1 year, 1 month ago

How AI image-generators work

THE FLURRY of images generated by artificial intelligence feels like the product of a thoroughly modern tool. Today “generative AI ” models put brush to virtual paper: publicly available apps, such as Midjourney and OpenAI’s DALL-E, create images in seconds based on text prompts. The models behind image-generators are trained on enormous datasets: LAION-5B, the largest publicly available one, contains 5.85bn tagged images. A model that has learned which types of pixel arrangement correlate to the word “hippopotamus” should be able to sample from its latent space to create a realistic image of the mammal. Adding more detail to the prompt—for example, “a renaissance-era oil painting of a green hippopotamus, somewhere along the river Nile” —requires the model to source additional layers of visual detail, such as image style, texture, colour and location, and to combine them correctly.

Hindustan Times

Discover Related