ChatGPT Images 2.0 is here! Text generation accuracy is greatly improved, making it easy to create marketing posters

OpenAI officially released ChatGPT Images 2.0 on Tuesday, significantly improving the accuracy of text generation, as well as the design aesthetics for posters and portraits. The model also introduced “thinking mode” for the first time, enabling image generation to include web search and multi-image batch output capabilities, bringing it fully in line with commercial application scenarios.

(Canva announces deep integration with Claude, enabling the transformation of AI drafts into finished design deliverables)

From making things up to perfect menus: AI has finally learned to spell

Looking back two years ago, the weaknesses of AI image generation models in text generation were almost universally known. As long as the prompt included text requirements, the output would often be filled with absurd spelling errors or even hallucinations. This was even worse in non-English Chinese, Japanese, and Korean languages.

Official announcement Korean poster mockup

Now, ChatGPT Images 2.0 can generate a promotional poster that can be used directly by vendors, with clear and accurate text. In recent years, researchers have actively explored new architectures such as (Autoregressive Models) that can self-regress, and its operating logic, understanding of text, and generation and verification capabilities have improved significantly.

Thinking mode goes live: web search and composition consistency are all covered

The most core upgrade in ChatGPT Images 2.0 is “thinking mode (Thinking Capabilities).” It is currently available to paid users of ChatGPT Plus, Pro, the business version, and the enterprise version. Once enabled, the model can instantly perform web searches to assist image generation. It can also create corresponding visual explanation graphics based on files the user uploads, and conduct self-review and optimization of the image content before official output.

In batch generation, under thinking mode, a single prompt can output up to eight images at once, and the images can maintain consistent character appearances, object styles, and overall visual style. This makes it suitable for comic panels, social media series image-and-text posts, and even interior design space planning drawings for various rooms.

Official announcement comic panel mockup

In terms of resolution, the new model supports up to 2K output, and also adds multiple aspect ratio options from 3:1 to 1:3, further meeting a variety of business needs.

Asian languages are greatly optimized—Chinese, Japanese, and Korean users are in luck!

Besides English, OpenAI specifically noted major improvements to Images 2.0 for Asian text, including clear enhancements in Japanese, Korean, and Chinese.

Test articles that circulated in Chinese tech communities a few days ago also verified the news. Multiple Zhihu creators conducted hands-on testing comparisons between GPT-Image-2 and the competing Google Nano Banana Pro, covering a range of scenarios such as Chinese poster design, e-commerce cover images, social media interface layouts, and data visual charts.

Zhihu article tests GPT-Image 2.0

The test results show that GPT-Image-2 clearly outperforms in the aesthetics of Chinese typefaces, layout hierarchy, and overall design feel. The generated poster styles are closer to real commercial materials, rather than template-like outputs with an obvious “AI look.” The article also points out that GPT-Image-2 shows higher detail accuracy in replicating the interface (such as game screenshots or messaging app screenshots) and in recreating real portrait scenes.

ChatGPT Images 2.0 fully opens up, and the API also launches

Currently, ChatGPT Images 2.0 has been providing basic functionality to all ChatGPT and Codex users free of charge starting this Tuesday, while paid users can unlock more advanced output effects. At the same time, OpenAI is also opening the GPT-Image-2 API. Pricing is calculated based on output quality and resolution tiers, offering integration flexibility for enterprises and developers.

It’s worth noting that the new model’s knowledge cutoff date is December 2025. For image generation prompts involving the latest current events, accuracy may be subject to certain limitations. In addition, the generation speed for complex compositions can’t be as immediate as typical text Q&A, but it still only takes a few minutes.

This article: ChatGPT Images 2.0 makes its debut! Text generation accuracy greatly improves, making it easy to produce marketing posters First appeared on Chain News ABMedia.