Papers
arxiv:2407.02392

TokenPacker: Efficient Visual Projector for Multimodal LLM

Published on Jul 2
· Submitted by sunshine-lwt on Jul 4
Authors:
,
,
,

Abstract

The visual projector serves as an essential bridge between the visual encoder and the Large Language Model (LLM) in a Multimodal LLM (MLLM). Typically, MLLMs adopt a simple MLP to preserve all visual contexts via one-to-one transformation. However, the visual tokens are redundant and can be considerably increased when dealing with high-resolution images, impairing the efficiency of MLLMs significantly. Some recent works have introduced resampler or abstractor to reduce the number of resulting visual tokens. Unfortunately, they fail to capture finer details and undermine the visual reasoning capabilities of MLLMs. In this work, we propose a novel visual projector, which adopts a coarse-to-fine scheme to inject the enriched characteristics to generate the condensed visual tokens. In specific, we first interpolate the visual features as a low-resolution point query, providing the overall visual representation as the foundation. Then, we introduce a region-to-point injection module that utilizes high-resolution, multi-level region-based cues as fine-grained reference keys and values, allowing them to be fully absorbed within the corresponding local context region. This step effectively updates the coarse point query, transforming it into an enriched one for the subsequent LLM reasoning. Extensive experiments demonstrate that our approach compresses the visual tokens by 75%~89%, while achieves comparable or even better performance across diverse benchmarks with significantly higher efficiency. The source codes can be found at https://github.com/CircleRadon/TokenPacker.

Community

Paper author Paper submitter

TokenPacker: Efficient Visual Projector for Multimodal LLM

🙌High-quality visual token reduction for MLLM!

🥰We introduce a simple yet effective visual projector, which can compress visual tokens by 75%~89%.🚀

tokenpacker.jpg

I think the performamce gain may be considered as too marginal. I would like to know the gain for more challenging datasets such as mm-vet, mmstar, and llava bench in the wild. Do you have any plan? As human sensory system, the more image tokens the more benefit.

·

I think that even if it doesn't improve much, the compression of 75% of the tokens alone make it really cool approach to project the visual tokens.
please release the checkpoints :)

Great

Sign up or log in to comment

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2407.02392 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2407.02392 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2407.02392 in a Space README.md to link it from this page.

Collections including this paper 15