OpenAI's Picture-Drawing Program Now Makes Images That Look Like Photographs

DALL·E 2 produces sharper, more realistic images than last year's version, and can edit an existing photo from a typed instruction. It is available only to a small group of invited testers.

OpenAI showed the second version of its image program today. DALL·E 2 does what the first one did — makes a picture from a written description — but at four times the resolution and with a realism the first version never approached. "An astronaut riding a horse in a photorealistic style" produces something that looks, at a glance, like a photograph of an astronaut riding a horse. The lab's examples include a teddy bear mixing chemicals as a mad scientist, a bowl of soup that is a portal to another dimension, and a fox in the style of Monet.

The program can also edit. Give it a photograph and an instruction — add a flamingo to this pool, remove the dog from this couch — and it makes the change, matching the light and shadows of the original. It can produce variations on an existing image, keeping the subject and changing the composition.

OpenAI is not releasing it publicly. Access is limited to a few hundred invited testers under rules that forbid making images of real people, violence, or political content, and the lab has built filters to refuse such requests. The stated reason is to study the risks before opening it up: a program that makes convincing photographs of things that never happened is an obvious tool for deception.

The program has trouble with text in images and with counting, and it sometimes reflects the biases of the photographs it learned from — "a lawyer" tends to come out as a man. But the testers who have posted results describe the same reaction: they typed something absurd, and the picture came back looking real.