Source-linked AI summary
StableWorld: Towards Stable and Consistent Long Interactive Video Generation
Ying Yang, Zhengyao Lv, Yujia Zeng, Tianlin Pan, Haofan Wang, Yueming Lyu, Binxin Yang, Hubery Yin, Chen Li, Jing Lyu, Ziwei Liu, Chenyang Si
TL;DR
Interactive video generation models can suffer progressive scene collapse because small frame-to-frame drifts accumulate within scenes and propagate over long horizons. StableWorld dynamically evicts degraded historical frames while retaining geometrically consistent references, and experiments across multiple world models report improved stability, consistency, and generalization with only slight inference-time overhead.
Problem
Current interactive video generation models suffer instability and temporal degradation during long-horizon generation, including accumulated drift and progressive scene collapse.
Method
StableWorld uses a Dynamic Frame Eviction Mechanism to maintain clean historical frame representations by filtering degraded frames and retaining geometrically consistent references.
Results
Extensive experiments across Matrix-Game 2.0, Open-Oasis, and Hunyuan-GameCraft 1.0 show improved stability, long-term consistency, and generalization across diverse interactive scenarios.
Takeaways & Limitations
StableWorld is evaluated across multiple interactive video models and reduces error accumulation while preserving motion consistency.
Takeaways & Limitations
ORB-based geometric similarity introduces approximately a 1.01×–1.02× inference slowdown under the default setting.
Abstract
from arXiv · showhide
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.
1 Introduction
Current interactive world models can progressively collapse over long horizons because small within-scene drifts accumulate over time. StableWorld addresses this by dynamically evicting degraded frames while retaining geometrically consistent references, improving stability across models and scenarios.
- Current world models commonly exhibit progressive scene collapse during long generation, especially in static or slowly changing environments.
- Small discrepancies between adjacent frames accumulate into drift from the initial clean state, eventually producing visible inconsistencies and scene collapse.
- Larger history windows mitigate degradation by retaining temporally cleaner early frames that reduce frequency-amplitude fluctuations relative to the first frame.
- StableWorld dynamically maintains clean historical representations by retaining recent frames for motion continuity and selecting less-degraded earlier reference frames.
- StableWorld reduces cumulative drift while preserving motion continuity across Matrix-Game 2.0, Open-Oasis, and Hunyuan-GameCraft 1.0 under diverse interactive conditions.
2 Related work
Related work develops video generation, interactive world simulation, and long-video generation, while addressing controllability, temporal coherence, and extended temporal horizons.
- Video generation methods learn spatio-temporal knowledge and produce visually coherent videos using diffusion and autoregressive approaches.
- Interactive world models predict environment states autoregressively from current states and actions, extending video generation toward world simulation.
- Long-video generation remains challenging because of computational costs and cumulative drift, motivating chunked generation and other long-context approaches.
3 Methodology
Interactive video generation combines diffusion denoising with autoregressive conditioning on selected historical frames and current actions. StableWorld addresses long-horizon drift by retaining geometrically consistent references and evicting degraded frames while preserving scene-transition flexibility.
- 3.1 Preliminary: Interactive generation produces each frame autoregressively from selected historical frames and the current action, while diffusion performs intra-frame denoising.The historical references are maintained in a memory buffer for sequential, real-time interaction.
- 3.2 The Reason for Scene Collapse: Small discrepancies between adjacent frames accumulate over time, causing drift from the initial clean state and eventual scene collapse.Drift from earlier views can also propagate into subsequently generated viewpoints.
- 3.2 The Reason for Scene Collapse: Larger history windows mitigate degradation mainly by preserving several clean early frames as reliable references for later generation.This stabilization is associated with reduced frequency-amplitude differences relative to the first frame.
- 3.2 The Reason for Scene Collapse: Rigidly retaining early frames can hinder drastic scene transitions, motivating a dynamic eviction strategy that balances drift reduction with scene-switching flexibility.Under identical action instructions, excessive early-frame retention fails to switch to a new scene.
- 3.3 Dynamic Frame Eviction via ORB-based Geometric Similarity: StableWorld compares ORB features between a reference frame and middle frames, filters matches with a ratio test, and verifies geometric consistency using Homography and Fundamental matrix RANSAC models.The resulting inlier ratios determine the similarity score used by the eviction procedure.
- 3.3 Dynamic Frame Eviction via ORB-based Geometric Similarity: When similarity remains above a threshold, the procedure checks farther frames; it evicts the farthest frame if all remain consistent, otherwise the frame before the first failure.Homography and Fundamental matrix inlier sets provide the geometric verification basis.
4 Experiments
Experiments evaluate StableWorld across three interactive video models and diverse settings, showing improved visual quality, stability, temporal consistency, and generalization with minimal overhead. Ablations identify suitable window-size, similarity-metric, and threshold choices.
- Evaluation Setup: Evaluation uses official datasets across Matrix-Game 2.0, Open-Oasis, and Hunyuan-GameCraft 1.0 under varied scenes and motions.The Matrix-Game evaluation covers 16 scenes with small and large motion actions.
- Qualitative Results: StableWorld outputs maintain higher scene stability, smoother motion, and more consistent temporal dynamics than baselines across all three models.The qualitative comparison includes Matrix-Game 2.0, Open-Oasis, and Hunyuan-GameCraft 1.0.
- User Study: StableWorld receives the majority of user-study votes, supporting gains in visual quality, temporal consistency, and motion smoothness.
- Window Size: Window sizes of 18 or 36 introduce residual artifacts from previous scenes, with the issue more pronounced at size 36.Excessive historical frames interfere with scene transitions.
- Similarity Ablations: ORB-based similarity achieves the strongest overall performance, while the default threshold 0.75 offers the best trade-off between visual quality and temporal stability.SSIM is overly sensitive to geometric perspective changes, whereas cosine similarity can miss scene changes and retain old-scene frames.
5 Conclusion and Future Work
The paper attributes scene collapse to accumulated inter-frame drift within the same scene and proposes StableWorld to reduce this error while preserving motion consistency. Experiments across three interactive video models report improved long-horizon visual quality, and future work targets training-time integration and longer stable generation.
- Conclusion: Scene collapse originates from inter-frame drift between adjacent frames that accumulates over time and creates large deviations from the original scene.
- Conclusion: StableWorld uses dynamic frame eviction to reduce error accumulation while preserving motion consistency across Matrix-Game 2.0, Open-Oasis, and Hunyuan-GameCraft 1.0.
- Future Work: Future work will integrate StableWorld into training and investigate discarding more drifted frames to extend interactive video length and stability.
A Algorithmic Implementation
The algorithm computes ORB-based geometric similarity between a reference frame and earlier frames, then evicts frames judged insufficiently similar before inserting the newly generated frame. Similarity is based on geometrically consistent matches verified by Homography and Fundamental matrix models.
- Inputs and Outputs: Algorithm 1 takes a latent-space sliding window, pixel frames, earlier frames, and a similarity threshold, then outputs an updated window.
- Similarity Computation: ORB features are extracted from the reference frame P0 and current frame Pk before descriptor matching and geometric verification.
- Similarity Computation: RANSAC estimates Homography H and Fundamental matrix F, and similarity is defined as s(P0, Pk) = max(rH, rF).
- Window Update: The procedure evicts the farthest latent frame and inserts the newly generated frame to maintain window size N.
- Thresholding: Frames with fewer than five valid correspondences receive similarity score 0, while the eviction threshold is set to 0.75.
B Additional Ablation
Additional ablations show that evicting recent frames harms short-term motion, while retaining too many historical frames limits adaptation to scene changes. Selectively removing earlier degraded frames better supports smooth transitions.
- Ablation Setup: The ablation compares evicting both earlier and recent frames with evicting only earlier frames under identical action conditions.
- Results: Evicting recent frames causes inconsistent motion and unstable short-term dynamics that hinder subsequent frame generation.
- Interpretation: Evicting recent frames and retaining too many frames both reduce adaptability to scene changes and limit smooth transitions to new environments.
- Conclusion: Selective removal of earlier degraded frames provides the more suitable strategy for preserving motion continuity and adapting to new scenes.
C More Qualitative Results
StableWorld strengthens stability during long-horizon generation across diverse outputs and motion settings, preventing cumulative errors while maintaining coherent motion.
- StableWorld effectively prevents cumulative errors across diverse long-horizon generation settings.The evaluation includes Matrix-Game 2.0 outputs and extremely long sequences under both small- and large-motion conditions.
- StableWorld produces notably more stable and temporally consistent interactive video sequences while maintaining coherent motion.
D Qualitative Results in Autoregressive Video Generation
Self-Forcing also exhibits scene-dependent error accumulation, especially in slowly changing scenes, while integrating StableWorld alleviates this degradation and improves sequence stability.
- Self-Forcing serves as the baseline for evaluating StableWorld in long-horizon autoregressive video generation.
- In Self-Forcing, small inter-frame drift accumulates within slowly changing scenes and eventually causes severe degradation.Catastrophic collapse is less common when scenes change frequently, but remains clearly visible in low-motion settings.
- StableWorld significantly alleviates error accumulation in Self-Forcing, resulting in more stable, consistent video generation.The integration uses geometric similarity to decide when earliest reference frames should be evicted.
E Limitation
The method incurs a slight inference-time slowdown, while qualitative results report stable generation across diverse environments, motion conditions, and long horizons.
- ORB-based geometric similarity introduces an approximately 1.01×–1.02× slowdown compared with the vanilla model under the default setting.The authors leave dynamically adjusting the sliding-window length for future exploration.
- Qualitative comparisons report more stable results and better visual quality than Matrix-Game 2.0, Open-Oasis, Hunyuan-GameCraft 1.0, and self-forcing.
- StableWorld maintains scene stability over thousands of frames under small-motion scenarios without drift or degradation.
- StableWorld preserves temporal consistency and avoids cumulative drift despite significant viewpoint and motion changes.