FreqForcing: Autoregressive Long Video Generation via Spectral Self-Anchoring

Researchers from Shanghai Jiao Tong University and Tencent HY Team propose FreqForcing, a training-free method for stabilizing autoregressive long-video diffusion. Unlike bidirectional systems such as Sora, HunyuanVideo, and Wan, autoregressive models repeatedly condition on their own imperfect outputs, causing color drift, motion stagnation, and collapse. The paper identifies this degradation as spectral energy drift concentrated in DC and low-frequency bands. Its Spectral Self-Anchoring mechanism uses two inference branches: local attention preserves high-frequency motion and detail, while anchor attention retrieves low-frequency structure from high-quality early frames; a 3D FFT-based low-pass fusion combines them, with temporal RoPE alignment maintaining valid positional ranges. Applied to the Wan2.1-T2V-1.3B base model, FreqForcing extends Self-Forcing trained on 5-second clips to stable 120-second generation, a 24-fold extrapolation, using the first 128 MovieGen prompts for evaluation. On VBench-Long at 60 seconds, it reaches 59.58 Dynamic Degree and 20.94 Overall Consistency, while at 120 seconds it records 58.97 and 20.98 respectively, outperforming training-free baselines and remaining competitive with training-based methods. The approach increases inference latency by only 16.5 percent on NVIDIA RTX A6000 GPUs, and ablations show that both attention sinks and spectral anchoring are necessary to suppress long-horizon drift.