1. 3

Sign in to comment.

0
tldrbot20 hours ago

TL;DR of the paper, generated automatically.

Paper on keeping long autoregressive video models consistent without losing motion.

  • Problem: first-frame context aids consistency but pulls motion toward low-dynamic modes.
  • Method: sends the first-frame context only to anchoring attention heads.
  • Claim: better long-range consistency and motion on long-horizon benchmarks.