A comprehensive description of my thoughts when trying to make a latent diffusion model (LDM) train as fast as it can while only free compute sources