FreqForcing: Autoregressive Long Video Generation via Spectral Self-Anchoring

Autoregressive video diffusion models are efficient for streaming video generation but suffer from error accumulation over long horizons. This manifests as color drift, motion stagnation, and eventual visual collapse. The authors investigate whether this degradation can be characterized and mitigated by analyzing the frequency-domain properties of the generated video.

The authors perform a frequency-domain analysis of autoregressive generation, identifying that error accumulation corresponds to a pronounced energy drift in the low-frequency bands (representing global layout and color tone). While existing techniques like 'attention sink'—which retains initial tokens in the key-value cache—help, they cannot fully eliminate this drift.

To address this, the authors propose FreqForcing, a training-free framework featuring Spectral Self-Anchoring (SSA). SSA employs a dual-branch attention mechanism: a local branch captures high-frequency dynamic motion, while an anchor branch maintains a cache of high-quality initial frames. These branches are fused in the frequency domain using a Gaussian low-pass filter, which injects stable low-frequency components from the anchor branch into the local attention output, effectively suppressing spectral energy drift without sacrificing temporal detail.

FreqForcing significantly improves long-horizon video stability. By applying SSA only during the early denoising steps—where global structure is established—the authors minimize computational overhead while achieving 24x extrapolation (extending 5-second pretraining to 2-minute generation). Quantitative evaluations on the VBench-Long benchmark show that FreqForcing outperforms existing training-free methods and remains competitive with more computationally expensive training-based approaches in terms of consistency and visual quality.

This work provides a principled, frequency-based explanation for why autoregressive video models fail over time. By offering a training-free solution that corrects spectral drift, FreqForcing enables the generation of longer, more coherent videos using existing pretrained models, which is critical for applications like interactive world models and embodied AI.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.