My Attempt at Blending Image Genres in Midjourney

Alex Tully
5 min readAug 7, 2023

--

I decided on a different direction for this piece. Rather than go for images that at least looked like real photos, what about stepping in to the surreal? I’d used the /blend function to achieve photographic compositions too complex for the bot (Rice Terrace Ziggurat and The Roll Monitor), but now I was looking to explore “bleeding” a non-photographic image into a photographic-looking one. Below’s my final result, and after that I’ll give an account of the process to make it.

My aim was to have a photo of a boy behind a school fence, but to include obviously unreal elements suggesting that the true nature of the place was a prison. In the past I’d tried to achieve it by blending a film noir style photo of a school with an oil painting of a prison. The chain-link fence was perfect as a linking element, but the results looked too photo-like. However I was still convinced that the way to go was to blend a photo of the real together with an illustration of the unreal. But I wondered if I’d get a better effect by changing the genres of each parent image.

I made the photo part infrared, in the hope that it’d be easier for the bot to handle the combination of “partially surreal + unreal” than “realistic + unreal”. My prompt was: infrared photograph of a depressed schoolboy sitting in a schoolyard with a schoolbag, rusting chain_link fence, Australian school_uniform, white polo_shirt — no necktie — s 50 — style raw

EDITED TO ADD: I put underscores between some of the words, and I explain why in this post.

And I got:

I should note here that Midjourney went screwy when I put in “ — no tie”. I guessed this was because “tie” can mean other things than “a thing you put around your neck”, so I changed it to “necktie” and got better results.

For the illustration parent image, I decided to go with psychedelic art, because this makes elements appear obviously unreal, and I hoped that those elements would retain their obvious unreality in the blend. At first I prompted: in a prison_yard, chain_link fence, asphalt, concertina_wire, watchtower, psychedelic, psilocybin_visuals — s 50 — style raw, and got the below four images.

Although the top right image had the elements how I wanted, none of these images looked psychedelic enough. So I decided to add extend the prompt with “fractals” and “Mandelbrot”. That made things too psychedelic, with Midjourney rolling the wire up into balls! So there was a bit of tweaking to do with the keywords to get the right balance, however eventually I got plenty of usable pictures. I finally settled on the below:

watchtower in a prison_yard, chain_link fence, asphalt, barbed_wire, psilocybin, fractals, Mandelbrot, psychedelic — s 50 — style raw

I knew I was getting warmer when I blended the above two images together and got the below result:

I then followed my tried-and-true method of using that as the first element of a prompt, followed by text sub-prompts for each of the two images that had gone into the blend (separated by double colons :: ). So this prompt was: <URL to the Above Image> infrared photograph of a depressed schoolboy sitting in a schoolyard with a schoolbag, rusting chain_link fence, Australian school_uniform, white polo_shirt:: watchtower in a prison_yard, rusting chain_link fence, asphalt, barbed_wire, psilocybin, fractals, Mandelbrot, psychedelic — no necktie, extra_limbs — s 50 — style raw

Here I ran into the same headache that had happened at the same point in The Roll Monitor: The AI deleted the schoolbag and the watchtower! Fortunately I could resolve it the same way as I had then, which was by including those items in a separate text sub-prompt (separated from the rest by double colons :: ) For emphasis, I’ve highlighted that sub-prompt in bold, and placed each sub-prompt on its own line. Here it is:
<URL to the Above Image>
infrared photograph of a depressed schoolboy sitting in a schoolyard with a schoolbag, rusting chain_link fence, Australian school_uniform, white polo_shirt::
watchtower in a prison_yard, rusting chain_link fence, asphalt, barbed_wire, psilocybin, fractals, Mandelbrot, psychedelic::
with a schoolbag, watchtower
— no necktie, extra_limbs — s 50 — style raw

From there it was just a matter of zooming out and then cropping.

I’ve titled it “Australian Public Education”, and it ticks all the boxes. The watchtower has been reduced to a weird triangle above a gate, but that’s still enough to bring menace into the photo. This is reinforced by the razor wire looping around overhead, but in an obviously unnatural way.

If you want Midjourney to make a photograph of one thing, with a surrealist undercurrent of something else, then please give this technique a try!

--

--

Alex Tully

Into Generative AI, but 100% Human-Written Blog (every word)・Bachelor’s in Maths・Master’s in Linguistics (@ANU 🇦🇺 )・Taught myself 🇯🇵 and 🇹🇭・Digital Nomad