MagicWorld: Towards Long-Horizon Stability for Interactive Video World Exploration

TL;DR

MagicWorld introduces a novel autoregressive interactive video model that enhances long-horizon stability and motion realism by incorporating flow-guided constraints and strategic training methods, supported by a new real-world dataset.

Contribution

It proposes innovative techniques for reducing motion drift and error accumulation in long-term interactive video generation, advancing the state-of-the-art in dynamic scene modeling.

Findings

Improves motion realism in complex dynamic environments

Reduces error accumulation over long interactions

Demonstrates superior performance on RealWM120K dataset

Abstract

Recent interactive video world model methods generate scene evolution conditioned on user instructions. Although they achieve impressive results, two key limitations remain. First, they exhibit motion drift in complex environments with multiple interacting subjects, where dynamic subjects fail to follow realistic motion patterns during scene evolution. Second, they suffer from error accumulation in long-horizon interactions, where autoregressive generation gradually drifts from earlier scene states and causes structural and semantic inconsistencies. In this paper, we propose MagicWorld, an interactive video world model built upon an autoregressive framework. To address motion drift, we incorporate a flow-guided motion preservation constraint that mitigates motion degradation in dynamic subjects, encouraging realistic motion patterns and stable interactions during scene evolution. To mitigate error accumulation in long-horizon interactions, we design two complementary strategies, including a history cache retrieval strategy and an enhanced interactive training strategy. The former reinforces historical scene states by retrieving past generations during interaction, while the latter adopts multi-shot aggregated distillation with dual-reward weighting for interactive training, enhancing long-term stability and reducing error accumulation. In addition, we construct RealWM120K, a real-world dataset with diverse city-walk videos and multimodal annotations to support dynamic perception and long-horizon world modeling. Experimental results demonstrate that MagicWorld improves motion realism and alleviates error accumulation during long-horizon interactions.