Gen-3's architecture uses diffusion transformers instead of the convolutional nets used in earlier models