I think you’re spot on. In my experience people confuse a models ability to solve some benchmark as a sign of its usefulness. Token throughput is often just as important from my personal usage. I am excited for more diffusion models to see how progress happens there.