OpenAI LP today introduced DALL-E 3, an artificial intelligence model that can generate images based on natural language prompts.

The model is the latest addition to a series of image-generating neural networks that the company first introduced in 2021. Compared with its predecessors, DALL-E 3 can generate higher-quality images. OpenAI says that the AI is also capable of understanding user prompts more accurately.

DALL-E 3 will initially become available through the paid versions of the company’s ChatGPT service. To use the new model, a customer must enter a description of an image into the ChatGPT interface. The chatbot analyzes the user’s description, creates a set of relevant AI prompts and then inputs those prompts into DALL-E 3, which in turn generates several different versions of the requested image.

OpenAI says DALL-E 3 is better at understanding natural language requests than its predecessors. In a demonstration video on the company’s website, a user asks DALL-E 3 to generate an image and then instructs the model to make about a half-dozen modifications. At one point along the way, ChatGPT draws on the information in the user’s prompts to generate a short story.

The new model can not only understand more complicated prompts than earlier neural neutrals but also generate higher-quality output. In one internal test, the company entered the same prompt into DALL-E 3 and a previous-generation neural network it released last year. The former model generated a higher-quality image.

“Modern text-to-image systems have a tendency to ignore words or descriptions, forcing users to learn prompt engineering,” OpenAI researchers wrote in a blog post today. “DALL·E 3 represents a leap forward in our ability to generate images that exactly adhere to the text you provide.”

OpenAI has also enhanced the built-in safety features. According to the company, DALL-E 3 declines requests to draw likenesses of public figures and includes mitigations against AI bias. OpenAI said that it developed the mitigations with help from external experts.

According to the company, DALL-E 3 rejects user prompts that ask it to generate images in the style of a living artist. It also includes a tool that will help OpenAI determine whether an image may have been generated by the model. Last month, Google LLC researchers detailed a similar tool called SynthID that uses a subtle watermark to track AI-generated files. 

DALL-E 3 is rolling out two years after the first neural network in the DALL-E series made its debut. Going forward, OpenAI will allow artists to have their images removed from the training datasets it uses to train new generative AI models. Artists can request that the company avoid using their content through a new form on its website. 

The new model is currently in preview. It’s set to become available through the paid Enterprise and Plus versions of ChatGPT early next month. Later in the fall, the company plans to make it accessible via an application programming interface, which will enable developers to incorporate it into their software. 

Image: OpenAI