Midjourney vs. Flux. Which AI image generator reigns supreme?

At Indie Hackers, we use Midjourney to generate all of our cover images, and it's been working out pretty well. But Midjourney now has some competition from Flux, the new AI image generator from Black Forest Labs.

Judging by the examples they’ve been putting out, it looks really good.

But, is it better than Midjourney?

That’s what I’ll be finding out in this article by using both models to generate some cover images.

My first muse today is the beautiful Mubashar Iqbal (aka Mubs), who just gave me a great interview giving his advice for new makers. For Mubs, the background image we will be using is this:

Let’s start with Midjourney. 

For those who don’t know, Midjourney is a Discord bot. So the way you use it is by joining their Discord server, typing /imagine, and then putting in your prompt.

The prompt we use to generate cover images for interviews is: 

Cover image for an article about a startup founder with a background of ___, URL_TO_ORIGINAL_PHOTO, blind contour illustration, background color ___. --ar 16:9

So for Mubs, the prompt is:

Cover image for an article about a startup founder with a background of https://s.mj.run/xlVJtCzMT7A, blind contour illustration, background color maroon. --ar 16:9

That returns these four options:

I really like version 3 (bottom left corner) because it actually looks like him and I’m digging the background, so let’s do some strong variations on that and see what we get:

These aren’t great because Mubs doesn’t wear glasses, so let’s try to tone it down with some subtle variations instead:

These all look pretty similar, but I like #4 (bottom right corner) the best because I think the beard and hair are most like Mubs’s. So, here’s the final output from Midjourney:

Now, let’s try Flux. 

To do so, I’m using Flux Pro via Replicate, a front-end that lets me use Flux’s best model for $0.055 an image. However, if you prefer to access it directly via the API, you can do so here. 

So, let’s try out the same prompt and see what we get:

That’s not great. Let’s see what happens if we get rid of the glasses, increase the number of diffusion steps to 35, and decrease the interval number (which makes it adhere to the prompt more) to 1.5:

That’s definitely better, but I still can’t get rid of the glasses. Let’s try raising the guidance number to 4, which should make it stick to the prompt even more:

MORE GLASSES! UGH! Let’s change the prompt to include “don’t under any circumstance put glasses on him” and set the guidance number to 5 and the interval number to 1, the maximum amounts allowed:

Ok, I give up. Flux refuses to take off Mubs’s glasses. So, score the interview article point for Midjourney.

Now, let’s try a more abstract cover image. 

For example, what if we had an article on AI agents? In this case, our prompt for Midjourney would be cover image of an AI agent. blind contour illustration, background color light blue. --ar 3:2. That gives us:

#1 (top left) is really good, so in just one prompt Midjourney gave us a really good abstract image.

When we run the same prompt in Flux, we get:

Not gonna lie; that thing is pretty cool. If I give it a bit more creative freedom by lowering the guidance number to 2.5 and increasing the interval number to 2.5, we get:

Sheeesh. That thing looks insane, but it’s impossible to say if it’s “better” than Midjourney. Midjourney’s is more abstract, while Flux’s looks like something straight out of Ex Machina. It just depends on which one is more your style. 

In any case, it’s amazing that just a few years ago, this was considered state-of-the-art AI image generation:

Now, the same prompt of “nursing school leader” generates this:

And with the amount of brainpower and money now in AI, things will only get better from here.