ChatGPT Images 2.0 fixes garbled text
- OpenAI’s ChatGPT Images 2.0, released April 21, swaps in the new gpt-image-2 model and finally makes in-image text readable enough for real diagrams. - The clearest proof is practical: OpenAI says text-heavy visuals, charts, posters, and multilingual layouts are now first-class outputs, not cleanup jobs. - That matters because DALL·E 2 and 3 shut down May 12, pushing image workflows onto models built for usable assets.
AI image models have been able to make pretty pictures for a while. The annoying part was everything that needed to be read. Labels broke. Menus turned into fake words. Diagrams looked right until you zoomed in and realized the captions were nonsense. That is the gap OpenAI is trying to close with ChatGPT Images 2.0, which launched on April 21 and runs on a new model called gpt-image-2. (openai.com) ### What was actually broken? Text inside AI-generated images has been the long-running embarrassment of the category. A model could nail lighting, composition, and style, but then misspell the headline on a poster or turn chart labels into mush. That made whole classes of work half-useful at best — slides, UI mockups, infographics, product flows, comics, and anythin(openai.com)enAI is explicitly pitching Images 2.0 as stronger on “text-heavy visuals” and “structured generation,” which is the tell here. (community.openai.com) ### What changed in this version? The new model is not just tuned to draw prettier images. It is tuned to produce usable assets. OpenAI describes gpt-image-2 as better at text rendering, multilingual output, layout, editing, and instruction-following. In plain English, that means you can ask for a poster in Japanes(community.openai.com)ething you can actually ship. (community.openai.com) ### Why does readable text matter so much? Because it changes the job from inspiration to production. If the words are wrong, the image is a draft. Someone still has to rebuild it in Figma, PowerPoint, Photoshop, or Illustrator. If the words are mostly right, the image becomes a real workflow tool. That is the jump(community.openai.com) lean into this too, with support for multi-turn editing and image generation inside broader app flows. (developers.openai.com) ### Is this just about English text? No — and that is a bigger deal than it sounds. OpenAI is calling out multilingual support as a headline feature, not a footnote. That matters because older image models often degraded even faster outside English, especially with denser scripts and mixed-language layouts. If multilingual text rendering really holds(developers.openai.com)and region-specific marketing assets. (openai.com) ### Did the benchmarks move too? Yes, and by a lot. In OpenAI’s developer forum announcement, the company says gpt-image-2 hit the #1 spot across all Image Arena leaderboards just hours after launch, including a +242 point lead in text-to-image. Benchmarks are never the whole story, but that size of gap suggests this was not a tiny cleanup pass. It looks more like a (openai.com)eople can inspect quickly — readable posters, cleaner layouts, and more faithful prompt execution. (community.openai.com) ### Why mention DALL·E right now? Because the handoff is immediate. OpenAI’s deprecation notices say DALL·E 2 and DALL·E 3 are scheduled to shut down on May 12, 2026. So this is not just a nice upgrade sitting beside the old stack. It is the replacement path. If you build on OpenAI’s image tools, the company is sig(community.openai.com)E line. (community.openai.com) ### So what is the real takeaway? The big news is not that AI can spell a little better. It is that image generation is getting pushed into documents, interfaces, diagrams, and other formats where correctness matters as much as aesthetics. When the text stops falling apart, a whole new set of boring, valuable use cases opens up. That is usually when a model goes from demo bait to infrastructure. (community.openai.com)
