Reproduce “DisCo: World Models with Discrete Camera Motion Control” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hongrui Huang, Junke Wang, Quanhao Li, Yu-Gang Jiang, Zuxuan Wu

DisCo converts camera trajectories into 27 discrete combinations of translation and yaw primitives, using them to condition a Wan 2.2 video world model and a four-step autoregressive student. The resulting system improves compositional camera control over continuous-action baselines, reaching an Action Score of 0.6332, an FVD of 220.51 on dynamic scenes, and stable long-horizon generation.

Key insights

  • Separating action features improves controllability: DisCo reduces average action-feature similarity to 0.7029, compared with 0.8826 for CameraCtrl and 0.9913 for Hunyuan-GameCraft.
  • Greater geometric detail is not always beneficial; replacing discrete embeddings with continuous Plücker features lowers Action Score from 0.6332 to 0.5563, while expanding the action vocabulary to 39 or 48 bins also hurts performance.
  • Balanced training matters: dynamic-only data gives the best FVD but weak action following, whereas combining dynamic and synthetic motion data produces the strongest overall trade-off.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.