Results from my experiment – with fine-tuning, diffusion models can beat pre-trained models almost 9 times larger, and autoregressive fine-tuned models 1.5 times larger