AirCache: Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference

Recent advancements in Large Visual Language Models (LVLMs) have gained significant attention due to their remarkable reasoning capabilities and proficiency in generalization. However, processing a large number of visual tokens and generating long-context outputs impose substantial computational overhead, leading to excessive demands for key-value (KV) cache. To address this critical bottleneck, we propose AirCache, a novel KV cache compression method aimed at accelerating LVLMs inference. This work systematically investigates the correlations between visual and textual tokens within the attention mechanisms of LVLMs. Our empirical analysis reveals considerable redundancy in cached visual tokens, wherein strategically eliminating these tokens preserves model performance while significantly accelerating context generation. Inspired by these findings, we introduce an elite observation window for assessing the importance of visual components in the KV cache, focusing on stable inter-modal relevancy modeling with enhanced multi-perspective consistency. Additionally, we develop an adaptive layer-wise budget allocation strategy that capitalizes on the strength and skewness of token importance distribution, showcasing superior efficiency compared to uniform allocation. Comprehensive evaluations across multiple LVLMs and benchmarks demonstrate that our method achieves comparable performance to the full cache while retaining only 10% of visual KV cache, thereby reducing decoding latency by 29% to 66% across various batch size and prompt length of inputs. Notably, as cache retention rates decrease, our method exhibits increasing performance advantages over existing approaches.

A Joint Framework Towards Class-aware and Class-agnostic Alignment for Few-shot Segmentation

Few-shot segmentation (FSS) aims to segment objects of unseen classes given only a few annotated support images. Most existing methods simply stitch query features with independent support prototypes and segment the query image by feeding the mixed features to a decoder. Although significant improvements have been achieved, existing methods are still face class biases due to class variants and background confusion. In this paper, we propose a joint framework that combines more valuable class-aware and class-agnostic alignment guidance to facilitate the segmentation. Specifically, we design a hybrid alignment module which establishes multi-scale query-support correspondences to mine the most relevant class-aware information for each query image from the corresponding support features. In addition, we explore utilizing base-classes knowledge to generate class-agnostic prior mask which makes a distinction between real background and foreground by highlighting all object regions, especially those of unseen classes. By jointly aggregating class-aware and class-agnostic alignment guidance, better segmentation performances are obtained on query images. Extensive experiments on PASCAL-$5^i$ and COCO-$20^i$ datasets demonstrate that our proposed joint framework performs better, especially on the 1-shot setting.