What changes now that DALL-E 2 is in beta, and how do its credits work?

DALL-E 2, OpenAI's system that creates images from a written description, is opening up. Yesterday the company announced that the tool is now in beta and that it will invite 1 million people from its waitlist over the next few weeks. Until now it has been a research preview for a limited group, which TechCrunch reports had recently passed 100,000 users.

The beta brings two big changes. DALL-E, free during the preview, now runs on a system of credits that you can top up with money. And for the first time, people may use the images they make in commercial work.

What OpenAI announced

OpenAI posted the news on its blog on 20 July 2022. The key points:

• Wider access: about 1 million waitlisted people are being invited, gradually.
• Credits: each user gets free credits every month and can buy more.
• Commercial rights: users can now reprint, sell and put their images on merchandise, including images made during the preview.
• Safety updates: stricter rules on faces and public figures, better filters and a new way of making images of people more varied.

Artists who need financial help will be able to request reduced-cost access, and OpenAI has opened an interest form for them.

How DALL-E 2 makes an image

DALL-E 2 was first shown in April. You type a prompt, a plain-language description such as "a watercolor of a lighthouse at dawn," and the system produces matching pictures. Under the hood it uses a diffusion model: it begins with random dots and changes them, step by step, into an image that fits the description. It can also edit a picture, changing part of it from a written instruction while keeping shadows and textures consistent, and make variations, new images inspired by one you give it.

In the beta, these features sit side by side with a new My Collection space for saving the images you like.

How the credits work

A credit is the unit you spend each time you ask DALL-E to make something. What one credit buys depends on the action:

At the paid rate, each credit costs about 13 cents, and each image about 3 cents. A credit is spent every time you press "generate" or "variations," whether or not you like the result, so it pays to plan prompts before you click.

Here is a simple way to make credits go further. Start with a clear prompt that names the subject, the style and the setting:

A cozy reading nook by a rainy window, warm lamp light, a sleeping orange cat on a blue armchair, digital painting

If one of the four results is close, spend your next credit on a variation of that image or an edit of one detail, rather than starting over with a new prompt.

New safety rules

OpenAI says it spent the preview working with researchers, artists and developers to understand the risks, and it is tightening several rules for the beta:

• Faces and public figures: uploaded images that contain realistic faces are rejected, as are attempts to depict public figures such as celebrities and politicians. OpenAI also uses techniques to stop the system from producing photorealistic faces of real people.
• Content filters: the filters that block violent, adult or political images have been made more accurate. The most explicit material had already been removed from the training data.
• Bias: a new technique, applied at the system level, makes images of people better reflect the world's diversity when a prompt names a person without saying their race or gender. OpenAI's example is "CEO."
• Monitoring: automated and human review continues.

TechCrunch notes that these questions affect the whole field. Image generators learn from millions of pictures gathered online and can absorb the biases in them. Google, for its part, has said it will not release its own system, Imagen, because of the risk of misuse.

Can you sell or publish the images you make with DALL-E?

Yes. Starting with the beta, OpenAI gives users full usage rights to commercialize their images, which it defines as covering reprints, sales and merchandise. This also covers images created during the research preview, when commercial use was not allowed. OpenAI says users have told it they plan to use DALL-E for children's book illustrations, newsletter art, concept art and characters for games, mood boards for design work and movie storyboards.

The content policy still applies to everything you make, so images must stay within OpenAI's rules on violence, adult content, politics and real people. And OpenAI describes the current pricing as a first phase, saying it may explore other options as it gathers feedback.

For a small business or a teacher, the practical upshot is that DALL-E can now produce usable illustrations for a few cents an image. The best way to get value from it is to treat each credit as a small experiment: write a precise prompt, keep the results you like in your collection, and build on them with edits and variations.