Deep Forcing: Training-Free Long Video Generation with Deep Sink and Participative Compression

Recent advances in autoregressive video diffusion have enabled real-time frame streaming, however, existing methods still suffer from visual error accumulation including visual fidelity and motion degradation over long-horizon. To address these challenges, we introduce Deep Forcing, a training-free extension of autoregressive video diffusion models that stabilizes long video generation through two complementary mechanisms. Deep Sink preserves approximately half of the sliding context window as persistent sink tokens and realigns their temporal RoPE phases to the current timeline, thereby maintaining global context during extended rollouts. Participative Compression performs importance-aware KV cache pruning, retaining only tokens that actively participate in recent attention while removing redundant or degraded history, effectively mitigating error accumulation under out-of-distribution lengths. Together, these components enable over 12× length extrapolation (e.g., 5s-trained → 60s+) without sacrificing inference speed, while improving visual fidelity and motion dynamics compared to prior methods. Our results demonstrate that Deep Forcing can achieve performance comparable to state-of-the-art training-based methods trained specifically for long video generation.

AI video generators have become impressively good at producing short clips, but ask them to keep going beyond a few seconds and the picture falls apart—colors drift toward oversaturation, fine details blur away, and motion grows erratic. The root cause is memory management: today's models follow a simple "first-in, first-out" rule that discards old frames blindly while leaning heavily on the newest ones, which unfortunately carry the most accumulated errors. We developed Deep Forcing, a method that fixes this without retraining the underlying video model—a major advantage, since training such models is enormously costly. Our approach has two parts: Deep Sink preserves a large group of early frames as stable visual anchors that ground the video throughout its run, while Participative Compression keeps only those past moments the model is actively paying attention to, rather than throwing memories away by age alone. With these two simple changes, video AI trained on just 5-second clips can now generate coherent minute-long videos—a 12× extension—matching methods that required expensive specialized training and bringing long-form interactive video generation within reach for anyone with a pretrained model.