Cycle-World Mitigating Error Accumulation in Long-term Video World Models via Reverse-Prediction Cycle Consistency

Jul 16, 2026

Upload by Ophelia

Autoregressive diffusion models can generate high-quality videos, but their sequential rollout causes unavoidable error accumulation. In long-horizon synthesis, small prediction deviations compound, triggering generative drift, structural collapse, and strong visual degradation. Cycle-World is proposed to enable stable, temporally consistent long-video generation by enforcing strict temporal reversibility during both training and inference. The method integrates reverse-prediction into training for causal embedding and uses the frozen reverse model for cycle-guided runtime correction with gradient-based refinement, validated on VBench for improved 60-second coherence.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages34

Reading Time1 h 26 min

Introduction

Error accumulation in autoregressive video generation

Structural hallucinations and physical inconsistency

Cycle-World framework and temporal reversibility

arXiv :2607 . 11836v1 [ cs .CV] 13 Jul 2026 Cycle-World: Mitigating Error Accumulation in Long-term Video World Models via Reverse-Prediction Cycle Consistency Zihan Su 1 ⋆, Teng Hu 1 ⋆, Jiangning Zhang2, Ruiyan Wang 1, Ran Yi 1†, Lizhuang Ma 1†, and Dacheng Tao3 1 School of Computer Science, Shanghai Jiao Tong University, Shanghai, China 2 Institute of Cyber-Systems and Control, Zhejiang University, Hangzhou, China 3 Nanyang Technological University, Singapore [https://szhcz.github.io/projects/Cycle-World/](https://szhcz.github.io/projects/Cycle-World/) Abstract. Autoregressive diffusion models have enabled high-quality video generation, yet their sequential nature inherently suffers from error accumulation. In long-horizon video synthesis, minor prediction deviations compound over time, inevitably leading to unconstrained generative drift, structural collapse, and severe visual degradation. To address this, we propose Cycle-World, a novel framework designed for stable and temporally consistent long-video generation. Our approach tackles error drift by enforcing strict temporal reversibility across both the training and inference phases. Theoretically, we demonstrate that forward generative drift can be strictly bottlenecked by a cycle-consistency objective. During training, we integrate an efficient reverse-prediction model to implicitly embed causal constraints into the forward generator, compelling it to produce reversible sequences that tightly adhere to the natural video manifold. At inference time, we repurpose this frozen reverse model as a runtime corrector. Through gradient-based cycle guidance, it iterativelyrefines the generated latent representations, actively suppressing accumulated errors before they are committed to the historical context. Extensive experiments on the VBench benchmark demonstrate that CycleWorld’s dual-phase synergy significantly mitigates error drift, achieving state-of-the-art overall generation quality and long-horizon temporal consistency in 60-second synthesis. Keywords: Video generation · Cycle consistency · Error accumulation 1 Introduction The field of video generation has witnessed unprecedented advancements recently, driven by powerful models such as Sora [36, 37], Seedance [12, 41], and ⋆ Equal contribution, † Corresponding authors. 2 Z. Su et al. 5s A boat sailing leisurely along the Seine River with the Eiffel Tower in background, Van Gogh style A fat rabbit wearing a purple robe walking through a fantasy landscape 30s an eye-level shot of a consistent orange-white British Shorthair in a bright warm-lit living room Cyberpunk cinematic, female agent with glowing cyber eye, rainy night neon 60s+ a tracking shot of a hiker with a red backpack on a a tracking shot of a Golden Retriever running mountain trail at sunrise Fig. 1: High-fidelity, long-horizon video generation with Cycle-World. By effectively bottlenecking generative drift via temporal reversibility, our framework strictly suppresses structural hallucinations inherent in forward-only models. CycleWorld maintains state-of-the-art visual quality, strict physical conservation, and temporally consistent object states over extended generation horizons. Kling [44], alongside pioneering open-source efforts like Wan [45], HunyuanVideo [29] . While these models primarily rely on bidirectional or full-sequence architectures that achieve remarkable visual quality, their non-causal nature fundamentally limits their flexibility for open-ended, sequential, and interactive generation. As the community pivots towards the paradigm of World Models [17, 25 , 34 , 42], there is a critical consensus that real-time interactivity, continuous generation, and causal reasoning are indispensable. Consequently, the field is experiencing a paradigm shift towards causal, autoregressive generation models [16, 23 , 57 , 61] . However, this shift towards causal autoregressive models introduces severe bottlen

FAQs

Why do autoregressive video world models suffer from long-horizon failure?

Because sequential prediction compounds small deviations over time, leading to generative drift and severe visual degradation during long-horizon rollout.

What problem does Cycle-World specifically target beyond generic noise drift?

Cycle-World targets structural hallucinations—irreversible violations of physical consistency such as entities appearing, clipping through objects, or objects vanishing without trace.

How does Cycle-World enforce temporal consistency during training and inference?

It integrates an efficient reverse-prediction model during training to embed causal constraints and uses the frozen reverse model at inference as a runtime corrector via gradient-based cycle guidance.

Jul 16, 2026

Upload by Ophelia