Image-generation models have a long history of bungling text. But while garbled letters used to be a clear AI tell, ChatGPT’s new image-generation tool is the best I’ve ever seen at rendering text.
I asked ChatGPT’s Images 2.0 model (available now to all ChatGPT users, including those on the free tier) to take some text from a recent story of mine and render it in pencil on a yellow legal pad and, well, it looks pretty much perfect to me:
Ben Patterson/Foundry
I also prompted it to create an infographic about AI tokens, instructing it first to search the web for accurate information and to use a serif font in a landscape 3:2 aspect ratio. Here’s what I got:
Ben Patterson/Foundry
Then I tasked Images 2.0 with creating another infographic, this time detailing the various Raspberry Pi models complete with specifications and other details:
Ben Patterson/Foundry
Finally, I asked the model to take a snapshot of me poolside and create a summer lookbook of outfits, starring me:
Ben Patterson/Foundry
OpenAI says Images 2.0 is its first image-generation model with “thinking” capabilities, meaning it can stop and ponder an image prompt before diving right in.
When it comes to text, Images 2.0 supports a variety of languages, including Japanese, Korean, Chinese, Hindi, Bengali, and others that employ non-Latin text.
It can also search the web for real-time information before rendering images, as well as create multiple images in one shot, good for rendering catalog images, comicbook-style panels, and storyboards.
OpenAI promises that Images 2.0 will deliver an “unprecedented level of specificity and fidelity,” meaning (hopefully) that it will do a better job at prompt adherence–that is, creating images that follow your prompts to the letter.
With this level of accuracy, Images 2.0 could offer an answer to the question I’ve long asked about image-generating models: What are they good for, aside from creating goofy memes or creepy deepfakes? What’s the actual, practical application?
Near-instant typesetting, infographic creation, and catalog rendering could be some of the solutions, although fixing a typo would require completely re-rendering the image.
It’s also possible that the more you experiment with Images 2.0 (I’ve only been playing with it for an hour or so), the more the rendered images may look same-y, which is why you’d likely need a skilled human prompter with an eye for design at the helm.
