Efficient Large Multi-modal Models via Visual Context Compression

While significant advancements have been made in compressed representations for text embeddings in large language models (LLMs), the compression of visual tokens in multi-modal LLMs (MLLMs) has remained a largely overlooked area. In this work, we present the study on the analysis of redundancy concerning visual tokens and efficient training within these models. Our initial experiments show that eliminating up to 70% of visual tokens at the testing stage by simply average pooling only leads to a minimal 3% reduction in visual question answering accuracy on the GQA benchmark, indicating significant redundancy in visual context. Addressing this, we introduce Visual Context Compressor, which reduces the number of visual tokens to enhance training and inference efficiency without sacrificing performance. To minimize information loss caused by the compression on visual tokens while maintaining training efficiency, we develop LLaVolta as a light and staged training scheme that incorporates stage-wise visual context compression to progressively compress the visual tokens from heavily to lightly compression during training, yielding no loss of information when testing. Extensive experiments demonstrate that our approach enhances the performance of MLLMs in both image-language and video-language understanding, while also significantly cutting training costs and improving inference efficiency.

Paper

Similar papers

Peer review

Reviewer PsS75/10 · confidence 4/52024-06-30

Summary

This paper shows that visual tokens are redundant in MLLM and can be compreseed by a large ratio without significantly hurting the model performance. Based on this observation, this paper studied several different approaches to compress visual tokens and identified that the simple average pooling method is the most effective one. Based on this, the paper further studied several different stage-wise MLLM training strategies, where the training starts with heavy compression and ends with no compression. The proposed training strategy could save the training cost by 16% while achieving even better performance than the baseline.

Strengths

1. The observation that visual tokens are redundant and can be largely compressed without significantly hurting the model performance is good. It could serve as a direction for future works. 2. The paper conducted thorough empirical studies about different compression methods, and proposed several training methods based on visual token compression. 3. The proposed method shows better performance than the baseline while reducing the training cost by 16%.

Weaknesses

1. It would be better to provide more discussions and investigations on why other visual compressors are significantly worse than average pooling. This may provide more insights on how to design good compressor. 2. The paper only studied LLaVA-1.5-7B. It would be better to show that if the method could scale up to larger models like 13B, 34B, or other structures such as Mini-Gemini [1]. Showing the efficiency on those large variants can better demonstrates the effectiveness of the method. 3. The 16% training efficiency improvement is marginal in practice, especially given that the four-stage training would be more cumbersome compared to the baseline. [1] Li, Yanwei, et al. "Mini-gemini: Mining the potential of multi-modality vision language models." arXiv preprint arXiv:2403.18814 (2024).

Questions

Please see the weakness section

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

Yes

Reviewer ENde6/10 · confidence 4/52024-07-14

Summary

The paper presents a novel approach to reducing redundancy in visual tokens within MLLMs by introducing the Visual Context Compressor (VCC) and Stage-wise MLLM Training. The VCC uses simple average pooling to compress visual tokens during training, enhancing efficiency without sacrificing performance. The Stage-wise MLLM Training progressively reduces compression as training proceeds, ensuring no loss of information during testing. The proposed methods demonstrate significant improvements in training efficiency and performance across multiple benchmarks.

Strengths

1. The paper addresses an underexplored area in MLLMs by focusing on the redundancy of visual tokens and introducing effective compression techniques. 2. The proposed methods achieve significant reductions in training costs while maintaining or improving performance on various benchmarks. 3. The paper provides thorough experimental validation across multiple benchmarks, demonstrating the effectiveness of the proposed methods.

Weaknesses

1. While the paper shows improvements across several benchmarks, the evaluation is primarily focused on visual question answering tasks. Additional evaluation on other multi-modal tasks could strengthen the claims. 2. The paper mentions potential information loss due to compression but does not provide a detailed analysis/visualization of how this might affect tasks requiring dense visual information. 3. This paper lacks some essential experiments. (1) The paper does not explore the effect of using larger input resolutions. For example, evaluating the method with 448×448 input images, which contain 2.25 times more visual tokens than the 336×336 input, could provide insights into the method's scalability. Designing a compression setting with a ratio of 225% for this input size and comparing it with the original LLaVA setting (compression ratio of 100%) would be valuable. (2) Testing the method on larger models such as LLaVA-13B or models with more input visual tokens (e.g., LLaVA-Next) could solidify the experimental section and demonstrate the robustness of the proposed approach. 4. The best stage-wise MLLM training scheme looks difficult to transfer to other model & data settings. The training scheme has too many options and variables. If the model size and dataset size increase significantly, the time cost and computation cost for finding the best scheme can become prohibitively large.

Questions

Please see the weakness. Minor issues: 1. What does the #Tokens mean in the tables? Is it inversely proportional to CR? 2. How do you compute the average performance in Table 5? 3. Some suggestions for future improvements: (1) multi-stage compression: use different compressors’ strides in the different positions of the LLM. For example, layer 1-3: stride=1; layer 4-12: stride=2; layer 13-24: stride: 4. (2) extend this technique to other modalities, such as video understanding.

Rating

6

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors have adequately discussed the limitations.

Reviewer dJy45/10 · confidence 4/52024-07-15

Summary

The paper presents a compelling study on redundancy of visual tokens in MLLMs and practical approaches to reduce them. The paper first verifies that one could eliminate up to 70% of visual tokens at testing time via simple average pooling with minial performance degradation. Then they experiments with several approaches for compressing the visual tokens, finding that a simple average pooling works the best. They also propose a staged training recipe, where we could save computation during early training stages and gradually remove compression.

Strengths

1. The idea is clean and simple. It is nice to empirically verify the redundancy in visual tokens. 2. The proposed staged training and the discussion around wider then deeper v.s. deeper then wider is interesting.

Weaknesses

- There seems to be an easy baseline missing from the discussion, such as 2D conv (as in Honeybee). Honeybee: Locality-enhanced Projector for Multimodal LLM - There are some ambiguous experimental details. Please see questions.

Questions

1. When doing average pooling, why choose the 1D average pooling instead of 2D average pooling over the grid features? 2. What is K in table 2? I cannot seem to be able to find discussion on tuning k v.s. stride in Section 4.2. 3. For ablation, why report only the performance on the 4 benchmarks? I suspect that the compression performance will also be quite dataset-dependent.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Area Chair MiZt2024-08-13

Dear reviewers, Please carefully read the rebuttal and the other reviews, then reply to the authors indicating whether your questions/concerns have been addressed. If not, please specify which questions/concerns remain unresolved so that the authors have a fair opportunity to address them before the author-reviewer discussion period ends in approximately 36 hours from now (August 13th, 11:59pm AoE). Best, AC

Authorsrebuttal2024-08-13

Invitation for Discussions and Clarifications

Dear reviewers, We wholeheartedly invite and encourage discussion. We would be grateful for any further questions or comments you might have, and are fully prepared to provide additional details or clarifications. Your feedback is invaluable to enhancing the quality of our work. As AC mentioned, we are dedicated to maximizing the remaining time in the author-reviewer discussion period to address any additional concerns you may have. Authors of paper 2615

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC