editing accuracy.
The latest model more accurately reflects the unique features of people and objects in photos and renders lighting and textures more naturally. A Sketch feature that lets users draw directly inside ChatGPT to create the images they want has also been added.
In particular, the ability to preserve the original's characteristics has been improved even when changing the background, style, or composition. Users can render a person in a reference photo with a new background or painting style, or keep a product's appearance while changing the shooting background and composition. The ability to edit only the requested parts while maintaining other detailed elements has improved even in images where the subject and background are complex.
Consistency in iterative editing has also improved. The company said it reduced instances where the previous edits changed or image quality degraded when refining product photos or advertising assets step by step. It also improved the ability to handle complex visual instructions and layout (placement), transparent backgrounds, and the content accuracy of images that contain real-world information.
A Sketch feature has also been added to let users fine-tune the creative process. Sketch is a new feature that allows users to draw directly inside ChatGPT and use that as References for the final image. When users type "@Sketch" in the ChatGPT chat window, the sketch tool opens and they can draw directly with a mouse. ChatGPT uses this drawing as a reference to generate refined images.
ChatGPT Image 2.5 will be rolled out to users of all plans for ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and the web. OpenAI said, "More than 3 billion images are currently being generated each week with ChatGPT Image and the GPT-Image model of our application programming interface (API)."
Features to assist creation have also been added. After selecting templates such as Poster or Merch, users can start production by entering the information to convey, design elements, and style. A feature that lets users comment directly on images to specify what to edit has also been added. When sharing images, users can include the prompts (instructions) used to create them. Recipients can apply their own photos and details to the same prompts to make additional edits or generate new images.
For developer APIs, OpenAI released GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. GPT-Image-2.5 Flare offers higher image quality than GPT-Image-2 while reducing generation latency by up to 50%. The company said it is suitable for a variety of uses, including social content, in-product image generation, visual search, rapid mockup production, and large-scale image generation.
Sunburst is a model for tasks that require fine adjustments and high precision during editing, even if generation takes longer. It can be used to produce campaign assets for actual ad execution or high-quality product images. Both models are provided via API.
OpenAI said it designed the system to prevent harmful outputs by checking prompts and images based on its existing safeguards. It will continue to apply C2PA metadata and invisible watermarks so that images generated with OpenAI tools can be identified.