OpenAI released an update to their image creation tool yesterday, called 4o Image Generation. The results are impressive, and I have to admit I really enjoy this image they included in their “limitations” section. It’s a delightfully weird collection of things. I almost want a framed poster.
It’s pretty good at making interesting layouts and copying a particular artist. Here are a few examples where I asked it to be late-career Matisse paper cutouts of Harbor Freight catalogs and supermarket circulars:
It does do a much more coherent job of producing accurate text, and arranging things and understanding space and layout than other current image generation tools. It feels like a big leap forward in terms of model comprehension, even if the visual results are not as polished or realistic as Midjourney. Though I still consider the earliest models like VQGAN+CLIP and early Midjourney versions to have more character, which has since been optimized away. Now Midjourney looks too glossy, too perfect.