ChatGPT’s new Images 2.0 model is surprisingly good at generating text

It was easy to distinguish between human-generated images and AI-generated ones. Just two years ago, you couldn’t use photo templates to create a menu for a Mexican restaurant without inventing new delicacies like enchuita, chorriros, puerto, and margarta.

Now, when I ask the new ChatGPT Images 2.0 template for a Mexican menu, it creates something that can be used immediately in the restaurant without customers noticing that something is wrong. (However, the $13.50 ceviche might make me question the quality of the fish.)

For comparison, this is the result I got from DALL-E 3 a couple of years ago (at that time, ChatGPT wasn’t generating images):

AI-based image generators have historically struggled with their spelling because they generally use diffusion models, which work by reconstructing images from noise.

“Diffusion models […] “We reconstruct certain inputs,” Asmelash Teka Hadgu, founder and CEO of Lesan AI, told TechCrunch in 2024. “We can assume that the writings on the image are a very, very small portion, so the image generator learns patterns that cover more of these pixels.”

Since then, researchers have discovered other image-generating mechanisms, such as autoregressive models, which make predictions about what an image should look like and work more like an LLM.

Unfortunately, OpenAI declined to answer a question at a press conference this week about what type of model is powering ChatGPT Images 2.0.

However, the company explained that the new model has “thinking capabilities”, which give it the ability to search the web, create multiple images from a single prompt, and double-check its creations – this allows Images 2.0 to create marketing assets of varying sizes, as well as multi-panel comic strips.

OpenAI also says that Images has a stronger understanding of displaying non-Latin text in languages ​​such as Japanese, Korean, Hindi, and Bengali. The model expires in December 2025, which may affect how accurately it generates certain claims that include recent news.

“Images 2.0 brings an unprecedented level of precision and precision to image creation. Not only can it visualize more complex images, but it actually effectively brings that vision to life, able to follow instructions, preserve desired details, and render the subtle elements that often break image paradigms: small text, icons, UI elements, dense textures, and subtle stylistic constraints, all at up to 2K resolution,” OpenAI said in a press release.

These capabilities mean that creating images isn’t as fast as writing a question on ChatGPT, but creating something as complex as a multi-panel storyboard still only takes a few minutes.

All ChatGPT and Codex users will have access to Images 2.0 starting Tuesday; Paid users will be able to create more advanced outputs. The company will also make the gpt-image-2 API available, with a price dependent on the quality and accuracy of the output.