We appreciate the reviewer's recognition of our work that "viewpoints are supported by experimental evidence", "theoretical analysis is detailed", "experiments is reasonable with clear logical progression and evidence"

1.Verification on larger dataset:(1) As suggested, we conduct experiments on a larger dataset. We curate a dataset with a total of 1000 samples, which is currently the largest for image editing metric evaluation. Images are randomly selected from MagicBrush and edited by 6 advanced editing models, which include the reviewer suggested ACE and Genartist model. The results on the newly expanded dataset is shown in Table1. Our BPM yield highest human alignment score on this dataset as well, validating its effectiveness and robustness.(2)Although our BPM metric calculation is fully automated, the user study required for human alignment test hinders further dataset scale-up.

2.Rationality behind $f_{position}$ and $f_{size}$:(1) Our intension is that previous metrics evaluate the entire image, thereby overlooking region-specific semantic requirements and failing to assess whether the fine-grained state changes of edited objects align with instructions. Size and position serve as typical state changes, thus we design corresponding function for their evaluation. (2) regarding how to evaluate the size and position change to make it more generalisable, we also compare GPT-4o that directly judge whether the editing obeying corresponding state change demanding. The results are shown in Table 2, our tailored functions $f_{size}$ and $f_{position}$ yield more human-aligned judgment, validating the trustworthiness.

3.More experimental results on BPM-based editing images: Apart from BPM evaluation, we further provide following results: (1)User study: we conduct user preference tests, which is provided in Table12 of Supp.M. BPM-based editing yield 91% of the samples being rated as better or equally good as the original compared to original ones, validating BPM serves as effective editing quality enhancement. (2) Evaluation with other metrics: We also evaluate images by other metrics, which is shown in Table3. As shown, our BPM-based editing images yield the better performance for all these metrics as well, validating the robustness of BPM quality enhancement.

4.Inclusion with more related works:(1)Suggested works ACE and GenArtist focus on new framework for image generation and editing, but they all use conventional metrics for evaluation. In contrast, our main contribution is the new evaluation metrics. (2)Experimental comparisons of testing them using our new metrics are included in Table 1, and we will add relative discussion in related work in final version.

We appreciate the reviewer's recognition of our work that "claim is solid", "metric is novel".

Limited scope of proposed metric, i.e., global editing and object attribute changing: 1) Thanks for recognize our proposed metric suitable for local object-oriented editing. 2) As suggested, we apply our BPM to global instruction-based editing including style change, environmental change. a) visualized results is in Fig. b)Quantitative verification we use the 160 global editing samples in PIE-bench for human alignment evaluation. Results are in Tab. BPM yield highest human alignment for such global editing cases. c) Comparison with GPT-4o(VLLM competitors): While applicable, GPT-4o(HQ-Edit) underperforms our BPM in global editing. d) Unlike metrics like FID(conventionally used for style transfer) that rely on paired GT before-and-after images (“after” indicates target stye image), our proposed metric operates without this constraint, offering greater practicality, our metric operates without this constraint, offering greater practicality.3) Regarding attribute changing(a)Clarify object identify step for attribute changing cases, source/target object $o_1,o_2$ would affiliate corresponding attributes, like, "yellow mug" as $o_1$, "blue mug" as $o_2$ for "turning yellow mug into blue" b) Attribute understanding Our contribution is decompose image into editing-relevant and irrelevant regions for separate evaluation, allowing any CLIP variant to be applied. In our rebuttal, we test Long-CLIP-specialized in fine-grained attribute understanding-on a subset of 80 attribute-changing samples (Tab). Results show BPM’s evaluation capability for such cases can be improved with more advanced image encoder.

Benefits of Human Alignment Test on two models comparison rather than single model: 1) As suggested, we report metric and human score correlation for each single model in Tab. Though our BPM yields highest correlation in this format as well, such calculation is questionable as output from distinct instruction lacks comparability, and all metrics show performance not high.2)The rationale behind our pair model preference calculation is a) Reduce Subjectivity influence: Scores among users exhibit variance of 1.06(out of 0-5 range), thus using absolute values for single model correlation calculation may introduce bias. In contrast, employing relative preferences, by introducing a reference, reduces subjectivity in user study to enhance the credibility of the results. b) Reduce cross-sample bias: our metric evaluates which model performs better for the same instruction, rather than comparing scores across different instructions (as we believe outputs from distinct instructions lack absolute comparability). By pair-wise comparison on identical instructions, we enable more accurate model-to-model comparisons explicitly than correlating scores from a single model's diverse outputs.

Details of the detector and segmentor We use Grounding-DINO 1.5 and SAM-large as our detector and segmentor respectively. Will add this in final version.

Corner case of $f_{position}$ and $f_{size}$ In Fig, for object addition type instruction like "add dogs to the right of the surfboard",1) regarding $f_{size}$ (Line 8 of Alg.1, main text), since original object $o_1$ is None, $S_{size}$ is set as 1.2) regarding $f_{position}$ (Line 7 of Alg.1, main text), since $pos_{st}$ is "right", LLM would replace $o_1$ by the reference object surfboard, and judge whether $o_2$ dog obey "right of" $o_1$ surfboard.

Fairness of visualized comparison with other metrics: (1)We provide more visualizations about the metric human alignment comparison Fig. Across diverse types of editing instructions involving local attribute editing, object addition, global style transfer, our BPM can yield higher human alignment compared to GPT-4o and other metrics.
(2)this mask comparison experiment in Fig 11.Supp only aims to justify the design choice of proposed route to locate mask.

Apply mask into diffusion to for edit is not new. Agreed, 1)Our main contribution is new image editing metric that includes a novel pipeline for precise edited-region mask acquisition.2)this application is not our core contribution. It aims to show BPM design principle has broader potential as a byproduct for image editing.

Add reference to CLIP directional similarity: Agreed, we will add reference and clarify in related work, fix overclaim point in final version. Thanks for recognize the rest of the proposed metric is novel.

Typo in Eq.2: Thanks. We will correct this typo and incorporate all other valuable non-technical writing suggestions in the final version.

We appreciate the reviewer's recognition of our work that "enabling more precise evaluation", "broad applicability", "paper is well written", "motivation is clear and easy to understand" and "contributes to the broader scientific literature".

1.More concise figure of BPM pipeline: Thanks for your valuable suggestions, according to which we provide a new version of Figure 2 in Fig.

2.Inference speed analysis: We provide inference speed comparison in Table7, Supp.Material.. (1) Region-Aware and Semantic-Aware Judge are independent and can conduct in parallel, and will not affect inference speed. (2) our BPM has comparable inference speed ($\textless$1s per image) with other metrics like CLIP-I, and largely shrink the time cost compared with MLLM-based evluation, such as GPT-4o(11.4s per image), showing efficiency.

3.Assigning different weights to $S_{semantic}$ and $S_{region}$: We conduct experiments with different weight factor $\alpha$ between $S_{semantic}$ and $S_{region}$, i.e., $BPM= \alpha*S_{semantic} + (1-\alpha)*S_{region}$, the results are shown in Table1. Among all combinations, $\alpha=0.7$ reaches the best human evaluation alignment, verifying that higher yet moderate semantic scale would yield better evaluation results.

We appreciate the reviewer's recognition of our work that "expected to have a high impacts on image editing domains", "claims are well evaluated with sufficient evidences", "human alignment testing is effective".

1.Robustness of BPM on more diverse image editing tasks: (1)Our BPM can apply to global image style transfer like weather change and environment change as well, we provide visualizations in Fig. For quantitative evaluation, we conduct our experiments on the global editing subset of PIE-bench, which contains 160 samples. The results are shown in Table1. Our BPM yields the highest human alignment compared to other metrics, showing its effectiveness and wild application scope. (2) We want to clarify the focus of this paper is introducing new metrics for Instruction-Based Image Editing, specifically targeting edits that require clear semantic changes. Suggested tasks like image enhancement or flare removal—which primarily improve image quality rather than alter semantics, fall outside the scope of our work.

2.Apply balanced factor between preservation and modification: Thanks for the suggestions. We conduct experiments with different balanced factor $\alpha$ between $S_{preserve}$ and $S_{modify}$, i.e., $S_{semantic} = \alpha*S_{preserve} + (1-\alpha)*S_{modify}$, the results are shown in Table2. Among all combination, $\alpha=0.5$ reaches the best human evaluation alignment, verifying that moderate scale between preservation and modification yield better evaluation result.

3.Distill our rules in MLLM to evaluate in a unified model for better impact: As suggested, we select most advanced closed-sourced MLLM GPT-4o and open-sourced MLLM qwen-vl-max-latest for our rule distillation. The results are shown in Table3, indicate (1) describe our design rules as prompt to guide GPT-4o via chain-of-thought ("GPT+rule", Line 2) improve human alignment of original GPT-4o(Line 1) by 2.7%, showing our rules, that separately evaluate the preservation and modification with explicit region decomposition, are essential for comprehensive evaluation and can assist to yield more trustworthy results. (2) Though our evaluation rules can improve GPT-4o human alignment, it still falls short of our tailored pipeline(3.0% performance drop, Line 2,5). This originates from that our pipeline endows LLM and segmentation tools for accurate editing region localization and explicitly evaluating preservation, modification separately, while MLLMs, even the most advanced GPT-4o, is sub-optimal for such localization task, as verified in our Fig.11,Supp. What' more, we further provide GPT-4o with our generated mask("GPT+rule+mask", Line 3) as condition for evaluation, but no significant performance improvement is observed, verifying that MLLM shows weaker evaluation interpretability than our pipeline and fail to fully utilize such accurate mask guidance. (3) Open-sourced MLLM Qwen(Line 4) lags behind in human alignment(12.4% performance drop compared to BPM). Besides, there is no large-scale ideal data for such role distillation training. The data should contain paired real and edited images along with human-annotated editing masks, and heavy human annotation on the quality judgment for modification, preservation, size and position change, so we don't do any learning like fine-tuning it. (4) Our rules witness consistent improvement in MLLMs evaluation, showing its wild application and potential for generalization towards more unseen tasks like more diverse image editing.