Aug 13, 20234 min read

/Describe Greebles, Please

There are two things I have recently discovered using the AI image engine, Midjourney, that made me pause when I recognized their power. One is very simple - the word "greeble." The other is much more complicated - Midjourney's "/describe" command. Let's get the easy one out of the way first.

The word "greeble" refers to the object of a practice started in film special effects. It involves cannibalizing model kits (or anything else small and fiddly) and gluing these bits - the greebles - on to a large plain surface to break up that surface and make it more visually interesting. This process is called "kitbashing." Science fiction movies pioneered this technique to jazz up spaceships and Death Stars in the days before computer modeling. See the examples below from (or inspired by) the original "Star Wars." The practice continues today, but now all the greebles are computer generated.

What I discovered, quite by accident, is that Midjourney LOVES the word "greeble." It's as if the AI is primed for that word and makes a lot of very strong and often exuberant choices when it receives "greeble" as part of a prompt. For example, the images in the gallery below were generated by just the word "greeble" and nothing else:

A friend of mine, Ugo Fist, described images like these as being of something and nothing at the same time, which struck me as being very appropriate and is perhaps the best description of AI-generated imagery that I've heard yet.

If you add "greeble" to a simple noun, as in e.g., "a cat made of greebles," MJ produces results like the ones in the gallery below where the technique was used for cat, ghost, crow, and shark.

MJ treats locations like cities a little differently when asked to greeble them. In these cases, the result is usually a highly detailed and intricate model of the location based on stereotypes for that location, often in a stacked, vertical way. Below are London, Japan, Amsterdam, and the Bay of Naples.

When instructed to make images of rural places out of greebles, MJ still really wants to jam a bunch of buildings and structures in there. Below are two renderings of rural Portugal, followed by Yellowstone Park and rural Tuscany.

Before we leave greebles, the gallery below consists of images made by feeding an image into MJ and requesting it to be greebled. The image that was the seed was a still from a stage production of "Frankenstein Revived" designed by accomplished set designer Ken MacDonald. What's interesting here is that the original photo, while very colourful and visually arresting, is not obviously a stage set. Somehow, though, MJ figured out that it is a stage set and produced images that look very much like set maquettes for immense space operas.

How I stumbled across the word "greebles" involves the other term mentioned in this post's title, the "/describe" command recognized by Midjourney. I only very recently learned of this command, with all of my MJ interactions previously limited to the "/imagine" command. With /imagine, you enter a text prompt, MJ thinks about it for a bit, and then produces a 2x2 grid of results based on the prompt. For example, the grid shown here was made with the prompt, "Portugal made of greebles, photorealistic, in the style of hasselblad 1600f." Once you have a grid of results, you can select one or more for enlarging and/or making variations of. A fairly new development involves the ability to take one of those results and zoom out or pan up, down, left, or right - very handy.

The "/describe" command does exactly the opposite. It allows you to feed an image in MJ and then the AI comes up with four prompts that could have generated that image. You can then generate images based on those prompts or, and this is where it gets really interesting, pick and choose words and phrases from among all the options to make an entirely new prompt. Pictured here is an image that I had generated with the /imagine command. When I used this image as a seed for /describe, I got the following results:

"1. the digital artist walterny is standing in an underground city, in the style of dark black and bronze, mexican folklore-inspired, concept art, intricately sculpted, mysterious jungle, lynd ward, 8k --ar 85:128

2. the lady in the city looks like she is standing in the middle of a cave, in the style of spatial concept art, mesoamerican influences, dark gold and dark black, realistic and hyper-detailed renderings, greeble, junglecore, 8k --ar 85:128

3. a piece of art about an evil woman from the city of kaziena, in the style of highly detailed environments, dark cyan and dark amber, mayan art and architecture, james paick, high contrast chiaroscuro, ps1 graphics, elongated figures --ar 85:128

4. a woman in a dark place, with lots of metal, in the style of mayan art and architecture, perspective rendering, intricately mapped worlds, highly stylized figures, contrasting backgrounds, lovecraftian, ancient egypt --ar 85:128"

The word "greeble" jumped out at me as something I was unfamiliar with and the rest is history. As you can see, though, the prompts as generated by MJ provide some insight into how the AI parses prompts and well as possibly fruitful words and phrases. I'm not thrilled by how often MJ invokes the names of artists in these prompts, but that concern is mitigated by my efforts to look up some of the artists mentioned and sometimes completely fail to see the stylistic connection. For example, two artists are invoked in the prompts above, Lynd Ward and James Paick. I can see some stylistic overlap with James Paick, but Lynd Ward's stuff seems completely unrelated, as shown with the image search below:

In my pursuit of the ethical use of the AI, though, it's very easy to recombine these prompts and simply leave the artists' names out. For example - "in the style of spatial concept art, dark gold and dark black, dark cyan and dark amber, realistic and hyper-detailed renderings, high contrast chiaroscuro, a man in a dark place, with lots of metal, perspective rendering, intricately mapped worlds, highly stylized figures, contrasting backgrounds, lovecraftian." Running this prompt produced the following grid:

So there you have it, greeble and /describe, two concepts that have been occupying my MJ time a lot as of late. I hope you find them as interesting and useful as I do.