Thank you for your very detailed comments and suggestions. However, due to the word limit of 10000 words, we had to delete a lot of content to extremely compress the reply.

A1

On CS3 and DGF design choices. We agree that clearer motivation is needed. CS3 captures multi scale temporal and structural patterns in neural signals, consistent with findings that multi band EEG features improve intent decoding [1]. DGF performs selective multimodal fusion through dual gating, which follows prior successes of gating and normalization strategies in multimodal learning [2,3,4]. Evidence:

1. In a 35 class editing type classification pre study, CS3 plus DGF achieved the best F1 0.285 and mAP 0.411.
2. Replacing an MLP encoder with CS3 alone raised F1 from 24.9 percent to 28.1 percent.
3. DGF beat simple concatenation by 1.2 mAP points and cross attention by 2.2 mAP points, which is important under low SNR neural signals.
4. In downstream editing, adding CS3 plus DGF improved CLIP I to 63.19 percent while keeping CLIP T comparable to a simple MLP plus concatenation baseline. We will integrate justification, citations, and ablations into the revised paper.

[1] Altaheri, H., Muhammad, G., & Alsulaiman, M. (2022). Physics-informed attention temporal convolutional network for EEG-based motor imagery classification. IEEE transactions on industrial informatics, 19(2), 2249-2258.

[2] Choi, Y., Uh, Y., Yoo, J., & Ha, J. W. (2020). Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 8188-8197).

[3] Yin, D., Ren, X., Luo, C., Wang, Y., Xiong, Z., & Zeng, W. Retriever: Learning Content-Style Representation as a Token-Level Bipartite Graph. In International Conference on Learning Representations.

[4] Rahman, W., Hasan, M. K., Lee, S., Zadeh, A., Mao, C., Morency, L. P., & Hoque, E. (2020, July). Integrating multimodal information in large pretrained transformers. In Proceedings of the conference. Association for computational linguistics. Meeting (Vol. 2020, p. 2359).

---

A2

On pairing EEG plus PPG with T5, and fNIRS plus Motion with CLIP. Rationale is not arbitrary. T5 provides fine grained token level semantics that help precise instruction following, which complements the fast neural dynamics in EEG and the lightweight hemodynamics from PPG. CLIP provides robust global semantics that align with slower cortex wide fNIRS signals and intentional head Motion.

Evidence from the same classification pre study.

1. EEG plus PPG with T5 outperformed the CLIP variant by 3.0 F1 points and 3.6 mAP points.
2. fNIRS plus Motion with CLIP outperformed the T5 variant by 5.5 F1 points and 5.2 mAP points.
3. In editing, the aligned pairing yielded higher CLIP T 25.88 percent than a heterogeneous swap 24.96 percent. We will add this rationale and the supporting numbers.

---

A3

On the DiT backbone and conditioning. Backbone is FLUX.1 dev, a pretrained diffusion transformer with about 860 million parameters at 512 by 512 resolution. We integrate CS3 encoders and DGF to produce conditioning vectors that enter the DiT through cross attention, analogous to text prompts. Editing uses latent inversion of the input image and then applies neural or text conditioning during denoising. We will provide exact architecture blocks, parameter counts, and a conditioning diagram.

---

A4

On the missing citation for near text pretraining. We will add citations and clarify a two phase modular pretraining. Phase one pretrains CS3 encoders per modality using public datasets such as Thinking Out Loud for EEG [1] and an open fNIRS resource [2], combined with our internal corpora that are disjoint from evaluation. Phase two aligns paired encoders such as EEG plus PPG and fNIRS plus Motion to instruction embeddings with a symmetric NT Xent objective. We will list dataset names, sizes, and disjoint splits in the appendix.

[1] Nieto, N., Peterson, V., Rufiner, H. L., Kamienkowski, J. E., & Spies, R. (2022). Thinking out loud, an open-access EEG-based BCI dataset for inner speech recognition. Scientific data, 9(1), 52. [2] Ning, M., Duwadi, S., Yücel, M. A., Von Lühmann, A., Boas, D. A., & Sen, K. (2024). fNIRS dataset during complex scene analysis. Frontiers in Human Neuroscience, 18, 1329086.

---

A5

On pretraining and fine tuning details and stability. Pretraining used about 12 thousand steps across 2 epochs with LoRA based adaptation. Fine tuning used about 8 thousand steps with a conservative schedule and optional layer freezing. Early stopping triggered at about 19 thousand total steps. Loss dropped quickly in the first 2 thousand steps, showed two small transients at about steps 3,001 and 4,001, then stabilized without divergence. Full configs and loss curves will be added to the appendix.

---

A6

On flow versus diffusion objective and CFG details. We clarify that Eq. 10 describes a diffusion objective, not a flow objective. The model uses the DDPM loss adapted to a DiT backbone. Inference uses classifier free guidance with scale w = 3.5. Null conditioning is applied by zeroing modality embeddings: text only uses an empty string, neural only passes zero vectors from the CS3 encoder, and multimodal zeros both branches. This design remains fully compatible with standard CFG pipelines. We will update the text and include concise schematics of the conditioning path for reproducibility.

---

A7

On Fig. 6 readability. We will revise the layout to show category groups and sub types more clearly, add labels, and if needed split into two subfigures. A short how to read caption will accompany the figure.

---

A8

On metric choice and addition of LPIPS. We acknowledge that L1 and L2 do not capture perceptual similarity perfectly, yet they remain common in editing literature for spatial fidelity, for example Emu Edit and In Context Edit. We now also report LPIPS for perceptual quality. LoongX Neural Signals: LPIPS 0.1427. LoongX Signals plus Speech: LPIPS 0.1330. The updated table and confidence intervals will be included with references to LPIPS and recent editing works.

---

A9

On reporting trade off curves rather than single numbers. We added an interpolation analysis following Imagic and inspired by InstructPix2Pix and Imagic [1]. Let e_ori be the reconstruction anchor and e_tgt be the instruction anchor extracted from neural plus speech. Conditioning is e_bar equals eta times e_tgt plus one minus eta times e_ori with eta in [0, 1]. Over 150 held out samples, eta about 0.75 gives the best alignment fidelity balance. We will include CLIP Text, CLIP Image, and LPIPS curves and the summary table in the supplement.

[1] Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., ... & Irani, M. (2023). Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 6007-6017).

---

A10

On what DINO measures. DINO features capture fine grained structural similarity that correlates with perceived preservation of identity, pose, and local geometry. This complements CLIP Image which emphasizes global semantics. In our ablations, higher DINO aligned better with perceived success on shape and motion sensitive edits. We will add a brief definition and a pointer to recent evidence on DINO guided structure preservation.

---

A11

On OmniControl as the only text baseline and fairness of comparisons. We chose OmniControl for its adaptable multi conditional design and favorable speed quality trade off. To strengthen fairness, we now include ICEdit, Bagel, and Flux Kontext on the same data and also report LoongX integrated with these backbones. As seen in the A2\A3 Table of responsing Reviewer ZTwA, LoongX on Flux Kontext with neural signals plus speech reaches L1 0.1905, L2 0.0804, CLIP I 0.8078, CLIP T 0.273. We also report sizes and training FLOPs. OmniControl based LoongX uses about 12.5B parameters and about 3.04 times 10 to the 9 training FLOPs. Bagel based LoongX uses about 7.8B. Flux Kontext is about 12.6B. Gains come mainly from multimodal conditioning rather than brute force scaling. Full tables and configs will be added.

---

A12

On adding more textual editing baselines. We agree. We have added recent state of the art text editing methods and have also provided LoongX results when integrated into these SOTA backbones for a fair comparison.

---

A13

On demonstrating hands free image editing in practice. Our live protocol mirrors data collection. The main differences are buffering latency and attention variability, which we mitigate using temporal smoothing of neural embeddings and a lightweight confidence filter that drops weak signals. A real time demo is under finalization and will be released with the camera ready. In cross subject tests with 5 new users, live success reached about 92 percent of intra subject performance, which indicates strong transfer.

---

A14

On missing related work DreamConnect. We will add DreamConnect to related work and clarify differences in sensing. DreamConnect uses fMRI, which is less accessible for everyday interaction. Our focus is on practical EEG and fNIRS. The conceptual link on neurosemantic alignment is valuable, and we plan to explore similar alignment analyses in future work.

Dear Reviewer,

Thank you for your detailed and constructive feedback. For your convenience, we summarize our key responses below:

1. Architectural justification: We clarified the rationale for using CS3 and DGF modules, referencing both prior literature and our comprehensive ablation studies. Results show CS3+DGF outperforms baseline and alternative approaches in editing-type classification, supporting our design choices.
2. Modality pairing (EEG/PPG–T5, fNIRS/Motion–CLIP): These pairings are guided by theoretical reasoning and preliminary classification results, which demonstrate that our matching strategy yields better discriminability and downstream editing performance than alternatives.
3. Model and conditioning details: The editing model is based on the pretrained FLUX.1-dev DiT backbone (~12B params, 512×512), conditioned via our proposed CS3 and DGF modules. Input image editing uses DDPM inversion for latent alignment.
4. Pretraining citations: We have added all missing references for pretraining datasets and clarified our two-phase (public and proprietary) neural-text pretraining strategy.
5. Training details and stability: The model was trained for ~22k steps over ~2 epochs with LoRA. Training was numerically stable, with loss curves and config details to be included in the appendix.
6. CFG/conditioning setup: We use standard diffusion objectives with classifier-free guidance (CFG). Null-conditioning is implemented by zeroing embeddings.
7. Visualization and figure clarity: We will revise confusing figures and improve layout/readability in the final version.
8. Evaluation metrics: While L1/L2 remain standard for editing literature, we also report LPIPS, CLIP-I, CLIP-T, and DINO (for fine-grained feature similarity), ensuring a comprehensive evaluation.
9. Trade-off analysis: Following Imagic/InstructPix2Pix, we report trade-off curves between fidelity and semantic alignment, varying conditioning strength.
10. Baselines and fair comparison: We expanded comparisons to include recent state-of-the-art text-based methods (ICEdit, Bagel, Flux-Kontext) and reported model sizes/FLOPs for fair benchmarking.
11. Hands-free usability: Our pipeline closely matches practical usage. Additional cross-subject results are provided and a real-time demo is being developed.
12. Missing related work: DreamConnect and other latest works are now discussed and cited in the revised related work section.

We hope this summary addresses your main concerns and provides clarity on our key improvements.

Looking forward to your feedback!

Q1 & W1

The method relies on strong assumptions about the correspondence between neural signals and editing intent, without sufficient validation. During data collection, participants merely read editing instructions aloud, assuming they simultaneously imagine the edits. This unverified reliance on internal mental imagery raises concerns about whether the captured signals truly reflect editing semantics. Moreover, EEG and fNIRS signals may be dominated by language processing and speech-related activity, potentially confounding the neural representation of editing intent. No control experiments (e.g., reading unrelated text or silent imagination) are included to isolate editing-specific neural patterns.

A1

To ensure that the recorded neural signals truly reflected editing-specific mental states rather than mere language processing, we provided explicit instructions and guided training sessions to familiarize participants with the process of vivid mental imagery. Specifically, participants were encouraged to not only read the prompts aloud but to actively and vividly simulate the visual editing actions in their mind’s eye. This imaginative engagement was reinforced through a brief practice session prior to recording, helping establish a consistent attention and cognitive strategy across subjects and ensuring that the brain activity captured by the neural signals corresponded only to intensive editing intentions.

Moreover, as presented in Table 4 in Section A Technical Appendix, we tracked participants’ attention levels using alpha-to-theta EEG ratios, which is an established neural index of attentional engagement, and found that most participants maintained moderate to high attention levels throughout the task. These attention scores serve as an additional quality control signal, indirectly validating that participants were cognitively invested in the task, beyond passive reading.

Additionally, performed an unseen silent EEG experiment with new participants (N=2). The results show that the silent EEG from non-imagined read-only conditions had significantly lower performance (0.42 CLIP-I) than the imagined condition (0.63 CLIP-I), supporting the correspondence between neural signals and editing intent.

Q2 & W2

The claim that neural signals outperform text prompts in intuitive visual tasks (e.g., background editing) is contradicted by qualitative results. In Figure 7 (row 1), the Neural and Neural+Speech methods fail to execute the core instruction “place the cat above it,” while the Text-based method achieves the intended spatial manipulation. This inconsistency suggests that neural signals may struggle with structural understanding.

A2

Regarding the observation in Figure 7 (row 1), it is important to note that while the neural signals and neural+speech methods performed better than the text-based method in several tasks, the text-based method did better in spatial manipulation tasks like “place the cat above.” However, our method excels in achieving the object manipulation like let the cat in the figure to "looking down", while text-based method failed to achieve that.

Our neural signal-based method can indeed surpass the text-based method in several scenarios, e.g., the object editing task showed significant improvement with CLIP-I scores of 0.6605 compared to 0.6350 for the text-based method.

We will further clarify in the revised manuscript that the neural signals are better suited for low-level visual edits and that neural signal-based and text-based methods can offer complementary strengths, not necessarily one being superior across all tasks.

Q3 & W3

The approach depends on specialized hardware (e.g., fNIRS, PPG, EEG), limiting its practicality and scalability in real-world settings.

A3

While our current approach utilizes specialized equipment such as fNIRS, PPG, and EEG, we strongly believe that the rapid evolution of wearable brain-computer interface (BCI) technologies will soon make such devices as ubiquitous and user-friendly as today’s wearable smart watches. In the near future, lightweight and affordable neural headbands are likely to become part of everyday life, enabling not only health monitoring but also seamless interaction with digital systems. In this context, basic applications like neural-driven image editing will no longer be constrained by hardware limitations. Enhancing accessibility has always been a central motivation of our research, and we envision our system evolving hand-in-hand with advances in wearable BCI to deliver practical, real-world value.

Q4

The dataset includes only 12 participants, limiting generalizability. Broader evaluation across diverse populations (e.g., users with disabilities, different ages, or cultural backgrounds) is necessary to assess robustness and applicability.

A4

We appreciate the reviewer’s concern regarding generalizability. While our current dataset consists of 12 primary participants, this sample size is comparable to or even larger than several published works in neural signal-driven generation. For instance, EEG2Video [Li et al., 2023] utilized data from only 6 participants, and DreamDiffusion [Liu et al., 2024] involved 10 participants, yet both demonstrated compelling results.

To further assess robustness, we conducted cross-subject experiments involving 7 new participants who were not included in model training. Notably, this group includes individuals from diverse age ranges (a middle school student and an elderly participant aged 67), as well as one participant with a physical disability (missing fingers on the dominant hand). Despite these variations, our model maintained 60.49% CLIP-I and 44.47% DINO scores of its intrasubject performance, demonstrating encouraging generalization across populations. More detailed information about all test data can be found in response to Reviewer 1, Question 1.

Table: Subject-Level Metrics – Unseen without Speech (Silence Experiment)

Table: Evaluation Metrics – Silence Subjects

Q1

Dataset bias problem. The authors collected paired dataset using 12 subjects, and set training and test setup. Are the subjects of training set data and test set data independent? If the training and test data set share same subjects, then it is intra-subject experiment which lacks generalizability. Please elaborate the dataset collection protocol.

A1

1. Test data

We are deeply grateful for your insightful and highly constructive feedback, which has significantly enhanced the completeness and scientific rigor of our paper. Regarding the dataset bias issue, our original dataset was collected from 12 participants (6 female, 6 male, mean age 24.5 ± 2.5 years), each contributing around 2,000 paired samples under carefully controlled experimental conditions (see Appendix A for detailed protocol). While our initial split ensured training/test separation, we acknowledge that the possibility of subject overlap could limit generalizability. To address this, we performed additional cross-subject evaluations with 5 new participants (3 male, 2 female, ages 13–63). The results confirm that the model maintains strong generalization when applied to unseen individuals, with performance trends on CLIP-I, DINO, and CLIP-T remaining consistent with those from the original test set. This provides evidence that our approach is not overly reliant on subject-specific neural signatures, but instead captures transferable semantic representations.

Table 1. Performance comparison on the original test set with 12 subjects and the unseen test set with 5 new subjects.

The basic background information and sample size of the participants, as well as the specific individual performance based on LoongX (Signals+Speech), are as follows:

Table 2. Basic information and performance of all subjects using LoongX (Signals+Speech).

2. Dataset collection protocol

Dataset collection protocol details have been supplemented into Section C Supplementary Dataset Details in the supplementary PDF file (Main text 3.1; A.2).

Q2

Usage of speech signal. In the method, the conditional model takes speech signal along with neural signals. Also in the results, output from only neural signal shows significantly low performance compared to neural+speech. It seems that speech signal plays major role in editing. Please elaborate this part and why neural signal only setup does not work.

A2

We appreciate the reviewer’s attention to the role of speech signals in our method. In our framework, speech serves as a complementary modality rather than the dominant one, supplying explicit high-level linguistic cues often absent from neural signals alone. EEG, fNIRS, and PPG effectively capture intent with fine temporal resolution, but they encode semantics implicitly and intrinsically. Compared with these neural signals, Speech supplements auxiliary semantic information for more abstract tasks or specific text editing tasks. For example, a short spoken instruction (e.g., “make the sky look like heaven”) provides a low-entropy prior that helps disambiguate editing intent, while neural signals control the global style and fine-grained visual effects. Importantly, our experiments show that neural signals alone yield robust and competitive performance, especially for low-level visual edits such as global style adjustments. For instance, the neural-only setup achieves a CLIP-I score of 0.6605, which is slightly higher than the 0.6374 observed when speech is included.

However, when it comes to tasks requiring more nuanced semantic comprehension, combining speech with neural signals leads to better alignment with the intended instruction, as indicated by an increase in the CLIP-T score from 0.2436 (neural-only) to 0.2588 (neural+speech), a relative improvement of about 0.015. This indicates that each modality brings unique strengths: neural signals excel at pixel-level control and global editing, while speech boosts complex semantic alignment. Overall, the gap between the two settings is moderate rather than substantial, suggesting that neural signals are already effective, but can be further enhanced by incorporating speech. We believe this underscores the complementary role of speech in our system, and we appreciate the opportunity to clarify this point.

Table 1: Automatic Evaluation of Editing Methods

We appreciate the reviewer’s careful assessment and the opportunity to clarify these critical aspects. With these additional results and protocol details, we believe the work now provides a more complete and reliable account of both methodological soundness and empirical validity. We hope these clarifications help the reviewer better appreciate the strength and significance of the proposed work.

Minor Q3

During the data. collection, are the authors properly taken the approval from ethics board?

A3

Yes, we have obtained ethical approval from the relevant institutional ethics review board before conducting the experiments, as stated in the manuscript and supplementary materials. All participants were informed of the procedure and gave their consent before participation. We will attach the specific institutional information to the camera-ready version.

A1

Thank you for your insightful question. The Diffusion Transformer (DiT) used in our experiments is based on the Flux.1-Dev pretrained model, and fine-tuned on our dataset using LoRA (Low-Rank Adaptation). We trained it for 12,000 steps over 2 epochs, using an Adam optimizer with a learning rate of 1e-4. The number of LoRA layers is 10, with a total of 28M parameters. We will add more detailed information on the model's architecture and the fine-tuning strategy in the revised manuscript.

A2\A3

We expanded our evaluation to include state-of-the-art text-based editing methods, with a particular focus on Flux-Kontext, which represents one of the strongest open-sourced text-driven editing frameworks at present. This addition provides a clearer perspective on where our multimodal approach stands relative to established text-only paradigms.

As the results in Table 1 demonstrate, text-only Flux-Kontext indeed performs strongly in semantic alignment (CLIP-T = 0.2728) and overall visual fidelity. However, our LoongX framework integrated into the same Flux-Kontext backbone (fine-tuned on L-Mind) further elevates performance across nearly all metrics, achieving the highest CLIP-I (0.8078) and CLIP-T (0.2730).

A4

Thank you for your constructive advice. We have conducted a systematic user study to comprehensively evaluate both Editability (how well the system followed the user's intention) and Content Preservation (how well the original content of the image was preserved after editing), as recommended. We recruited 10 annotators (5 male, 5 female, ages 18-40), each of whom independently rated the edited images generated under different conditioning methods (see Table 2 and Table 3). Methods included both baselines (OmniControl, Flux-Kontext) and our proposed LoongX framework under different conditions. The scoring was performed in a double-blind manner and on a 0–5 scale, where 0 indicates "completely unacceptable" and 5 indicates "excellent".

The combination of neural signals and speech (Editability: 4.63, Content Preservation: 4.40) outperforms either modality alone. While text inputs offer high editability (4.55) due to their explicit semantic guidance, neural signals exhibit superior content preservation (4.28) by capturing implicit visual preferences. By integrating both, our method achieves a better balance between user intention and visual fidelity.

A5

We have expanded the explanation of our evaluation setup and the editing task categories, and we have also added further discussion of failure cases, as detailed in Section A.5 of the supplementary material.

Our evaluation framework consists of four representative categories of image editing tasks, as visualized in Figures 8–Figure 11 in the supplementary material:

1. Global Editing – e.g., changing the overall color tone, saturation, or lighting of the entire image.
2. Background Editing – e.g., modifying or replacing background elements while keeping the subject unchanged.
3. Object Editing – e.g., altering specific foreground elements like clothing color, object size, or replacing an item.
4. Text Editing – e.g., modifying embedded textual content within images.

Each category is evaluated both quantitatively and qualitatively to assess alignment with user intent and perceptual fidelity.

In addition, we have carefully analyzed and reported failure cases in the supplementary material. These cases highlight limitations such as:

(1) Highly imaginative prompts that lie far outside the training distribution (e.g., “long-legged space creature”),

(2) Ambiguous instructions lacking critical semantic detail (e.g., whether the original background should be retained), and

(3) Non-standard input formats (e.g., panoramic aspect ratios) that can challenge spatial reasoning.

These failure cases highlight current limitations in generalization and input robustness, and we view them as opportunities for future improvement.

A6

Our diffusion transformer is not trained from scratch; instead, we initialize it using the publicly available FLUX.1-dev pretrained weights from Black Forest Labs. This model adopts a U-Net-style DiT architecture with transformer blocks applied at each resolution scale. The full model consists of approximately 860 million parameters, operating at 512×512 resolution, and is pretrained on high-quality text-to-image datasets. Leveraging this foundation allows us to benefit from its robust visual representation capability while focusing our training on the unique neural-conditioning aspects of our task.

To adapt FLUX to our multimodal setting, we employ LoRA (Low-Rank Adaptation) for parameter-efficient fine-tuning. This enables us to fine-tune the model on our neuro-conditioned dataset without the need for full backpropagation through all layers, preserving generalization while reducing computational cost. Our conditioning pipeline feeds into the FLUX transformer via our proposed DGF fusion module, which supports both text embeddings (via T5-XXL and CLIP) and neural encodings (via CS3 encoder), as described in Section A.4 of the supplementary material.

A7

We appreciate the reviewer’s suggestion and agree that recognizing the limitations of our method is essential for a balanced evaluation. In the supplementary material (Section A.5: More Failure Cases), we expand on failure cases with qualitative examples in Figure 12.

Our analysis highlights three main scenarios where limitations emerge:

1. Abstract or highly imaginative instructions: For concepts far beyond the training distribution (e.g., “morph the dog into a long-legged space creature”), outputs become semantically inconsistent or implausible due to the lack of concrete grounding.
2. Vague or under-specified prompts: Ambiguity in instructions (Figure 12b) can cause mismatches with user intent, such as uncertainty over whether to retain or replace a background. This underscores the need for improved intent disambiguation via feedback or uncertainty modeling.
3. Non-standard input formats: With unusual aspect ratios (Figure 12c), spatial misalignment or degraded quality may occur, likely due to training data bias. Future work could address this with augmentation or scale-invariant architectures.

A8

The quantitative evaluation covered tasks including global texture changes, object removal, and background modifications. We report the following results:

• Global texture editing: CLIP-I: 0.6605
• Object removal: DINO: 0.4812
• Text edits: CLIP-T: 0.2588

A9

We acknowledge the issues raised regarding the color preservation in Figure 7, where certain edits, like "place the cat above," were not fully captured by our model. We have provided an extended limitations discussion to include more examples like this, where certain fine-grained details or semantic richness were not fully preserved. Additionally, we are addressing the issues causing these artifacts, which are largely due to the backbone model's capacity and the inherent challenges in handling complex image semantics.

A10

As confirmed, the submission of supplementary material alongside the main paper was in accordance with the conference's submission guidelines. Thank you for your detailed review.

Dear Reviewer,

We are deeply grateful for your encouraging words and constructive suggestions. Your positive assessment of our work and recognition of its impact on multimodal editing mean a great deal to us.

We sincerely appreciate your acknowledgment of the additional experiments and clarifications, and as advised, we will include extended results, detailed setups, and qualitative comparisons with different backbones in the final version.

Thank you again for your thoughtful engagement and kind support!

Warm Regards,

The Authors