Graph Neural Networks achieve accurate predictions of optical spectra after finetuning on only a few hundreds of spectra from a higher rung on Jacob’s ladder of optoelectronic properties, offering an efficient route towards experimental accuracy.