OpenAI rolls out ‘ChatGPT Images 2.0’
The neural network learned to perfectly fit text into a frame and independently search for data on the network.
OpenAI has released a major update to its visual AI. ChatGPT Images 2.0 shifts its focus to creating complex layouts, infographics, and drawings.
This version was previously secretly tested under the code name “duct tape”.
ChatGPT Images 2.0 is designed to improve how AI interprets and executes visual prompts, with a focus on accuracy, structure, and visual consistency. It can more accurately position objects, follow detailed instructions, and maintain a consistent layout in complex scenes.
A significant addition is the ability to reason about queries before generating images. In supported modes, it can also use web information, generate multiple image results from a single query, and refine the results to make them more consistent.
This shifts the system from basic image generation to a more structured visual reasoning tool.
What’s new in version 2.0:
- Perfect text. The neural network no longer distorts letters. It accurately renders dense text blocks, website screenshots, and small fonts.
- Multiple languages. Text is correctly generated not only in Latin, but also in Japanese, Korean, Chinese, Hindi, and Bengali.
- “Thinking” mode. Before creating an image, the AI can search for relevant information online (such as weather or statistics) to create an accurate infographic based on real and current data.
- Freedom of formats. Images can be shot in up to 2K resolution and in any aspect ratio – from narrow screens (1:3) to wide panoramas (3:1).
The update was released to everyone on April 21st. The second version is available to users in the ChatGPT app, via the developer API (model gpt-image-2), and in the Codex app.
Source:
