Stream4D: 4D-Consistency for Streaming Autoregressive Diffusion Video Models
Stream4D targets the failure mode where long video rollouts become geometrically inconsistent or simply stop moving.
The paper says streaming autoregressive diffusion video models can drift over time because they are trained around local frame prediction, not a coherent dynamic world. It argues that static 3D reconstruction rewards can reward frozen scenes, since real object motion looks like reconstruction error. Stream4D swaps in a feed-forward 4D reconstruction reward and adds a motion prior to favor natural scene flow while penalizing jitter and artifacts. The authors report better 4D reconstruction, stronger motion preservation, and higher human-aligned preference across several autoregressive video backbones and generation horizons. HF Daily Papers' note
The paper says streaming autoregressive diffusion video models can drift over time because they are trained around local frame prediction, not a coherent dynamic world. It argues that static 3D reconstruction rewards can reward frozen scenes, since real object motion looks like reconstruction error. Stream4D swaps in a feed-forward 4D reconstruction reward and adds a motion prior to favor natural scene flow while penalizing jitter and artifacts. The authors report better 4D reconstruction, stronger motion preservation, and higher human-aligned preference across several autoregressive video backbones and generation horizons. HF Daily Papers' note
score 5