For years, putting readable text inside an AI-generated image was basically impossible. Ask any image generator to produce a business card, a poster, or a meme with legible words and you'd get back a mess of squiggly pseudo-letters that looked like someone tried to write while riding a bus. OpenAI's new Images 2.0 model, now rolling out inside ChatGPT, is changing that.
According to TechCrunch's hands-on coverage, Images 2.0 handles in-image text significantly better than what came before. That's not a small thing. Designers, social media managers, and marketers have been running a two-step workaround for years - generate the image, drop it into Canva or Photoshop, add the text manually. If Images 2.0 can collapse that into a single prompt, that's a real workflow change for a large slice of ChatGPT's user base.
What This Solves in Practice
Think about the kinds of images that need embedded text: event flyers, product mockups, social media graphics, ad concepts, infographics. These are the exact assets that small teams and solo creators produce constantly, usually by stitching together multiple tools. A model that renders "Summer Sale - 40% Off" accurately on a storefront graphic, without manual editing, removes a genuine friction point.
The text-in-image problem has always been partly about how image generators are trained - they learn to produce pixel patterns that look like writing without actually understanding that letters have rules. Fixing this likely required both architectural changes and more targeted training data. OpenAI hasn't published a technical breakdown, but the results speak to meaningful model-level improvements, not just a surface-level tweak.
Where DALL-E 3 Left Things
DALL-E 3, the model that previously powered ChatGPT's image generation, was already a step up from earlier versions at short text strings - a single word or two-word label could sometimes come through cleanly. But anything longer or more design-specific tended to degrade. Images 2.0 appears to push that ceiling considerably higher, though real-world testing across varied prompts will tell us how far the improvement actually extends.
For anyone who produces visual content regularly, this is worth testing now. The gap between "good enough to ship" and "needs manual cleanup" is where most of the time goes.