Kamera: Unified Position-Invariant Multimodal KV Cache for Training-Free Reuse

The paper introduces Kamera, a novel approach for enhancing position-invariant multimodal key-value (KV) caching, which enables the reuse of cached data without the need for re-encoding. By implementing a low-rank conditioning patch alongside each cached chunk, Kamera addresses the loss of cross-chunk conditioning that traditional KV caches suffer from, significantly improving multi-hop reasoning accuracy while reducing computational overhead. This method demonstrates substantial performance gains on benchmarks like MM-NIAH and two-page doc-QA, making it particularly valuable for practitioners working on multimodal AI systems that require efficient resource utilization during processing.