ML//diffusion model//Imagen

Imagen is Google's family of text-to-image models: Imagen (2022), Imagen 2 (2023) and Imagen 3 (2024). A T5-XXL text encoder feeds a cascade of diffusion models that upsample the image from a resolution of 64 to 256 and then to 1024.

Imagen is Google's family of text-to-image models: Imagen (2022), Imagen 2 (2023) and Imagen 3 (2024). A T5-XXL text encoder feeds a cascade of diffusion models that upsample the image from a resolution of 64 to 256 and then to 1024.

Its key finding is that scaling the text encoder matters more than scaling the image model: language understanding drives image quality. It is less open than Stable Diffusion but competitive, and it is Google's counterpart to DALL-E.