Instruction-Guided Visual Masking

Instruction following is crucial in contemporary LLM. However, when extended to multimodal setting, it often suffers from misalignment between specific textual instruction and targeted local region of an image. To achieve more accurate and nuanced multimodal instruction following, we introduce Instruction-guided Visual Masking (IVM), a new versatile visual grounding model that is compatible with diverse multimodal models, such as LMM and robot model. By constructing visual masks for instruction-irrelevant regions, IVM-enhanced multimodal models can effectively focus on task-relevant image regions to better align with complex instructions. Specifically, we design a visual masking data generation pipeline and create an IVM-Mix-1M dataset with 1 million image-instruction pairs. We further introduce a new learning technique, Discriminator Weighted Supervised Learning (DWSL) for preferential IVM training that prioritizes high-quality data samples. Experimental results on generic multimodal tasks such as VQA and embodied robotic control demonstrate the versatility of IVM, which as a plug-and-play tool, significantly boosts the performance of diverse multimodal models, yielding new state-of-the-art results across challenging multimodal benchmarks. Code, model and data are available at https://github.com/2toinf/IVM.

Paper

Similar papers

Peer review

Reviewer zLHv6/10 · confidence 4/52024-07-02

Summary

In this paper, the author introduce Instruction-guided Visual Masking (IVM), a generic and powerful visual grounding method that enhances broad multimodal instruction following tasks in a plug-and-play way. By masking out all instruction-irrelevant image regions, IVM effectively injects superior visual grounding ability to downstream LMMs non-intrusively, significantly boosting both commercial and open-sourced LMMs and achieving state-of-the-art results across numerous challenging multimodal benchmarks.

Strengths

Overall, the motivation of this paper is commendable, as it addresses an interesting problem and provides rich illustrations that facilitate understanding. Furthermore, the work is supported by a substantial number of experiments to validate its effectiveness.

Weaknesses

1. The routine of retraining models by preparing new datasets, as described in this paper, can be considered somewhat old-fashioned. 2. The technical insight of this paper is relatively weak, and no strong novelty technical methods are proposed.

Questions

I have a few minor concerns that I would like the author to address. 1. The routine of retraining models by preparing new datasets, as described in this paper, can be considered somewhat old-fashioned. Recent literature has proposed similar grounding datasets, such as Ferret (GRIT) [1] and Kosmos-2 (GRIT) [2]. However, these two works are not cited. There should be some new discussion on motivation about the dataset in this paper. 2. Why does the instruction-related area between 90% and 100% in Figure 5 increase abnormally significantly? It is necessary to consider whether these data needs to be cleaned. 3. Formulas 1 and 2 in the paper are somewhat deliberately mystifying. I understand that the purpose is to learn an optimal model from the hand-labeled data and the auto-generated data. This process is somewhat similar to the pseudo-label curriculum learning proposed in CLIP-VG [3]. It is suggested to increase relevant discussion and citations. 4. I have some confusion regarding Figure 6 of the paper, which requires modification. Specifically, (1) the discriminator and generator should be two separate processes, while in Figure 6, the image, text, and attention are depicted as being fed to the model simultaneously; (2) It is unclear whether the LLM in the figure represents one model or two models since there are two LoRAs shown but only one LMM is illustrated. 5. A framework diagram of the inference model used in the downstream experiment should be drawn according to Figure 6, so as to illustrate how the IVM assists the model in inference. 6. Since the paper claims that IVM is a plug-and-play method for assisting LLM, so I am curious about how this paper performs on the RefCOCO/+/g dataset after incorporating IVM. After all, the motivation behind this paper stems from the visual grounding task, and there has been a number of research on grounding multimodal large language models, such as Ferret [1], Kosmos-2 [2], and LION [4]. 7. Other writing issues, such as some vocabulary in the paper should be consistent. For example, (a) line 135 RefCoCo should be changed to RefCOCO to maintain uniformity; (b) Lora is used in Figure 5, while LORA is used in the main text, and it is recommended to use LoRA uniformly in order to be consistent with the original article; (c) LMM and LLM are used in this paper, however LLM has not the full vocabulary, and it is suggested maybe better to use LLM and MLLM uniformly. These rudimentary errors should not appear in NeurIPS submission papers. On the whole, this paper is a relatively solid work, the overall presentation is good, literature review is relatively sufficient, so I currently give a positive scores. Hope the author can address my concerns, and I will decide whether to raise down or raise up my rating according to the author's rebuttal reply. -- [1] You, Haoxuan, et al. "Ferret: Refer and Ground Anything Anywhere at Any Granularity." The Twelfth International Conference on Learning Representations. 2024 [2] Peng, Zhiliang, et al. "Kosmos-2: Grounding multimodal large language models to the world." arXiv preprint arXiv:2306.14824 (2023). [3] Xiao, Linhui, et al. "CLIP-VG: Self-paced Curriculum Adapting of CLIP for Visual Grounding." IEEE Transactions on Multimedia (2023). [4] Chen, Gongwei, et al. "Lion: Empowering multimodal large language model with dual-level visual knowledge." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

see weaknesses.

Authorsrebuttal2024-08-12

Thanks for increasing the score!

We really thank the reviewer for increasing the score, and sorry for the confusion in the rebuttal phase! Regarding the remaining concerns, we provide the following detailed responses. >**Explanation about Eq. (1) and Eq.(2)** - Due to the space limits in the rebuttal phase (6000 characters), we did not include very detailed discussions on Eq. (1) and Eq. (2) in the rebuttal. The introduction of Eq.(1) and Eq. (2) is inspired by recent offline IL work [2] that tries to learn a good model jointly from a high-quality near-expert dataset (human data) and a mixed-quality suboptimal dataset (machine-generated data), which typically consists of two stages: - `Discriminator Training`. Eq. (1) tries to train a discriminator $d$ to distinguish between human data $\mathcal{D}_E$ and the machine-generated data $\mathcal{D}_D$. By doing so, the output of the discriminator $d$ becomes a confidence value for data quality ($d$ will output near 1 or near 0 if the data can be clearly identified as human or machine data; $d$ will output around 0.5 if the annotation quality is hard to judge), see Figure 9 (b) for discriminator output statistics. - `Discriminator-weighted Supervised Learning`. After training the discriminator, the $d$ value can be an adaptive weight function to reprioritize the training annotations in Eq. (2), where high-quality annotations will be more fitted than low-quality ones. In this sense, the side-effect of bad annotations in the mixed-quality machine data can be largely filtered out, see Figure 9 (a) for detailed experimental evidence. Here, all annotations are pre-collected rather than gradually updated during training as CLIP-VG does. So, we're more insensitive to the compounding errors than CLIP-VG. >**Additional discussions on Q6** - We are really sorry for the confusion here. Due to time limits, we tested IVM only on the validation split rather than all refcoco splits. Now, we're trying to implement IVM-13B by scaling IVM on more human and machine data. After that, we will conduct more thorough experiments in the future. >**I expect the authors to make corresponding revisions regarding the aforementioned issues in the final version.** - Sure! We will consider all these helpful suggestions when revising our paper! Thanks to the reviewer again for the efforts and engagement in the discussion phase, and open to any further comments! [2] Discriminator-weighted offline imitation learning from suboptimal demonstrations. ICML 2022

Reviewer zLHv2024-08-12

Thanks for the authors' responses. That's all.

Thanks for the authors' responses. That's all.

Reviewer EbuK6/10 · confidence 4/52024-07-12

Summary

This paper introduces the Instruction-guided Visual Masking (IVM), a versatile visual grounding model designed to improve alignment between textual instructions and specific image regions. It outlines the development of a visual masking data generation pipeline and a new learning technique, Discriminator Weighted Supervised Learning (DWSL), which prioritizes high-quality data samples to enhance performance on multimodal tasks.

Strengths

1. This paper introduces instruction-guided visual masking (IVM). The IVM-enhanced multimodal models can focus on task-relevant image regions to better align with complex instructions. This implies that the model can become more sensitive to instructions. 2. Figures 2 and 3 are helpful to understand the method. 3. This paper has collected a richer and more complex visual grounding dataset.

Weaknesses

1. The IVM model architecture shown in Figure 6 is not conducive to understanding the approach. 2. It is recommended to compare more methods on a multimodal benchmark.

Questions

1. The paper mentions that the IVM-based model does not show significant gains on the GQA, SQA, and VQAv2 benchmarks. It suggests that these benchmarks may not depend heavily on grounding abilities, which raises some doubts for me. Intuitively, the results for simple visual inputs related to instructions should be better. 2. Additionally, I am concerned that the improved performance of the proposed method over comparative models might largely be attributed to the use of a larger Visual Grounding dataset. I am interested in seeing how other models would fare if they were trained or fine-tuned using the same dataset.

Rating

6

Confidence

4

Soundness

4

Presentation

2

Contribution

3

Limitations

Limitations and broader impact have been discussed.

Reviewer EbuK2024-08-12

I feel that my concerns were not fully addressed. The explanations provided seem more intuitive and lack solid theoretical or experimental support, such as Q1 and Q2. Given this, I decided to lower my rating to weak accept.

Authorsrebuttal2024-08-12

Thanks for the further comments

We thank the reviewer for the efforts and engagement in the discussion phase. Regarding the remaining concerns, we provide additional responses as follows: >**Additional responses to Q1** - For Q1, we provide the analysis of the relationships about data quantities w.r.t different `instruction-relevant visual ratios (IVR)` (the ratios of preserved pixels in the original images) generated by IVM on different multimodal benchmarks including the simple GQA, SQA, VQAv2 (due to time limits, we evaluated on 1/10 subset of the VQAv2 dataset, if the reviewer would like to see results on full datasets, we will try our best to finish the experiments by the author rebuttal deadline) and the more complex V* and EgoThink benchmarks. |Benchmark|<20% (IVR) |20%-40% (IVR)|40%-60% (IVR)|60%-80% (IVR)|>80% (IVR)| |---|---|---|---|---|---| |V*|100%|0%|0%|0%|0%|0%| |Egothink| 67% | 14% | 4% | 3% | 12%| |VQAv2 (40K samples)| 13% | 16% | 12% | 10% | 49%| |GQA| 17% | 14% | 8% | 4% | 57% | |SQA| 0% | 5%| 11% | 7% | 77% | - These statistics demonstrate that most visual contents in GQA, SQA and VQAv2 benchmarks are instruction-relevant (>80% IVR). Existing base MLLMs can easily focus on the correct image areas to follow the instructions, rather than be distracted by the minor visual distractors (for example, see Figure b in the PDF in **General Response**). - On the contrary, only a small ratio of visual contents in the challenging V* and Egothink benchmark are instruction-relevant (<20% IVR). In this case, MLLMs are more likely to be distracted by a lot of irrelevant visual content if the MLLMs' visual grounding ability is not that strong. With the IVM assistance, however, the performance can be greatly enhanced via the explicit surgical visual grounding. >**Additional responses to Q2** - For Q2, `LISA is the only comparable baseline in our setting`. Specifically, most MLLMs cannot be directly finetuned or trained in our setting as they are not primarily designed to predict dense image heatmaps given complex instructions that require strong reasoning ability. Among all these MLLMs, LISA is the most relevant work that can be fairly compared, as LISA is specifically tailored for reasoning segmentation, which shares some similarity with our setups. Indeed, we have empirically compared the baseline method LISA (please check the Human+machine label (SL) baseline in Figure 9 (a) in our paper for details). Here, we summarize the results more directly (detailed comparisons of diverse data quantities and data components can be found in Figure 9 (a)). |Training object & data|IVM on IVM-Mix-1M (Ours)|LISA on IVM-Mix-1M| |---|---|---| |Improvements on V* benchmark|**+26.2**|+16.2| `If the reviewer has any further detailed concerns or requires any other experimental supports, please do not hesitate to point them out. We will be more than happy to address them and improve the quality of our paper.`

Authorsrebuttal2024-08-13

Hi! The discussion is approaching to close. So, hope our additional responses successfully address your remaining concerns. If not, please do not hesitate to point them out. We would really appreciate any further comments that can improve the quality of our paper!

Reviewer bwCu8/10 · confidence 5/52024-07-13

Summary

This paper presents IVM (Instruction-guided Visual Masking). The key idea is that we could mask out the instruction-irrelevant regions in the given image. The trained model is tasked to mask out the irrelevant regions, enforcing the multimodal model to focus on the task-related visuals. Such grounding-centric approach is effective in enhancing different multimodal models. The paper also details their solution to create a large number of reliable pixel-level labels. The paper presents a MoE pipeline with various visual grounding models to collect reliable labels. The paper further introduces DWSL for IVM training. DWSL helps IVM training to focus more on higher quality training data.

Strengths

- The proposed IVM is sound and effective. I found the problem formulation of generating heatmap interesting. The experiments clearly demonstrate the effectiveness of the proposed method. - The MoE pipeline is well-developed and can be applied to both labeled and unlabeled data. - The IVM architecture with discriminator training can effectively reduce the impact of low-quality data. - The paper is well-written and solid.

Weaknesses

I don't find significant problems in the paper. One possible improvement for this paper is that it would be interesting to provide some typical failure cases of the proposed method.

Questions

see weakness

Rating

8

Confidence

5

Soundness

3

Presentation

3

Contribution

4

Limitations

N/A

Reviewer zLHv2024-08-12

I am not convinced by reviewer bwCu's comments and rating, as it seems that the reviewer bwCu was aware of the author's identity and deliberately gave high marks.

After reading the comments of the other reviewers, I am not convinced by reviewer bwCu's comments and rating. To be specific, - (a) The reviewer bwCu claims that his confidence in this paper is "absolutely certain about your assessment. You are very familiar with the related work." However, the reviewer bwCu did not give some valuable and in-depth comments to this paper. I think, if the reviewer bwCu is familiar with this field, he should be able to find the defects of this paper and make some valuable comments. - (b) Reviewer bwCu gave nothing but praise for this paper; - (c) Judging from the comments of the other three reviewers, there are still more or less problems in this paper. However, reviewer bwCu directly gave the extreme rating of "Strong Accept". To sum up, it seems that the reviewer bwCu was aware of the author's identity and deliberately gave high marks.

Reviewer EbuK2024-08-12

I agree with reviewer zLHv's perspective on the concerns regarding reviewer bwCu's comments and rating. Compared to the comments from other reviewers, reviewer bwCu's comments seem overly positive and lack the professional suggestions expected. This raises my curiosity about the reason behind the reviewer bwCu's strong accept rating.

Reviewer 5SdR6/10 · confidence 3/52024-07-14

Summary

For the purpose of precise instruction following performance in LLM, this paper proposes a versatile grounding model that is compatible with diverse multi-modal models. Leveraging the LLM, a visual masking data generation pipeline is built and 1 million image-instruction pairs are constructed. On top of it, an Instruction-Guided Visual Masking (IVM) model is trained to focus on task-relevant regions that align with complex instructions.

Strengths

1. An IVM model is proposed to enhance multimodal instruction following via nuanced surgical visual grounding. Overall, this model is simple yet effective. Such a model can be seamlessly incorporated into a multimodal model to boost the performance of downstream tasks. 2. A dataset creation pipeline with a mixture of experts is carefully devised and along with it an 1-M dataset is built. 3. To effectively utilize the dataset, a discriminative weighted supervised learning training strategy is devised to select the high-quality dataset pairs. 4. Extensive experiments have been conducted to validate the effectiveness of the proposed approaches on various tasks, e.g. visual understanding, reasoning segmentation model and real robot control.

Weaknesses

1. This work proposes a visual grounding model. However, it seems that the comparison with state-of-art visual grounding methods (VG) is missing. I understand this task involves slight differences with reasoning segmentation (RS) tasks. But will the VG task evaluation be more direct? 2. This approach relies heavily on the human manually labeled data, without which the performance will significantly drop. Therefore, it is more like a dataset creation work. Finetuning on this dataset, other compared works might also achieve similar or even better performance. 3. The derived framework is a little bit heavy, with LLM and heavy visual rendering head involved . There might be a more efficient network framework to achieve similar performance.

Questions

Refer to the questions in weakness section. Other questions related to technical details. 1. How does this approach adapt to images with distinct image resolution? Is there any specifically designed relative position embedding? 2. Is there any plan to extend this work to more complicated robot action? The pick and put are very simple action. The real world application will involve more convoluted operations such as visual navigation or sequential manipulation.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

Authors adequately addressed the limitations.

Reviewer zLHv2024-08-12

The author's response is polite, and I appreciate the author's efforts during this rebuttal process.

I noticed that the author did not fully respond to some of my questions, such as explaining equations 1 and 2 in Q3, etc. Additionally, the authors' results in Q6 appear to be insufficient compared to several baselines. Nevertheless, Considering the author's response is polite, and I appreciate the author's efforts during this rebuttal process. Therefore, I would like to raise my rating by 1 point to `weak accept`. I expect the authors to make corresponding revisions regarding the aforementioned issues in the final version.

Authorsrebuttal2024-08-12

Thanks for the responses!

Thanks to the reviewer again for the efforts and engagements, and open to any further comments!

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC