Google introduced the Imagen neural network. She creates the craziest photos from any text

Google announced Imagen neural network for creating photos from text queries.

So far, the company does not allow testing the capabilities of the neural network on its own, but has published several unusual options. Each example has a corresponding query.

A few examples:

Imagen is based on the diffusion method – initially, a text request creates a schematic drawing with a resolution of 32×32 pixels, which is subsequently increased to 256×256 and up to 1024×1024 pixels. At the same time, the neural network adds new elements to the drawing, thanks to which the picture is of high quality.

The developers compared Imagen with a similar DALL-E 2 neural network from OpenAI. Testers when comparing images from Imagen and DALL-E 2 more often chose Imagen for more accurate and reliable work. [The Verge]