Dual-Signal Adaptive KV-Cache Optimization for Long-Form Video Understanding in Vision-Language Models

TL;DR

Sali-Cache is a proactive memory management framework for vision-language models that uses dual-signal filters to optimize KV-cache usage, enabling efficient long-form video understanding without accuracy loss.

Contribution

It introduces a novel a priori caching strategy combining temporal and spatial filters, reducing memory usage while maintaining model accuracy in long video processing.

Findings

Achieves 2.20x memory compression ratio.

Maintains 100% accuracy on BLEU, ROUGE-L, and Exact Match metrics.

Enables long-form video processing on consumer hardware.

Abstract

Vision-Language Models (VLMs) face a critical memory bottleneck when processing long-form video content due to the linear growth of the Key-Value (KV) cache with sequence length. Existing solutions predominantly employ reactive eviction strategies that compute full attention matrices before discarding tokens, resulting in substantial computational waste. We propose Sali-Cache, a novel a priori optimization framework that implements dual-signal adaptive caching through proactive memory management. By integrating a temporal filter based on optical flow analysis for detecting inter-frame redundancy and a spatial filter leveraging saliency detection for identifying visually significant regions, Sali-Cache intelligently manages memory allocation before entering computationally expensive attention operations. Experimental evaluation on the LLaVA 1.6 architecture demonstrates that our method achieves a 2.20x compression ratio in effective memory usage while maintaining 100% accuracy across BLEU, ROUGE-L, and Exact Match metrics. Furthermore, under identical memory budget constraints, Sali-Cache preserves context-rich features over extended temporal durations without degrading model performance, enabling efficient processing of long-form video content on consumer-grade hardware.