Slowly and systematically convinced myself that denoising diffusion models are basically as close as we know so far to The Right Way to do generative and restorative models today. Unfortunately, that means I talked myself out of my own GPU.