Researchers have proposed a new method for training autoregressive diffusion video models to generate coherent and dynamic videos. The previous approach used a static critic that penalized object motion as reconstruction error, but this led to unnatural or static motion in the generated videos. Stream4D replaces the static critic with a feed-forward 4D reconstruction reward that explicitly models scene dynamics, allowing for more realistic motion. This method is combined with a motion prior that rewards natural scene-flow magnitude and penalizes jitter and non-rigid artifacts. --- Why it matters: This matters to researchers in AI because it addresses a significant limitation of previous video generation methods, which struggle to produce coherent and dynamic videos. Stream4D's approach can improve the quality and realism of generated videos, making them more suitable for applications such as virtual reality and video editing. Source: https://arxiv.org/abs/2608.19556