Pack and Force Your Memory: Long-form and Consistent Video Generation

2025-10-02

AI Summary

Long-video generation faces two major challenges: difficulty in modeling long-range temporal dependencies and severe error accumulation during autoregressive decoding. To address these, we propose MemoryPack and Direct Forcing—two novel, linear-complexity mechanisms. MemoryPack enables dynamic, text-image joint-guided memory modeling via learnable contextual retrieval, significantly improving inter-frame consistency. Direct Forcing aligns training and inference through a single-step approximation strategy, effectively suppressing error propagation. Both components preserve scalability while enhancing temporal coherence and structural stability. Extensive experiments demonstrate substantial improvements in long-video generation quality, achieving state-of-the-art performance across multiple benchmarks—including COYO-Video, WebVid-10M, and UCF101—without increasing computational overhead. Our approach advances autoregressive video generation toward practical deployment.

Technology Category

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

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Long-form video generation presents a dual challenge: models must capture long-range dependencies while preventing the error accumulation inherent in autoregressive decoding. To address these challenges, we make two contributions. First, for dynamic context modeling, we propose MemoryPack, a learnable context-retrieval mechanism that leverages both textual and image information as global guidance to jointly model short- and long-term dependencies, achieving minute-level temporal consistency. This design scales gracefully with video length, preserves computational efficiency, and maintains linear complexity. Second, to mitigate error accumulation, we introduce Direct Forcing, an efficient single-step approximating strategy that improves training-inference alignment and thereby curtails error propagation during inference. Together, MemoryPack and Direct Forcing substantially enhance the context consistency and reliability of long-form video generation, advancing the practical usability of autoregressive video models.

Problem

• Addressing long-range dependency capture in video generation
• Mitigating error accumulation in autoregressive decoding processes
• Enhancing temporal consistency and reliability for long videos

Innovation

• MemoryPack models dependencies with textual and image guidance
• Direct Forcing improves training-inference alignment efficiently
• Combined techniques enhance consistency in long-form video generation