QuFiH: Hybrid low-bit quantization and block-level parameter efficient fine-tuning for video hashing

Current video hashing methods lack large-model-based deep feature representation capabilities, resulting in suboptimal retrieval performance on large-scale video databases. To address this, we propose Video-HLM (Video-Hashing Large Model), built upon the pretrained Video-LLaMA. However, its computational demands hinder practical deployment. To balance efficiency and performance, we develop a novel video hashing framework that fuses hybrid low-bit Quantization and block-level parameter-efficient Fine-tuning (QuFiH). Specifically, QuFiH employs 4-bit quantization for the foundation model and 2-bit quantization for the hash head, optimizing storage and performance. Additionally, we propose block-level LoRA/Propulsion, reducing redundant parameters while maintaining model expressiveness with lower computational overhead. Furthermore, we explore a “distill-then-finetune” strategy, combining knowledge distillation with the downstream task fine-tuning to enhance generalization and retrieval performance. Rich experiments on public datasets demonstrate QuFiH’s superiority. It compresses model parameters by 6.84× while outperforming state-of-the-art hashing methods in mAP@100, even surpassing the full-precision Video-HLM. QuFiH achieves dual advancements in efficiency and accuracy, offering a practical solution for large-scale video retrieval in resource-constrained environments. Code is available at: https://github.com/kydbj/QuFiH.