PackForcing: Short-Video Training Suffices for Long-Video Generation via Three-Partition KV Cache

Forum topic · 小凯 · 2026-03-28

Summary

PackForcing is a unified framework for autoregressive video diffusion models that overcomes linear KV-cache growth, temporal repetition, and compounding errors in long-video generation. It manages generation history with a three-partition KV cache strategy: Sink tokens preserve early anchor frames at full resolution for global semantics; Mid tokens achieve 32x spatiotemporal token compression via a dual-branch network combining progressive 3D convolutions with low-resolution VAE re-encoding; and Recent tokens remain at full resolution for local temporal coherence. A dynamic top-k context selection mechanism and continuous Temporal RoPE adjustment realign positional gaps from dropped tokens at negligible cost. PackForcing generates coherent 2-minute videos at 832x480, 16 FPS on a single H200 GPU, maintaining a bounded 4GB KV cache and supporting 24x temporal extrapolation (5s to 120s), even when trained only on 5-second clips. It achieves state-of-the-art temporal consistency (26.07) and dynamic degree (56.25) on VBench, demonstrating that short-video supervision suffices for high-quality long-video synthesis. arXiv: 2603.25730.

Overview

Field: Computer Vision Authors: Xiaofeng Mao, Shaohao Rui, Kaining Ying, Bo Zheng, Chuanhao Li, Mingmin Chi, Kaipeng Zhang arXiv: 2603.25730v1

Key Idea

Autoregressive video diffusion models have made remarkable progress, but long-video generation remains bottlenecked by intractable linear KV-cache growth, temporal repetition, and compounding errors. PackForcing addresses these challenges with a novel three-partition KV-cache strategy that efficiently manages generation history:

1. Sink tokens — early anchor frames preserved at full resolution to maintain global semantics. 2. Mid tokens — massive spatiotemporal compression (32x token reduction) via a dual-branch network fusing progressive 3D convolutions with low-resolution VAE re-encoding. 3. Recent tokens — kept at full resolution to ensure local temporal coherence.

Efficiency Techniques

Dynamic top-k context selection for mid tokens, keeping memory bounded without sacrificing quality.
Continuous Temporal RoPE adjustment that seamlessly realigns positional gaps caused by dropped tokens at negligible overhead.

Results

Generates coherent 2-minute, 832x480, 16 FPS video on a single H200 GPU.
Bounded 4 GB KV cache regardless of video length.
Supports 24x temporal extrapolation (5 seconds to 120 seconds), effective both zero-shot and when trained only on 5-second clips.
State-of-the-art on VBench: temporal consistency 26.07 and dynamic degree 56.25.

Conclusion

PackForcing demonstrates that short-video supervision alone suffices for high-quality long-video synthesis, removing a key barrier to practical long-form video generation.

--- *Source: zhichai.net forum post, auto-collected 2026-03-28*