AI draws dog-walking baby radish in a tutu

These pictures were generated from the command "An illustration of a baby daikon radish in a tutu walking a dog"

Avocado-shaped armchairs, radishes in tutus and cats in sunglasses are among some surreal works of art created by an artificial-intelligence system.

Dall-E was trained to generate images from short text-based descriptions, using 12 billion images and their captions sourced from the internet.

Creator OpenAI previously produced a text generator that turns short phrases into stories, poems and articles.

Experts said the results of its latest innovation were impressive, if patchy.

• Have we seen our AI future?

• 'Dangerous' AI offers to write fake news

• DeepMind AI agent could turbocharge YouTube

Combining an understanding of both language and pictures will be the key to making AI smarter and the models devised by OpenAI make good strides towards this, most researchers agree.

These pictures were created from the phrase "An armchair in the shape of an avocado"

OpenAI blogged Dall-E - named after surrealist Salvador Dali and the little animated robot Wall-E from the film of the same name - was a diverse artist, "creating anthropomorphised versions of animals and objects" and "combining unrelated concepts in plausible ways".

But it admitted the success of the model depended on how the descriptions were phrased.

"As more objects are introduced, Dall-E is prone to confusing the associations between the objects and their colours, and the success rate decreases sharply," researchers wrote.

Users can create their own illustrations, including combining animals such as cats and snakes

Users are also invited to create their own images, changing parameters such as colour, shape, and design.

AI systems could have "significant, broad societal impacts", the Californian company, started in 2015, blogged.

And the potential for bias, the impact on jobs and other longer-term ethical challenges would need to be considered.

Attempts to combine a snail and a harmonica were not always successful

Google deep-learning engineer Francoise Chollet tweeted Dall-E's work was "really cool".

But Georgia Tech associate professor and AI researcher Mark Riedl tweeted that googling some of the descriptions, including "burrito animals" and "cartoon daikon", could throw up similar images, suggesting Dall-E had copied rather than created some its work.

He also highlighted some of its failures, such as attempts to combine a snail and a harmonica.

But he added: "To be honest, I am not sure how to blend a snail and a harmonica either."

He later told the BBC that he was impressed by what the system could do.

"The ability to use concepts in new ways is an important part of creativity and intelligence," he said.