Testing GPT-4o's Image Capabilities for Education

OpenAI announced their new GPT4o image generator, calling it a "game-changer for educators." They showed some incredible examples, including a correctly labeled cross-section of a cell.

Our team here at AI for Education loves an image generator, so we all dug in to explore further and found that the technology is not quite there yet.

There's no doubt that the new 4o image generator is a major improvement on the Dall-E integration. Being able to directly prompt GPT4o to create the image instead of the LLM prompting an image generator means more control over the outputs.

We also love the improvements in text rendering and the consistent characters. And Mandy DePriest on our team got it to correctly render the lyrics to Abba's Dancing Queen, her personal benchmark.

But when we used the same prompt from the original post to try and duplicate the labeled cell, we repeatedly came up short. Despite trying various prompting strategies, it could never accurately create a diagram of a cell and label the parts.

For comparison's sake, we ran the same prompt through multiple generators, with poor results. While image generators are improving quickly, it's clear they still have a lot of room for improvement.

But sometimes, mistakes are the best teachable moments. As Hannah Ketteman on our team shared, one of her favorite things is to produce an image like these and then have students find and correct the errors - a great way to assess students while building their AI literacy.

What do you think of 4o's new image capabilities? What are your favorite applications and those you are hoping to improve?