Long-Horizon Video Generation

Long-horizon video generation studies how to extend generated videos across many scenes or minutes of streaming while preserving subjects, environments, motion, and events. Common approaches include autoregressive diffusion with compressed or retrieved memory, sparse KV-cache policies, training on model-generated rollouts, and geometry- or action-conditioned world models. These methods support multi-shot storytelling, interactive world simulation, and camera-controlled exploration, with benchmarks such as VBench-Long testing consistency, visual quality, and control over extended rollouts. A central tension is balancing persistent memory and natural motion with bounded compute, especially when models trained on short clips must remain stable far beyond their training horizon.