DivPrune introduces a diversity-based pruning technique that significantly reduces inference time in Large Multimodal Models (LMMs) by intelligently removing redundant visual tokens.

• Reduces thousands of visual tokens while maintaining model performance
• Achieves up to 90% reduction in visual tokens with minimal accuracy impact
• Implements a novel approach using token diversity metrics instead of content importance
• Demonstrates improved efficiency across multiple LMM architectures

This engineering breakthrough directly addresses a critical limitation in multimodal AI deployment: the high computational cost and latency caused by processing numerous visual tokens.

DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models