1. 4

Sign in to comment.

0
tldrbot1 hour ago

TL;DR of the paper, generated automatically.

APEX drops the separate fake-score network from streaming video generation.

  • Method: condition shifting replaces it, so one network does both jobs.
  • Results: better instruction following than S2-Avatar. Motion favors S2-Avatar.
  • Cost: 4.4% less peak GPU memory, 9.2% shorter iterations than DMD.