This reminds me when people were doing crazy stuff to improve the first Stable Diffusion model by swapping layers, interpolating weights, documenting which layer was most responsible for the quality of the hands etc. At the end the final models had dozens of different ancestors.