Midjourney
A brief post on generating images with Midjourney
Midjourney is a research lab led by David Holz who co-founded Leap Motion. They’ve recently released an artificial intelligence image generator of the same name that is capable of creating high quality images based solely on a textual description of its contents.
There is very little information available regarding the technology behind Midjourney. There is some speculation that its architecture is based on Generative Adversarial Networks (GAN), which were invented in 2014 by Ian Goodfellow[1]. GANs use two neural networks; one that generates an image and another that evaluates it based on existing examples of the target result, giving it a score for accuracy. After evaluating the image, the data propagates back to the first neural network which learns from the feedback and modifies the image for the next iteration of the loop. The process is repeated until the generated image satisfies the criteria of the second neural network.
Midjourney is not the first image generator to gain popularity in recent years. DALL·E and its successor DALL·E 2 by OpenAI uses a 12-billion parameter version of the popular GPT-3 transformer model[4].
Midjourney is currently in open beta and anyone can register and generate about 25 images for free.
Below are some of my own experiments with it.
Comments