The paper introduces TokenPacker, an efficient visual projector designed for Multimodal Large Language Models (MLLMs). TokenPacker aims to bridge the visual encoder and the LLM by generating condensed visual tokens while preserving fine-grained visual details. The method employs a coarse-to-fine scheme, first downsample visual features to a low-resolution representation and then uses high-resolution region-based cues to enrich this representation. This process effectively reduces the number of visual tokens by 75% to 89% while maintaining or improving performance across various benchmarks. The approach is evaluated on multiple datasets, including CC-558K, Mini-Gemini, and LLaVA-1.5, demonstrating superior efficiency and effectiveness in high-resolution image understanding. The source code is available at https://github.com/CircleRadon TokenNamecker.The paper introduces TokenPacker, an efficient visual projector designed for Multimodal Large Language Models (MLLMs). TokenPacker aims to bridge the visual encoder and the LLM by generating condensed visual tokens while preserving fine-grained visual details. The method employs a coarse-to-fine scheme, first downsample visual features to a low-resolution representation and then uses high-resolution region-based cues to enrich this representation. This process effectively reduces the number of visual tokens by 75% to 89% while maintaining or improving performance across various benchmarks. The approach is evaluated on multiple datasets, including CC-558K, Mini-Gemini, and LLaVA-1.5, demonstrating superior efficiency and effectiveness in high-resolution image understanding. The source code is available at https://github.com/CircleRadon TokenNamecker.