iLLaVA: An Image is Worth Fewer Than 1/3 Input Tokens in Large Multimodal Models

In this paper, we introduce iLLaVA, a simple method that can be seamlessly deployed upon current Large Vision-Language Models (LVLMs) to greatly increase the throughput with nearly lossless model performance, without a further requirement to train. iLLaVA achieves this by finding and gradually merging the redundant tokens with an accurate and fast algorithm, which can merge hundreds of tokens within only one step. While some previous methods have explored directly pruning or merging tokens in the inference stage to accelerate models, our method excels in both performance and throughput by two key designs. First, while most previous methods only try to save the computations of Large Language Models (LLMs), our method accelerates the forward pass of both image encoders and LLMs in LVLMs, which both occupy a significant part of time during inference. Second, our method recycles the beneficial information from the pruned tokens into existing tokens, which avoids directly dropping context tokens like previous methods to cause performance loss. iLLaVA can nearly 2$\times$ the throughput, and reduce the memory costs by half with only a 0.2\% - 0.5\% performance drop across models of different scales including 7B, 13B and 34B. On tasks across different domains including single-image, multi-images and videos, iLLaVA demonstrates strong generalizability with consistently promising efficiency. We finally offer abundant visualizations to show the merging processes of iLLaVA in each step, which show insights into the distribution of computing resources in LVLMs. Code is available at https://github.com/hulianyuyy/iLLaVA.