We appreciate the reviewer for careful comments and provide our responses below. All changes in the revision are marked in red.

Q1: Although the use of GPT-4o for evaluation is efficient in practice, it lacks guaranteed consistency. The evaluation results could be influenced by underlying changes in GPT-4o, rendering the results unreliable. An alternative solution to this issue could be to train an independent evaluation model.

A1: Thanks for this comment. The version of GPT-4o we used is '2024-08-06'. After multiple (5 times) tests, we found that the fluctuations in the accuracy of the three metrics are within 1%, and the score fluctuations are within 0.05 on Real-Edit benchmark. This demonstrates the stability of GPT-4o. However, we fully agree with the reviewer's suggestion to train an independent evaluation model. In the future, we will explore fine-tuning a specialized evaluation model based on existing open-source multimodal large model.

We have included the above discussion in Section D.3 of the appendix in revision.

Q2: The RewardEdit-20K dataset continues to utilize images generated by InstructPix2Pix's model. For some challenging editing samples, scores cannot exceed 2 points. Even if the reward model can enhance performance in other samples, can it generate better results for these difficult cases generated by InsPix2Pix?

A2: Thank you for your insightful comment. To investigate whether our reward model can generate better edited images for challenging editing samples in InsPix2Pix, we first randomly selected 500 samples from RewardEdit-20K with scores not exceeding 2. Then, we used our reward model to generate edited images based on the original images and instructions of these samples, and scored them using GPT-4o. The experimental results are shown in the table below. "Original" represents the average scores of the original edited images of these samples across three metrics, while "Ours" represents the scores of the edited images generated by our reward model. From the table, it can be observed that the edited images generated by our method significantly outperform the original edited images on all three metrics, indicating that our method can generate better results for these difficult cases.

Tab. C: Comparison of edited images for challenging samples in InsPix2Pix.

We have integrated the above results and analysis in Section C.3 of the appendix in revision.

Q3: There is a lack of detailed analysis on the individual impact of each perspective reward on the improvement of editing models. Including such insights in an ablation study could strengthen the paper's contribution.

A3: Thank you for your suggestion. We agree that analyzing the impact of each perspective reward is beneficial and we conduct additional experiments for training on each perspective separately. As shown in the table below, the following score reached 3.40 with only the instruction following reward, the preserving score reached 3.54 with only the detail preserving reward, and the quality score reached 3.95 with only the generation quality reward. These results demonstrate the effectiveness of each perspective reward.

Tab. D: Ablation of each perspective reward. 'IF', 'DP' and 'GQ' are instruction following reward, detail preserving reward and generation quality reward.

We have included the above results in Section C.4 of the appendix in revision to make our contributions clearer.

Thank you for your rebuttal. However, there are still some concerns that need to be addressed.

About A3

The ablation study in A3 is lacking in certain aspects. Specifically, I have two questions:

1. The three evaluation metrics for each individual reward seem similar. Please analyze why the metrics increase even when the reward is not related to the specific metric.
2. Why is the evaluation of the full model inferior to that obtained by adding individual rewards?

About A5

There appears to be a bug between Equation line 281 and Figure 5. From line 281, the latent behaves as query only, then the resulting output of the cross attention $Z''$ eliminates the detailed information of the source image. However in Figure 5, $Z''$ is directly sent to the diffusion model. Per my understanding, the structure of Figure 5 is not feasible, as the diffusion model can not touch the details of the input source image, making it unable to perform image editing.

Suggestion

It is highly recommended to label each equation with a numerical label, as this is a standard practice.

Summary

I am inclined to reject this paper as the readability could be significantly enhanced.

We thank the reviewer for his careful and helpful comments on our work. We provide our responses below to answer the reviewer's questions. All changes in the revision are marked in red.

Q1: The multi-view reward mechanism used in this paper relies entirely on GPT-4o’s evaluation. Although GPT-4o demonstrates strong capabilities in understanding and generating natural language, it may not fully capture the subtle nuances of human perception regarding image editing quality.

A1: Thank you for this comment. We agree that the annotation/evaluation from GPT-4o is not as good as human annotation. However, human annotation is very expensive and time-consuming, making it unsuitable for large-scale data generation. In contrast, GPT-4o-based data generation is scalable with reasonable quality. Moreover, our experiments demonstrate that using multi-view rewards generated by GPT-4o can still significantly improve the model's image editing performance, indicating the reliability of our method.

We have included the above discussion in Section D.3 of the appendix for clarification.

Q2: The cost for this reward is expensive as it using GPT-4o .

A2: Thanks for this comment. The price for GPT-4o is \$2.5 per 1 million tokens. Evaluating a single sample costs \\$0.004, and evaluating the entire Real-Edit benchmark costs approximately \$2.24, which is acceptable/negligible compared to training costs such as GPU expenses. Frankly, using GPT-4o incurs additional costs compared to free metrics, but we believe it's worth it. As the most powerful multimodal model, GPT-4o provides more accurate evaluations of editing performance. In the future, we will attempt to train specialized editing evaluation models to replace GPT-4o.

Q3: The human evaluations showed slightly lower scores than those generated by GPT-4o, though both showed consistency in ranking. Could the authors provide insights into why this discrepancy exists?

A3: Thanks for this comment. We guess that the reason for this discrepancy may be that human evaluators often have higher expectations (e.g., setting a higher bar for a perfect score of 5) and subjective perceptions, making them more critical of details and quality. This results in lower scores compared to GPT's evaluations.

We have added the above analysis in Section 6.3 in revision.

Q4: Given the reliance on the RewardEdit-20K dataset, a more detailed release plan for this dataset (and possibly pretrained models) could be helpful.

A4: Thanks for this comment. We are currently organizing the data and code and plan to publicly release all code (including training, testing, evaluation code) and pretrained models on GitHub within the next two months. Additionally, the multi-perspective reward dataset RewardEdit-20K and the evaluation benchmark Real-Edit will also be made available on Huggingface.

Q5: While the approach shows significant improvements, an analysis on instances where the multi-reward mechanism fails or provides subpar results would be beneficial. Understanding the limitations could offer insights for future iterations or refinements of the method.

A5: Thank you for your suggestions. We agree that adding an analysis of the failed cases would be beneficial for our method. To explore the limitations of our method, we collected and analyzed failed cases (please see Section A of the appendix in the revision). The analysis revealed two main limitations of our method. The first limitation is that during testing, even when the given multi-perspective reward scores are all 5, the generated edited image does not always achieve a score of 5. This indicates that the reward information does not always perfectly guide the model, especially in some complex cases. The second limitation is that our method has difficulty accurately understanding the quantifiers and spatial position words in the instructions, as shown in Fig. 10 in revision. This may be due to the model's insufficient understanding of fine-grained language features. In future work, we will explore ways to improve the model's understanding of fine-grained semantics for image editing.

We have added the above analysis to Section A of the appendix in the revision.

We thank the reviewer for helpful comments on our work and provide our responses to the reviewer's questions below. All changes in the revision are marked in red.

Q1: Both the quantitative assessment in Table 1 and the training strategy are based on the same scoring strategy for the GPT-4o, making the evaluation of the results overly dependent on the a priori of the VLLM and making it difficult to objectively validate the strengths of this method. One solution to this dilemma is to use other assessment metrics (e.g., CLIP scores) on the three dimensions to be adopted for comparison with other methods.

A1: Thanks for this comment. Evaluating edited images presents significant challenges due to the diversity and uncertainty of possible edited images for a given original image and instruction. GPT-4o is currently the most powerful multi-modal understanding model, capable of accurately parsing editing instructions and comprehensively understanding the content of the original images and the editing requirements. Therefore, we use GPT-4o for automated image editing evaluation. To verify the reliability of GPT-4o, we also performed human evaluations and obtained consistent results.

Additionally, we agree with the reviewers that it is necessary to evaluate the dimensions of Following,Preserving, and Quality based on existing evaluation metrics. We recalculated the performance of existing methods and our method on Real-Edit based on the CLIP score and the FID score, with the results shown in the table below. Specifically, the CLIP feature similarity between the edited image and the instruction is the following score, the similarity between the original and edited images is the preserving score, and the FID between the original and edited images is the quality score. The table shows that our method still achieved promising results and improvements over the baseline.

Tab. A: Comparison of different methods based on existing evaluation metrics.

However, these metrics also have limitations: 1) When the editing instruction and the images are complicated, CLIP/FID score can not accurately represent the following/preserving/quality of the edited image, e.g., CLIP can not distinguish left/right. 2) the range of the following score and preserving score is relatively small, which may make it hard to distinguish performance differences between methods.

We have included the above discussion and results in Section C.1 of the appendix in the revision.

We thank the reviewer for careful comments on our work and provide our responses below. All changes in the revision are marked in red.

Q1: The paper lacks a comparison with RL in T2I methods like DPO-Diffusion[1]. If I understand correctly, I believe that the Multi-Reward Framework is conceptually similar to methods like DPO-Diffusion. Therefore, I think it is reasonable and necessary to articulate the comparisons and distinctions between these approaches, especially the method difference.
[1] Wallace, B., Dang, M., Rafailov, R., Zhou, L., Lou, A., Purushwalkam, S., Ermon, S., Xiong, C., Joty, S., & Naik, N. (2023). Diffusion Model Alignment Using Direct Preference Optimization. arXiv preprint arXiv:2311.12908.

A1: Thanks for the insightful comment. We agree with the reviewer's suggestion that we should clarify the differences and connections between our method and methods like DPO-Diffusion. Both DPO-Diffusion and our proposed Multi-Reward approach fundamentally aim to optimize the quality of generated images through feedback mechanisms. The main differences between our Multi-Reward and DPO-Diffusion are as follows:
(1) Granularity of feedback. DPO-Diffusion's preference feedback is expressed as relative preferences, such as `Image A is better than image B', therefore the feedback signal only has two possible states. In contrast, our Multi-Reward uses absolute numerical values and detailed text description for feedback signals (For examples, see Appendix Section B).
(2) Applicability of feedback. DPO-Diffusion is only applicable to situations with a single feedback value, whereas our approach can simultaneously incorporate multi-perspective feedback information, including instruction following, detail preserving and generation quality.
(3) Training stability. We directly use feedback information as an additional condition while still employing the original Diffusion Loss. This approach is simple and effective, avoiding the training instability that DPO can introduce to the diffusion model.

We have integrated the above detailed comparison into Section D.4 of the appendix in the revision.

Q2: The article lacks some novelty. Of course, high-quality and abundant data can effectively enhance model performance, so I am uncertain about the extent to which the proposed Multi-Reward Framework improves upon traditional methods. For example, in Table 1, additional editing data (0.02M) was used for training. I believe it is fair to compare it with the exact same data using the same baseline method; otherwise, it is difficult to convince me whether the improvement comes from the high-quality data or the architecture. Perhaps experiments on the unprocessed REWARDEDIT-20K data could be added for comparison.

A2: Thank you for your thoughtful feedback. To clarify, we are not working on using MLLMs to filter high-quality data from InsPix2Pix. The 20K samples in our RewardEdit-20K dataset are randomly sampled from InsPix2Pix. Our motivation is that constructing a perfect image editing dataset is challenging, and the ground truth in existing image editing datasets often contains issues. Therefore, we propose using multi-perspective rewards to rectify the inaccurate supervision. To more fairly demonstrate the role of multi-perspective rewards, we conducted the ablation experiments shown in the table below. The experimental results indicate that, with the same data, using multi-perspective rewards significantly improves performance compared to the baseline, demonstrating the effectiveness of multi-perspective rewards.

Tab. B: Ablation study of editing data.

We have included the above clarification and results in Section C.2 of the appendix in the revision to make our contributions clearer.