LIVE: Long-horizon Interactive Video World Modeling

Autoregressive video world models predict future visual observations conditioned on actions. While effective over short horizons, these models often struggle with long-horizon generation, as small prediction errors accumulate over time. Prior methods alleviate this by introducing pre-trained teacher models and sequence-level distribution matching, which incur additional computational cost and fail to prevent error propagation beyond the training horizon. In this work, we propose LIVE, a Long-horizon Interactive Video world modEl that enforces bounded error accumulation via a novel cycle-consistency objective, thereby eliminating the need for teacher-based distillation. Specifically, LIVE first performs a forward rollout from ground-truth frames and then applies a reverse generation process to reconstruct the initial state. The diffusion loss is subsequently computed on the reconstructed terminal state, providing an explicit constraint on long-horizon error propagation. Moreover, we provide a unified view that encompasses different approaches and introduce progressive training curriculum to stabilize training. Experiments demonstrate that LIVE achieves state-of-the-art performance on long-horizon benchmarks, generating stable, high-quality videos far beyond training rollout lengths.

Modern AI systems can learn to imagine how a scene will change after a user or agent takes actions, such as moving a camera through a room or navigating in a game. This ability is important for building interactive simulators for robots, games, and virtual environments. However, current systems often work well only for short videos: small mistakes made at each step gradually build up, causing the generated video to become blurry, unrealistic, or inconsistent over time. This paper introduces LIVE, a new training method for long-horizon interactive video generation. The key idea is to teach the model not only to predict future frames, but also to recover the original starting scene from its own generated video. This encourages the model to keep its mistakes under control instead of letting them grow over time. Experiments on real-world indoor videos and interactive game environments show that LIVE produces more stable and higher-quality videos over much longer sequences than previous methods, without requiring an extra teacher model.