I put 7 leading AI image generators to the test with the same prompt — here’s the winner

[admin]

I put seven leading artificial intelligence image generators to the test on the same three prompts to see which performed best on prompt adherence, realism and text rendering.

I put 7 leading AI image generators to the test with the same prompt — here’s the winner : Read more

 

[claytoncasey01]

I put seven leading artificial intelligence image generators to the test on the same three prompts to see which performed best on prompt adherence, realism and text rendering.

I put 7 leading AI image generators to the test with the same prompt — here’s the winner : Read more

I'm tired of seeing these articles doing these "same prompt" tests like it's apples to apples. Different models require somewhat unique ways to get the best results. A great result from Midjourney probably won't give a great result with Flux or SDXL for example.

Please start comparing with proper prompts for each and discuss the differences so people might learn something new.

 

[RyanMorrison]

I agree that there are different prompting techniques for different models but I try to write simple descriptive prompts that are relatively generally applicable. Also, this is closer to how most 'average' users might utilize AI so is a good test of how they perform.

I will do some more detailed comparisons with custom prompts as I agree that is also a good way to explore the potential. I did point out the limitations of this prompt for Midjourney by relying solely on prompting without utilizing parameters or specific keywords.

 

[M5MSM]

Would you be able to show us the results with the other 2 AI generators you mentioned?

 

[binzer]

I plugged all 3 prompts into Google's Image FX. It outright rejected the second 2 prompts. I believe this is because the Beatles and children were mentioned.

I quite like its results for the first prompt, but I've never used a more frustrating image generator (Link to ImageFX result).

 

[goatonastick]

This seems more like a test of how one specific form of prompt writing works on several models.
This prompt seems too long for some models, such as flux where they dont recommend more than 22 tokens, and some models, like the ACTUAL most customizable model, Stable Diffusion, works better with simple words or phrases separated by commas.
A great opportunity to teach people some good general rules, like how simpler and more concise prompts lead to better adherence, or how important prompt order is, is lost because the author just wanted to cut and paste a description and see what happens. Now, they are led to think there’s only one or two models that actually work, and the rest aren’t even worth their time.
I also noticed that, besides two of the scenes being in the streets of London, the article didnt clearly express their goal. Was it how closely they followed the prompt? Because the author often complained about the lack of realism in some.
Was is how realistic they were? Because the author didn’t use any terms to generate realistic scenes in the second two prompts, and even made the mistake of using the term “realism” in both of them, which are actually bad practice for trying to generate photorealistic images, because no normal person describes an actual photo as “realistic”, therefore the models often include references of drawings and illustrations, which are what the words realism/realistic tend to be used for. It’s even to a point that it’s recommended that you put “realistic” in the negative prompt (when possible) to ensure you get photographs rather than artwork.
It’s suggested you specifically mention the words photo, such as in the first prompt, or, for more precision, a focal length like 35 mm (or even a specific brand and model of lens!), because that sort of description would only be attached to actual photos in the training set.

 

[StefonR]

Unfortunate that all of the AI have a bias to interpret attractive as white or Caucasian. If this is the implicit bias it has, who knows what other bias has been trained into it and it will be acting on as these systems get more deeply integrated into society.

 

[iampat]

I put seven leading artificial intelligence image generators to the test on the same three prompts to see which performed best on prompt adherence, realism and text rendering.

I put 7 leading AI image generators to the test with the same prompt — here’s the winner : Read more

I ran these through imagen 3, default config, no edits to prompts. Only the first prompt was worth a hoot, the others were very cartoonish and had terrible compositions and jumbled text

 

[Wolfonacable]

Unfortunate that all of the AI have a bias to interpret attractive as white or Caucasian. If this is the implicit bias it has, who knows what other bias has been trained into it and it will be acting on as these systems get more deeply integrated into society.

Really? How did you reach this conclusion? The results from the first prompt returned varied results with, among others, what I assume was a mixed-race woman and a woman of east Asian heritage. I also, and please correct me if I'm wrong, didn't see any mention in the prompts of attractiveness...

 

[GerryP]

Unfortunate that all of the AI have a bias to interpret attractive as white or Caucasian. If this is the implicit bias it has, who knows what other bias has been trained into it and it will be acting on as these systems get more deeply integrated into society.

I haven't tried all the image generatgors. I prefer to master one rather tool than spread myself too thin. I've been working with DALL-E a lot. It defaults to a male caucasian unless specifically told to do otherwise. Similaily, it's people are all slim, young and way too pretty. When prompting DALL-E, specify what you want in these areas.

 

[we be digi]

Was this sponsored content. Accuracy to prompt may have been the best measure but I was most impressed by Freepik Mystic, ideogram ranks lower in my opinion.