Autoregressive and Diffusion Are Not Mutually Exclusive

burny_tech · x · 2026-07-17

The repost argues that autoregression and diffusion are not an either-or dichotomy; they can coexist, for instance, by combining them through approaches like **diffusion forcing**. The author further suggests that the real dividing line might not be "autoregressive vs. diffusion," but rather **discrete tokens vs. continuous tokens**. The original quote highlights a common industry heuristic: autoregression wins when memory bandwidth is cheaper, whereas diffusion has the upper hand when FLOPS are cheaper.

Original post →

More from Research

Research channel →