FlowNar: Scalable Streaming Narration for Long-Form Videos

2026-05-30

AI Summary

This work addresses the challenge that existing large models face in streaming long-form video narration, where computational and memory costs grow linearly with video length, hindering scalability. To overcome this, the authors propose FlowNar, a framework featuring a dynamic context management mechanism that prunes redundant visual information while preserving critical temporal dependencies through a novel Cross Linear Attentive Memory (CLAM) module. This design ensures bounded memory consumption and constant computational complexity regardless of input duration. The study also introduces a self-conditioned evaluation protocol and associated metrics that better reflect real-world deployment scenarios. Experiments demonstrate that FlowNar significantly outperforms strong baselines on Ego4D, EgoExo4D, and EpicKitchens100, enabling processing of videos up to ten times longer with a threefold increase in throughput.

Abstract

Recent Large Multimodal Models (LMMs), primarily designed for offline settings, are ill-suited for the dynamic requirements of streaming video. While recent online adaptations improve real-time processing, they still face critical scalability challenges, with resource demands typically growing at least linearly with video duration. To overcome this bottleneck, we propose FlowNar, a novel framework for scalable streaming video narration. The core of FlowNar is a dynamic context management strategy for historical visual context removal, combined with our CLAM (Cross Linear Attentive Memory) module for streaming visual history retention, ensuring bounded visual memory usage and computational complexity, crucial for efficient streaming. We also introduce a realistic self-conditioned evaluation protocol and complementary evaluation metrics to assess streaming narration models under deployment-like conditions. Experiments on the Ego4D, EgoExo4D, and EpicKitchens100 datasets demonstrate that FlowNar substantially improves narration quality over strong baselines while being highly efficient, supporting processing of 10$\times$ longer videos and achieving 3$\times$ higher throughput (FPS). The code is available at https://github.com/zeyun-zhong/FlowNar.

Problem

• streaming video narration
• scalability
• long-form videos
• resource efficiency
• Large Multimodal Models

Innovation

• streaming video narration
• dynamic context management
• CLAM
• scalable LMM
• bounded memory