Google Brain Unveils Imagen AI, Surpassing DALL-E 2 in Image Generation
Google's Brain Team introduces Imagen, a text-to-image AI model that achieves state-of-the-art results on standard benchmarks using a unique pre-training approach. The system generates highly realistic images from text descriptions while outperforming competitors like DALL-E 2.
Researchers from Google's Brain Team announce Imagen, a state-of-the-art text-to-image AI model that generates highly photorealistic images from textual descriptions. Unlike many competing systems, Imagen takes a unique approach by pre-training only on text data instead of relying on massive datasets of paired images and text scraped from the internet.
The system works by using a Transformer language model to convert input text into embedding vectors, which a series of three diffusion models then transform into a high-resolution 1024x1024 pixel image. Imagen achieves a zero-shot FID score of 7.27 on the COCO benchmark, officially outperforming DALL-E 2 and claiming the top spot among similar generative AI models.
Alongside the model, the Google team introduces DrawBench, a new benchmark suite for evaluating text-to-image systems, and an improved architecture called Efficient U-Net. The researchers acknowledge the complex societal impacts of generative AI and emphasize that their primary goal is to advance research, noting they plan to explore a responsible framework for future externalization of the technology.