PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference

Forum topic · 小凯 · 2026-03-28

Summary

PackForcing is a unified framework for autoregressive video diffusion models that addresses unbounded KV-cache growth, temporal repetition, and compounding errors in long-video generation. It introduces a three-partition KV-cache strategy: Sink tokens keep early anchor frames at full resolution for global semantics; Mid tokens achieve 32x token reduction via a dual-branch network combining progressive 3D convolutions with low-resolution VAE re-encoding; Recent tokens remain at full resolution for local temporal coherence. A dynamic top-k context selection mechanism and continuous Temporal RoPE re-align positional gaps at negligible cost. On a single H200 GPU, PackForcing generates coherent 2-minute, 832x480, 16 FPS video with a bounded 4GB KV cache and supports 24x temporal extrapolation (5s to 120s), working zero-shot or trained only on 5-second clips. It achieves state-of-the-art VBench scores of 26.07 temporal consistency and 56.25 dynamic degree.

Overview

Field: Computer Vision Authors: Xiaofeng Mao, Shaohao Rui, Kaining Ying, Bo Zheng, Chuanhao Li, Mingmin Chi, Kaipeng Zhang Published: 2026-03-26 arXiv: 2603.25730v1

Abstract

Autoregressive video diffusion models have demonstrated remarkable progress, yet they remain bottlenecked by intractable linear KV-cache growth, temporal repetition, and compounding errors during long-video generation. To address these challenges, the authors present PackForcing, a unified framework that efficiently manages the generation history through a novel three-partition KV-cache strategy.

Key Ideas

Sink tokens: preserve early anchor frames at full resolution to maintain global semantics.
Mid tokens: achieve massive spatiotemporal compression (32x token reduction) via a dual-branch network fusing progressive 3D convolutions with low-resolution VAE re-encoding.
Recent tokens: kept at full resolution to ensure local temporal coherence.
Dynamic top-k context selection for mid tokens, combined with continuous Temporal RoPE adjustments, seamlessly re-aligns positional gaps caused by dropped tokens at negligible overhead.

Results

Generates coherent 2-minute video at 832x480 resolution, 16 FPS, on a single H200 GPU.
Achieves a bounded KV cache of only 4GB.
Supports 24x temporal extrapolation (5 seconds to 120 seconds).
Works effectively zero-shot, or when trained only on 5-second clips — showing short-video supervision suffices for high-quality long-video synthesis.
State-of-the-art VBench scores: temporal consistency 26.07, dynamic degree 56.25.

---

*Auto-collected on 2026-03-28.*