Reverse-engineering Midjourney to get those creative prompt words

When I give a presentation about generative AI and the world it is about to create for us I talk about how all of us are prompt engineers now, or at least will become that very soon. And one of the things that is the hardest about being a prompt engineer is to stretch your vocabulary and dig out those words that you know what they mean, but rarely use.

Midjourney images are created from random noise and pixel predictions, but each word carry a large weight. Sometimes you will stumble upon a picture that looks great, but you have no idea what words it was that made it look the way it did. Here's an example:

/imagine behind the scenes of an AI image generation factory

I wanted to use this prompt to generate an illustration for a slide about image generating AI's, and this prompt generated, among others, this quite cool image.

Although it looks great, I think it could look even cooler but I have no idea how to replicate this specific style. The initial prompt also generated a bunch of other images, in very different styles, but I love the complexity, tonality and lighting of this specific image. So what do I do? Fortunately, Midjourney has a great solution.

It is called the /describe command. With it you hand an image to Midjourney, and it offers you a few alternative prompts that could be used to generate something similar. I type /describe and press enter. Discord then asks me to upload an image to send to Midjourney, so I select the one I got above.

After pressing enter once more the image is uploaded and I get proposed prompts back.

It even offers handy buttons below the image to press, and if I click either of those it will run that proposed prompt for me.

When you press one of them, a remix dialogue comes up. This dialogue box offers the ability to modify the prompt before it is generated, by adding or removing words as you please.

What I often end up doing is trying a few of the prompts as they are proposed, just to see what they end up creating for me. But in this case, one of them immediately gave me this very cool, but slightly different image, that I just love.

In the end the prompt that Midjourney proposed and that I run, without modifying it, was this one:

a man working at an office in a factory, in the style of luminous 3d objects, atmospheric perspective, aerial view, photo-realistic landscapes, cinematic sets, hyper-realistic water, steampunk

One of the benefits of this approach is that not only do I get nice images, I also learn new words for my prompts. There are a few keywords/tokens in there that I think have an outsized effect, namely: "man working", "office", "factory", "luminous", "atmospheric perspective", "aerial view", "steampunk", but of course the whole prompt was important. Nonetheless, the next time I struggle to replicate a certain style that I want to achieve, I now have a few more words to use.

Let me know what you thought of this technique and if you end up getting any nice pictures, or learn new keywords, in the process. I'll also link to a few previous articles I have written about Midjourney in case you're looking to get started.

Other Midjourney articles I have posted

• Getting started with Midjourney
• How adjectives can shape your Midjourney AI photographs
• Using weights to tweak your Midjourney prompts
• Using image prompts and seeds to put patterns on things in Midjourney