What does “a pink guinea pig picking a gala outfit from a wardrobe full of sharks” look like according to an AI? Lately, a number of AI art generators have challenged people online to come up with the most unexpected descriptions to through artificial intelligence generate fictional, sometimes highly grotesque imagery. But how does the technology work? And what might it be used for in the future?

There are several different ways to use AI to generate images. One method involves using deep learning algorithms to train computers to recognize patterns in data. Once trained, the algorithm can then produce similar results when presented with new data. Another approach uses generative adversarial networks (GANs), which involve two separate models competing against each other. A third technique uses reinforcement learning, where the system learns by trial and error.

Two applications that have recently received attention in the media are DALL-E and Midjourney. DALL-E developed by OpenAI uses a deep learning model and a version of GTP-3 modified to generate images. Midjourney is developed by a smaller research team, and in a similar way utilizes neural networks for image generation. Both applications are in open beta for private users. If you want to try out DALL-E you can sign up for access on this link. 

Backpack generated in Midjourney using text prompt “techwear backpack made of plastic and precious metals”

How can this be relevant technology for the fashion industry?

Image content and prototyping

Generating images based on large masses of data has the potential to become valuable technology in fashion in many different areas. Stock photo services might see themselves replaced by AI applications in the future as the technology gets more refined and precise. AI generators have the ability to take large numbers of images to create a new, realistic-looking one yet at the same time stretch human imagination beyond its limitation and biases. If a product designer e.g. wants to create a large number of iterations of a prototype, AI image generators can deliver this in minutes or even seconds. In this article you can see an example of how an AI fed with image data of shoes can be trained to output new nonexisting designs from it.

Trend analysis and forecasting

Activating masses of data that might be hard to quickly overview with the human eye can be used within trend analysis and forecasting where numbers and statistics don’t provide a quick or good enough overview. Photo and image production will probably in the future be partly replaced by AI-generated imagery or work complementary to other technologies and manual work. 

Digital art and character design

The film, games, and digital entertainment industries will be able to use similar applications in creating characters or environments. At the current state of the open betas, one can already now use the output as concept art and inspirational input whereas in the future it’s easy to imagine an AI within a more strict and refined framework creating production-ready assets. The approach of being able to quickly iterate and modify images through manual feedback also goes way beyond what humans can possibly achieve in the same timeframe. 

Midjourney using text prompt: “sweater with NASA typography and mystic necklace”

Stable Diffusion using text prompt “bright blue jacket with neon knitted bands”

Midjourney using text prompt: “a vaporwave ecommerce store with shoes made of clay and electricity”

Stable Diffusion using text prompt: “A pink guinea pig picking a gala outfit from a wardrobe full of sharks”