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

2026-07-13

AI Summary

Autoregressive diffusion models often suffer from drift, structural collapse, and quality degradation in long-horizon video generation due to the accumulation of prediction errors. This work proposes a two-stage cycle-consistency framework that enforces temporal reversibility constraints by incorporating a causal regularizer during training and leveraging a frozen backward-prediction model for gradient-guided latent correction during inference. This approach compels the generated trajectories to remain tightly aligned with the true video manifold, effectively mitigating error accumulation. To the best of our knowledge, this is the first method to achieve strict control over forward-generation drift, significantly improving overall quality and temporal consistency for 60-second videos on the VBench benchmark and establishing a new state-of-the-art performance.

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 iteratively refines the generated latent representations, actively suppressing accumulated errors before they are committed to the historical context. Extensive experiments on the VBench benchmark demonstrate that Cycle-World'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.

Problem

• error accumulation
• long-horizon video generation
• temporal consistency
• generative drift
• video world models

Innovation

• cycle-consistency
• reverse-prediction
• error accumulation mitigation
• temporal reversibility
• long-horizon video generation