Google Researchers Unveil Imagen, a Highly Advanced Text-to-Image AI Model
A new AI system called Imagen combines large language models with diffusion techniques to generate highly realistic images from text descriptions, outperforming competitors in human evaluations.
Researchers introduce Imagen, a new text-to-image diffusion model that achieves an unprecedented level of photorealism and deep language understanding. The system combines the text-processing power of large transformer language models with the high-fidelity image generation capabilities of diffusion models to produce highly accurate visuals from written prompts.
The team discovers that scaling up the language model component yields significantly better image-text alignment and sample fidelity than simply increasing the size of the image diffusion model. Thanks to this approach, Imagen achieves a new state-of-the-art FID score of 7.27 on the COCO dataset without ever training on COCO data, with human raters finding the generated images to be on par with real COCO images in terms of text alignment.
To rigorously evaluate this new technology, the researchers also introduce DrawBench, a comprehensive benchmark designed to test text-to-image models on complex prompts. In side-by-side comparisons against recent methods like VQ-GAN+CLIP, Latent Diffusion Models, and DALL-E 2, human raters consistently prefer Imagen for both overall sample quality and image-text alignment.