Trialled by Daniel Zuboff
Strategy at Loomery
Generative AI isn’t just confined to text. Its ability to create images has been accelerating for quite some time now. At first, Generative Adversarial Networks (GANs) were at the forefront of research, but now diffusion-based models are capable of producing detailed pictures from text prompts. Open AI, Microsoft, Nvidia and others have built progressively more capable systems, but Midjourney has recently made an especially important advance with their implementation of version 5 of their underlying neural network.
We’ve already explored the text-to-image generation capabilities of Runway’s image and video offerings in Turbocharging video editing with AI, but Midjourney specifically focuses on image generation and manipulation with a bunch of tools to create, modify, and merge images. In a kind of experiment, I’ve given it many of the same inputs I used in the Runway article as a way of comparing the two models.
Midjourney doesn’t provide its service on a standalone website, instead, you need to access the hardware and interact with the software through Discord. Once you join Midjourney’s official account and start a monthly subscription (they used to offer a free trial but halted it due to misuse), you’re free to interact with the Midjourney bot via a list of commands. You can find an easy-to-read guide to all the things you can ask the bot to do on the Midjourney website.
You’ll have to join one of the newcomer rooms after paying to begin generating. A basic subscription (with a monthly cost of $12) gives you 3.33 hours of fast generation time. Typing /settings gives you access to many useful options, but you can also specify these in your image prompts which we’ll come to later.
I’ve chosen the latest MJ version 5 model for all the following images. There are also easy-to-select options for quality (how long you give the computer to generate the image), Style (how biased the software is towards artistic and abstract images) and whether you’d like to remix images (modify generated images or ones you provide as inputs).