The model inferred meaning much the same way it infers meaning from text. Short phrases can generate intricate images accurate to author intent using stable diffusion.

The models themselves in those studies leveraged stable diffusion as the mechanism of image generation, but instead of text prompts, they use fMRI data training.