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

Noosaga Placements

• Computer Visionsubfield95%
  The paper focuses on video generation, a core computer vision task, and is evaluated on VBench, a standard computer vision benchmark. The primary arXiv category is cs.CV.
  arXiv:2607.11836v1 [cs.CV]Extensive experiments on the VBench benchmark demonstrate that Cycle-World's dual-phase synergy significantly mitigates error drift
• Diffusion Modelsframework95%
  The paper explicitly builds upon autoregressive diffusion models, which are a type of diffusion model. The forward generator is a diffusion model, and the method is designed to improve its long-term generation.
  Autoregressive diffusion models have enabled high-quality video generationthe forward generator Gθ and reverse model Rϕ are strictly frozenactual synthesis in diffusion models involves an iterative denoising process
• Deep Learningsubfield90%
  The method is built on autoregressive diffusion models, a deep learning technique, and uses concepts like latent representations, gradient-based optimization, and distribution matching distillation.
  Autoregressive diffusion models have enabled high-quality video generationWe propose Cycle-World, a novel framework designed for stable and temporally consistent long-video generationDuring training, we integrate an efficient reverse-prediction model to implicitly embed causal constraints into the forward generator
• Vision Transformers and Foundation Modelsframework70%
  The paper situates its work within the context of large-scale video generation models like Sora, Wan, and HunyuanVideo, which are often considered foundation models. The method is designed to be model-agnostic and applicable to pre-trained models.
  driven by powerful models such as Sora, Seedance, and Kling, alongside pioneering open-source efforts like Wan, HunyuanVideoIn the era of large-scale foundation models, such retraining is often computationally prohibitiveCGI can be seamlessly plugged into any off-the-shelf autoregressive video generator

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.

Paper Context