> During inference, generating sequences ranging from 16 to 4096 tokens incurs a 16× to 4700× increase in FLOPs compared to AR baselines.

I wonder why the increase in FLOPs has such a wide spectrum? Naively, I'd have expected the FLOPs to increase linearly with the number of tokens. OTOH, it sort of makes sense because because diffusion models are not autoregressive, as their name suggests.