LIVE: Long-horizon Interactive Video World Modeling

TL;DR

LIVE introduces a novel cycle-consistency approach for long-horizon video prediction, effectively controlling error propagation without additional teacher models, leading to superior long-term video generation quality.

Contribution

It proposes a cycle-consistency objective for long-horizon video modeling, eliminating the need for teacher models and improving stability and quality in extended video predictions.

Findings

Achieves state-of-the-art long-horizon video prediction performance.

Produces stable, high-quality videos beyond training rollout lengths.

Outperforms prior methods in long-term video generation benchmarks.

Abstract

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 an 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.