GPT-4: Introducing DeepFloyd IF, a cutting-edge open-source text-to-image model that delivers photorealistic images with advanced language understanding. The model consists of a frozen text encoder and three cascaded pixel diffusion modules, generating images at resolutions of 64×64, 256×256, and 1024×1024 pixels. Utilizing a T5 transformer-based text encoder and a UNet architecture, DeepFloyd IF achieves a zero-shot FID score of 6.66 on the COCO dataset, outperforming current state-of-the-art models and showcasing the potential of text-to-image synthesis.
Read more at GitHub…