ImageGPT 2.0 is on the top rankings of most of the “AI arena” testing websites. Those ones give the users a prompt and two images results from that prompt from two randomly selected models and ask “which of these follows the prompt better” or “which of these two images are better” (which are two separate questions and depend on what the particular site is trying to assess). So it’s probably the top ranking model by consensus right now, though these tests are biased towards explicit prompt following. You might think that’s the obvious thing to rank by, but I wouldn’t be so sure. Midjourney arguably produces the most beautiful images of any image generator, and sometimes it ignores parts of your prompts to do so. So they’re really just different tools rather than purely better or worse.
Midjourney, incidentally, is not included in these competitions because they don’t have an API and therefore they cannot be automated to generate thousands of images from prompts for these tests. Someone would have to do it manually. I would be really eager to see how Midjourney does, because it’s its really its own unique thing and very different from how something like ImageGPT or gemini / nano banana work. The latter are definitely better at following a specific prompt, but I find that midjourney’s images are more beautiful and the midjourney interface is a much more interesting journey of creative exploration. It’s almost two entirely different tools and workflows even though the end result is a series of images.
I generate about 90% of my images on midjourney even if the autoregressive nano banana and imagegpt 2.0 are “better.” I have a separate thread about midjourney specifically if you’re interested.