I continued my exploration of what MidJourney is capable of by generating some 16:9 aspect ratio images centered on the theme of science fiction concept art, a genre the system seems to be especially well versed in.
MidJourney tends to put human figures and female faces into its compositions without specific prompting, likely based on the prevalence of these images in its training data. It’s tricky to remove these figures from generated images by changing prompts or redrawing areas - rather, it’s bests to just reroll and see if the next set of images remove the unwanted aspects.
As with my first exploration, I found that simple prompts were usually best. If the prompt is too specific, the system will not produce what you want, focusing instead on seemingly random nouns. It does not seem to understand the structure of the language but rather seems to pick keywords from the prompt. You basically have to play roulette with it for a few iterations until it produces a good initial image from which you can create variations.
There is a new feature in v5.2 where specific areas of an image can be replaced by providing a new prompt. While this does sometimes improve areas of images that are undesirable, it rarely produces exactly what a specific prompt asks for, so until language recognition is improved in general, this feature doesn’t seem very useful.
It will be extremely useful when these AI image generation systems can be combined with other AI systems so that edits can be made to the images using natural language. It’s inspiring to see the variety and quality of images these systems produce, but it’s also annoying when a small aspect of the image can’t be changed without conventional imaging editing techniques.
Here are a few of the images generated during this run.