Cycle-World: Mitigating Error Accumulation in Long-Term Video World Models via Temporal Reversibility

Forum topic · 小凯 · 2026-07-15

Summary

Cycle-World is a new framework for stable, long-horizon video generation that addresses the error accumulation problem inherent in autoregressive diffusion models. Minor prediction deviations in sequential generation compound over time, causing generative drift, structural collapse, and visual degradation. Cycle-World enforces strict temporal reversibility in both training and inference. Theoretically, the authors prove that forward generative drift can be strictly bottlenecked by a cycle-consistency objective. During training, an efficient reverse-prediction model implicitly embeds causal constraints into the forward generator, forcing it to produce reversible sequences that stay close to the natural video manifold. At inference, the frozen reverse model is reused as a runtime corrector: through gradient-based cycle guidance, it iteratively refines generated latent representations, suppressing accumulated errors before they enter the historical context. Experiments on VBench show that the two-phase synergy significantly mitigates error drift, achieving state-of-the-art overall quality and long-term temporal consistency in 60-second synthesis. Paper: arXiv 2607.11836 by Zihan Su, Teng Hu, Jiangning Zhang, Ruiyan Wang, and Ran Yi.

Paper Overview

Field: Computer Vision Authors: Zihan Su, Teng Hu, Jiangning Zhang, Ruiyan Wang, Ran Yi arXiv: 2607.11836

Introduction

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.

Key Ideas

Problem: Error drift in long-term autoregressive video generation — small prediction errors accumulate and degrade quality over time.
Cycle-World: A framework enforcing strict temporal reversibility across both training and inference phases.
Theory: Forward generative drift can be strictly bottlenecked by a cycle-consistency objective.
Training: An efficient reverse-prediction model implicitly embeds causal constraints into the forward generator, forcing it to produce reversible sequences that closely follow the natural video manifold.
Inference: The frozen reverse model is reused as a runtime corrector. Through gradient-based cycle guidance, it iteratively refines generated latent representations, actively suppressing accumulated errors before they are committed to the historical context.

Results

Extensive experiments on the VBench benchmark demonstrate that Cycle-World's two-phase synergy significantly mitigates error drift, achieving state-of-the-art overall generation quality and long-term temporal consistency in 60-second video synthesis.