In January 2021, OpenAI introduced DALL-E, a neural network that the company claims can “take any text & create an image from it,” according to Ilya Sutskever, chief scientist and co-founder of OpenAI. These included concepts it may have never chanced up-on during training.
Now the company is back with an even more impressive version of the previous program called DALL-E. According to the OpenAI website, “DALL-E 2 is a new A.I. system that can create realistic images & art from a natural language description.
Astronauts on horses and more
The website offers many examples of A.I system’s work and they are simply impressive. For example, the words “an astronaut,” “ride a horse,” and “in a photorealistic style” produced the image shown above.
The amazing thing is that A.I. system understands where to place the astronaut on the horse and what a photorealistic image is. OpenAI further explain that “DALL-E 2 can perform realistic edits on existing images from a natural language caption. It can add & remove elements while taking shadows, reflections & textures into account.
The examples show how the program can add features or make them disappear entirely in an image while still leaving in good condition. Finally, OpenAI reveals that “DALL-E 2 can take an image & create different variations of it inspired by the original.” This is illustrated with a painting that differs slightly with each version, while remaining true to the general characteristics of the painting.
How does DALL-E achieve all this?
OpenAI states that “DALL-E 2 learned the relationship between images & text used to describe them. It uses a process called “diffusion” that starts with a pattern of random dots and gradually transforms that pattern into an image as it recognizes certain aspects of that image. Even better, OpenAI’s content policy does not allow users to create violent, adult, or political content, among other categories, making the program safe for everyone.