StableWorld: Towards Stable and Consistent Long Interactive Video Generation

TL;DR

The work tackles instability in long interactive video generation, where small inter-frame drift within a persistent scene leads to progressive degradation. It first shows that error accumulation originates from drift inside the same scene and propagates across time. It then introduces StableWorld, a Dynamic Frame Eviction mechanism that maintains a sliding window of frames and latent tokens, preserving geometrically consistent frames while evicting degraded ones; eviction is guided by a geometry-based similarity $s(P_0, P_k)=\max(r_H, r_F)$ computed from ORB features with RANSAC on both the Homography $\mathbf{H}$ and the Fundamental matrix $\mathbf{F}$ and an eviction threshold $\theta=0.75$, preventing drift from stacking. StableWorld is model-agnostic and yields substantial gains in visual quality and temporal consistency across Matrix-Game 2.0, Open-Oasis, and Hunyuan-GameCraft 1.0 with only ~1.0–1.02× overhead.

Abstract

In this paper, we explore the overlooked challenge of stability and temporal consistency in interactive video generation, which synthesizes dynamic and controllable video worlds through interactive behaviors such as camera movements and text prompts. Despite remarkable progress in world modeling, current methods still suffer from severe instability and temporal degradation, often leading to spatial drift and scene collapse during long-horizon interactions. To better understand this issue, we initially investigate the underlying causes of instability and identify that the major source of error accumulation originates from the same scene, where generated frames gradually deviate from the initial clean state and propagate errors to subsequent frames. Building upon this observation, we propose a simple yet effective method, \textbf{StableWorld}, a Dynamic Frame Eviction Mechanism. By continuously filtering out degraded frames while retaining geometrically consistent ones, StableWorld effectively prevents cumulative drift at its source, leading to more stable and temporal consistency of interactive generation. Promising results on multiple interactive video models, \eg, Matrix-Game, Open-Oasis, and Hunyuan-GameCraft, demonstrate that StableWorld is model-agnostic and can be applied to different interactive video generation frameworks to substantially improve stability, temporal consistency, and generalization across diverse interactive scenarios.

Figures (21)

• Figure 1: StableWorld: producing stable and visually consistent interactive videos across diverse environments such as natural landscapes and game worlds, while preserving continuous motion control and preventing long-term scene drift.
• Figure 2: Visualization of progressive scene collapse over time across different world-simulation models.
• Figure 3: Accumulated frame-to-frame drift in Matrix-Game 2.0 and Open-Oasis.
• Figure 4: Frequency amplitude difference between the anchor frame and different target frames. (a) Default setting: window size = 9. (b) window size = 36. (c) window size = 90. (d) window size = 9 with the first clean 3 frames retained.
• Figure 5: Comparison of sliding-window updates. When the window slides, the vanilla model simply discards the oldest frames, causing errors to accumulate over iterations. In contrast, StableWorld continuously filters out degraded frames while maintaining geometrically consistent ones, resulting in more stable and consistent interactive video generation.