Meta increases the accuracy of automated image creation with Make-A-Scene

If you draw an outline of a sandy beach under a sunny sky with surfboards, label each of these elements with keywords and describe the scene in a sentence, the result is amazing. Mark Zuckerberg, the creator of Facebook, presented a new Meta AI development project on Thursday, July 14, in a demonstration video. Make-A-Scene is a generator of artistic creations powered by artificial intelligence such as Craiyon, Dall-E or Imagen.“It will eventually be used to tell digital stories, build immersive virtual worlds in the metaverse, or even create artwork in NFT.” states a press release from Meta AI.

First, the sentence is analyzed by a neural network that examines each word and links them together to develop context. After understanding the information, the AI ​​synthesizes the image using a set of GANs (generative adversarial networks). This helps to create photorealistic images, even if the depicted scene is filled with incongruous elements. Adding a sketch drawn by the creator allows you to control the composition of the image. According to several testers, the sketch-to-text combination is 99.54% closest to the expected result, compared to 66% using text alone. Upon arrival, the photo has a resolution of 2048 x 2048, or about 4.2 million megapixels.

This is a boon for machine learning to better understand how people perceive the world. “Using artificial intelligence to enhance human creativity is a powerful use of technology”, thinks Meta AI. The company also works with generative artists who use algorithms to design their creative works, such as Sofia Crespo or Scott Eaton.

For transparency, the company used publicly available data to train Make-A-Scene so the AI ​​community could analyze and understand how the tool works.

So far, there is no information about the availability of the tool to the general public.