Different AI image models can give you widely varying results, even with identical prompts. I did some testing using the same prompt across several AI image tools: Nano Banana, ChatGPT, Midjourney, and others. Overall, the models have all advanced enough that the results were reasonably good across the board (with a few notable failures). In this post, I’ll give you the prompts and results I used to compare different AI image models.
AI image models compared
These are the specific models I used for testing. Except for Midjourney, I generated all of these within Flora. Flora makes these sorts of comparisons much easier, since you can add the prompt once, select multiple models, and generate them all at the same time.
Nano Banana 2GPT Image 1.5Midjourney v7 or v8.1Flux 2 ProRecraft v4Seedream 5.0 Lite
Prompt 1: A single person
For my first test, I wanted a photorealistic image with a focus on a single person. As I expected, all of the results were pretty good; this is something all the tools can generate now. I like the aesthetic and pose for some results better than others, but none of these were failures.
Editorial photo, an Indian woman, mid 40s, dark hair pulled back in a single long braid, wearing a kurti with a small burgundy geometric pattern, walking in past desks in a modern open office, trees visible through the windows
Prompt 2: Photorealistic objects
For my second test, I used a prompt for two objects: a hard hat and a pair of safety glasses. There were a few issues with these images (the plywood in the Nano Banana image looks too worn, for example), but they weren’t bad. All of the models followed the directions for the focus objects, setting, and camera angle.
Editorial photo, a yellow hard hat and a pair of safety glasses on a stack of plywood sheets on a construction site, low angle
Prompt 3: Illustration
For this prompt, I used the style of “digital illustration” as a starting point. However, Recraft and Flux both failed to generate illustrations initially; they generated photorealistic images instead. I had to change the prompt to “vector illustration” for those image models. With illustrations, you see more variety in the default styles. Obviously, I don’t have a lot of detail in the prompt about the style; that would affect the similarity more.
While it’s not my favorite of this set, I was pleasantly surprised at how well ChatGPT did. The most recent updates to ChatGPT’s image model have made it much more comparable to other tools. It doesn’t always have the yellow filter, and the overall illustration style has greatly improved. His back blends into the back of the bench in this example though, so this would require some editing or iteration to be usable.
Digital illustration, modern corporate style, a Black man typing on his phone, thoughtful expression, wearing a dark blue t-shirt, sitting on a bench in a park, medium shot from the waist up, viewed at a side angle
Prompt 4: Group of people plus text
For my final prompt, I wanted to challenge the models. This time, I prompted for a group of people, plus I asked for text on a whiteboard in the background. The more people you add to an image, the more likely it is that something goes wrong in the generation.
The Recraft image here is the biggest failure out of all of the images for all of the prompts. All of these images have some issues though; this is still where the image models require a lot more iteration and editing to get it right (if it’s possible at all). Most of these images look very obviously AI, especially in the text. Midjourney looks a little better partly because it didn’t actually try to make readable text.
Candid photo of a racially and ethnically diverse team of designers discussing solutions, sitting around a light wooden table in an office conference room, gray walls, whiteboard with brainstorming mind map on one wall
Reflections on the differences
Reflecting on these results, I still find myself drawn to Midjourney. I like the overall look of Midjourney more, and I find that it’s easier to generate a wider range of faces. But Nano Banana is great, and better than Midjourney for consistency (although I may revise that opinion after I test the new Midjourney v8.1 a little more). ChatGPT has improved enough to be genuinely useful now, although it still looks more like stock photography. Flux, Recraft, and Seedream are all useful in some situations too.
If you’re going to pay for a single image tool, I recommend Flora or Freepik. Even though they don’t have access to Midjourney, both Flora and Freepik give you access to multiple image models. You have a lot more flexibility that way.
Your thoughts?
Which examples do you think turned out the best? Which models do you prefer working with? Let me know in the comments.
Further reading
Check out my comparison of AI image generation tools for more information on the strengths and weaknesses of different tools.
