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

2026-03-26

AI Summary

This work addresses the limitations of autoregressive video diffusion models in long-form video generation—namely, linearly growing KV cache memory, temporal redundancy, and error accumulation. The authors propose PackForcing, a framework that manages generation history via a three-partition KV cache strategy (Sink, Mid, and Recent tokens), integrated with a dual-branch spatiotemporal compression network (combining 3D convolutions and low-resolution VAE re-encoding), dynamic top-k context selection, and a continuous Temporal RoPE positional alignment mechanism. Using only 4 GB of KV cache on a single H200 GPU, PackForcing generates 2-minute videos at 832×480 resolution and 16 FPS, achieving a 24× temporal extrapolation (from 5 to 120 seconds). This is the first demonstration that short training clips can support high-quality long-video synthesis, setting new state-of-the-art results on VBench for temporal consistency (26.07) and motion dynamism (56.25).

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Natural Language Processing: Generation

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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, we present PackForcing, a unified framework that efficiently manages the generation history through a novel three-partition KV-cache strategy. Specifically, we categorize the historical context into three distinct types: (1) Sink tokens, which preserve early anchor frames at full resolution to maintain global semantics; (2) Mid tokens, which achieve a massive spatiotemporal compression (32x token reduction) via a dual-branch network fusing progressive 3D convolutions with low-resolution VAE re-encoding; and (3) Recent tokens, kept at full resolution to ensure local temporal coherence. To strictly bound the memory footprint without sacrificing quality, we introduce a dynamic top-$k$ context selection mechanism for the mid tokens, coupled with a continuous Temporal RoPE Adjustment that seamlessly re-aligns position gaps caused by dropped tokens with negligible overhead. Empowered by this principled hierarchical context compression, PackForcing can generate coherent 2-minute, 832x480 videos at 16 FPS on a single H200 GPU. It achieves a bounded KV cache of just 4 GB and enables a remarkable 24x temporal extrapolation (5s to 120s), operating effectively either zero-shot or trained on merely 5-second clips. Extensive results on VBench demonstrate state-of-the-art temporal consistency (26.07) and dynamic degree (56.25), proving that short-video supervision is sufficient for high-quality, long-video synthesis. https://github.com/ShandaAI/PackForcing

Problem

• long video generation
• KV-cache growth
• temporal repetition
• compounding errors
• autoregressive video diffusion

Innovation

• PackForcing
• KV-cache compression
• long video generation
• temporal extrapolation
• hierarchical context management