DeepFloyd-IF: A Pixel-Based Triple-Cascaded Diffusion Model for Photorealistic Text-to-Image Generation 05-04