Thank you for the insightful comment. We provide our responses as follows.

Weakness 1 & Question 1

We argue that our work is not merely engineering-driven, but offers profound insights into style-aligned generation and other attention sharing approaches within the MM-DiT framework.

Vanilla attention sharing methods employ identical position embedding for reference and target images. However, this practice in MM-DiT leads to content leakage and degraded text controllability, a fundamental challenge affecting all attention sharing methods in this architecture. To the best of our knowledge, our work is the first to address this issue by introducing ShiftPE and the SSA component, providing general insights for future attention sharing approaches.

Furthermore, our method is supported by solid data analysis. As shown in Figure 4 of the paper (Default RoPE part), the attention heatmap reveals that using identical position embedding for reference and target images causes the model to attend only to spatially aligned regions. ShiftPE, which assigns non-overlapping shifted position embedding to reference and target images, effectively decouples their spatial attention. This is validated in Figure 4 (ShiftPE part), where the attention becomes more appropriate and disentangled. Moreover, ablation studies and visualizations in Table 2 and Figure 7 demonstrate the effectiveness of our approach. We plan to conduct deeper theoretical analysis in future work.

Weakness 2 & Question 2

Although our method is simple and effective, we did not claim that it is lightweight or efficient in the paper.

Below are the actual measured inference latency and GPU memory usage: "Original" denotes Flux generating two images directly, while "Ours" denotes our method generating both the reference and target images simultaneously (also two images in total), the experiment is conducted on a single H20 using bf16 precision:

As shown in the table, both the inference latency and GPU memory overhead remain within reasonable bounds. More importantly, our approach requires no training, thus naturally avoiding the expensive training process.

Weakness 3 & Question 3

For [1], we conducted a detailed comparison using the official code and recommended configurations. The experimental results are as follows:

As can be observed, [1] barely performs meaningful image editing and fails to make substantial content changes given a reference image, resulting in a lower $S_{text}$ and higher $S_{sty}$ and $S_{dino}$ values. The reasonable range for $S_{sty}$ is 0.4–0.8 and for $S_{dino}$ is 0.3–0.6; values beyond these ranges indicate excessive similarity between the target and reference images, suggesting content leakage, which is further confirmed by the low $S_{text}$.

[2] is a subject-driven image editing method that requires both a scene image and an object image as input, and thus is not applicable to style-aligned image generation tasks.

We will include the results of [1] to the Table 4, discuss and cite [2] in the related work section (Section 2) in the revised manuscript.

[1] Unveil Inversion and Invariance in Flow Transformer for Versatile Image Editing.

[2] PS-Diffusion: Photorealistic Subject-Driven Image Editing with Disentangled Control and Attention.

We thank the reviewer for the constructive comments. We provide our responses as follows. We have summarized the reviewer's weaknesses and questions into four aspects, which we will address in turn.

Weakness 1

We thank the reviewer for the valuable feedback. We agree that Figure 3, which illustrates the overall pipeline of our method, currently lacks sufficient integration with the main text. To address this, in the revised manuscript, we will add:

• A dedicated reference to Figure 3 in Section 3.3
• incorporate explicit descriptions and discussions of the figure within the ShiftPE (lines 147–148) and SSA (lines 155–156) subsections.

These revisions will strengthen the connection between the figure and the methodological narrative, thereby improving clarity and readability.

Weakness 2

We thank the reviewer for the insightful comments. Due to an oversight on our part, We did not clearly articulate the motivations and theoretical analysis of the method. Below, we clarify the design rationale for each component:

1. Motivation of ShiftPE: Vanilla attention sharing methods employ identical position embedding for reference and target images. However, this practice in MM-DiT leads to content leakage and degraded text controllability, a fundamental challenge affecting all attention sharing methods in this architecture. As shown in Figure 4 of the paper (Default RoPE part), the attention heatmap reveals that using identical position embedding causes the model to attend only to spatially aligned regions. To mitigate this, we propose ShiftPE, which assigns non-overlapping shifted position embedding to reference and target images, effectively decoupling their spatial attention. This is validated in Figure 4 (ShiftPE part), where attention becomes more appropriate and disentangled. Moreover, ablation studies and visualizations in Table 2 and Figure 7 demonstrate the effectiveness of our approach.

2. Motivation of SSA: In standard MM-Attention, Q, K and V are derived from image and text tokens. However, text tokens do not carry style information, and computing their projections for attention introduces redundant computation. To improve efficiency and focus on image-relevant features, SSA shares only the Q, K and V derived from image tokens, eliminating unnecessary computation while enhancing style consistency.

3. Motivation of Flexible Strength Control: Practical applications often require fine-grained control over style consistency. To this end, we introduce a scaling factor $\lambda$ that modulates the strength of style consistency, enabling flexible and user-controllable style-aligned image generation without retraining.

We will include these explanations in the revised manuscript to improve clarity.

Weaknesses 3

We agree that discussing limitations in the main text enhances transparency and readability. In the revised manuscript, we will move the limitation section from the auxiliary material to the main body of the paper, placing it at the end of Section 5 (Conclusion) for better visibility.

Weaknesses 4 & Question 1

Our method allows users to provide reference image as style input for generation.

Our method inherently requires the Q, K and V features from the reference image during the generation process to produce style consistent images. When the reference image is generated by Flux, we naturally have access to its intermediate Q, K and V. For a user-specified reference image, we obtain approximate cached Q, K and V by applying noise at different timesteps $t$ and feeding the noisy latent into the diffusion model.

The algorithm proceeds as follows:

Algorithm for user-provided reference image
Input: user-provided reference image I, num of inference timesteps T
Output: Cached Q, K, V of I
Step1 --> Initialize a random noise: noise ~ N(0, 1)
Step2 --> latent = vae_encode(I)
Step3 --> cached_qkv = {}
Step4 --> for t = T, T-1, ..., 0 (T timesteps) do:
              noise_input = t * noise + (1 - t) * latent
              Q_t, K_t, V_t = Flux(noise_input, t, others(e.g., prompt embeds, guidance scale, ...))
              cached_qkv[t] = (Q_t, K_t, V_t)
Step5 --> return cached_qkv        

The above algorithm extends our method to support user-provided images as style references. Moreover, our approach demonstrates strong performance. We conduct an evaluation on StyleCrafter's test set (20 distinct style images and 20 prompts, forming 400 test cases) and compare with StyleCrafter, StyleShot, and CSGO. The results are summarized in the table below:

Experimental results show that our method significantly outperforms existing approaches in both text controllability ($S_{{text}}$) and style consistency ($S_{{sty}}$, $S_{{dino}}$). Moreover, the compared methods require large-scale training datasets and model training, and our method is training-free, demonstrating clear practical advantages.

We would like to thank you again for your thoughtful comments and valuable suggestions.

With the discussion period approaching its close, if there are any further points you would like us to clarify or address, we would be happy to provide additional explanations.

We sincerely appreciate the time and effort you have dedicated to evaluating our work.

Thank you for the insightful comment. We provide our responses as follows.

Weakness 1

We agree that analyzing failure cases is crucial for deeper understanding and improvement. Due to limitations in this rebuttal (can't upload pdf file), we cannot show visual examples, but we observe that failures often involve content leakage—e.g., when the reference image is "A bowl with apples" and the target is "A pencil", the generated image may incorrectly include the bowl and apples alongside the pencil, indicating content leakage problem. This insight highlights the need for better disentanglement in future work.

Regarding the reviewer’s example—target prompt: a bike on top of ocean, with ocean from the reference image—we have tested this scenario and found that AlignGen produces visually coherent results: a person riding a bicycle over the ocean, with the water texture and style well aligned to the reference. This demonstrates effective alignment under semantically challenging conditions.

To further evaluate robustness and uncover underlying mechanisms, we conducted more comprehensive testing using three challenging prompt types generated by ChatGPT (20 cases per type, 4 prompts per case):

• Long and Complex Prompts: Intricate scenes with multiple objects, characters, and actions (e.g., A high-altitude weather control station on a cliff with turbines, mid-air holograms, and engineers in pressure suits calibrating conduits during a storm, in low-poly style).
• Ambiguous Style: Vague stylistic descriptions (e.g., with a hazy echo) instead of explicit style labels (e.g., in oil painting style).
• Conflict Prompts: Semantic mismatches between reference image and target image (e.g., reference: bicycle in cloud library; target: fish walking in city).

Quantitative results are summarized below, where Normal Test Dataset refers to the test set used in the paper:

For Ambiguous Style prompts, compared to the Normal Test Dataset, style consistency decreases slightly , while $S_{text}$ drops more noticeably—due to ambiguous style descriptions also impairing content clarity, leading to less coherent and focused outputs. For Long and Complex and Conflict Prompts, compared to the Normal Test Dataset, style consistency is somewhat reduced, yet performance remains strong, thanks to Flux’s robust generation capacity in handling complex or conflicting inputs.

Furthermore, we will conduct a more in-depth theoretical analysis based on these results in future work to explore the underlying hidden mechanisms.

Weakness 2

We thank the reviewer for the suggestion.

We propose Flexible Strength Control in Section 3.4, which introduces a parameter $\lambda$ to achieve a trade-off between prompt alignment and style fidelity. While we recommend setting $\lambda = 1.1$ in the paper for good performance, in certain extreme cases, manual tuning of $\lambda$ is still required to achieve optimal results. Developing an adaptive mechanism to automatically balance prompt alignment and style fidelity would indeed be highly desirable.

In the future, we will actively explore an automated mechanism to balance prompt alignment and style fidelity.

We thank the reviewer for the constructive comments. We provide our responses as follows. We have summarized the reviewer's weaknesses and questions into four aspects, which we will address in turn.

Weakness 1

We argue that our method is not a simple incremental modification, but offers distinct academic value and deep insights into attention sharing strategies and the MM-DiT architecture.

Vanilla attention sharing methods employ identical position embedding for reference and target images. However, this practice in MM-DiT leads to content leakage and degraded text controllability, a fundamental challenge affecting all attention sharing methods in this architecture. To the best of our knowledge, our work is the first to address this issue by introducing ShiftPE and the SSA component, providing general insights for future attention sharing approaches.

Moreover, beyond ShiftPE and SSA mentioned in Weakness 1, Flexible Strength Control (Section 3.4) and Delicate Attention Replacement (Section 3.5) also contribute significantly to the final performance, as shown in Table 2 and Figure 8 in paper. These contributions should not be overlooked.

Weakness 2 & Question 1

We constructed three types of challenging prompts using ChatGPT for more comprehensive testing—Long and Complex, Ambiguous Style, and Conflict Prompts—with 20 cases per type (4 prompts per case). The design principles are as follows:

• Long and Complex Prompts: Intricate scenes with multiple objects, characters, and actions (e.g., A high-altitude weather control station on a cliff with turbines, mid-air holograms, and engineers in pressure suits calibrating conduits during a storm, in low-poly style).
• Ambiguous Style: Vague stylistic descriptions (e.g., with a hazy echo) instead of explicit style labels (e.g., in oil painting style).
• Conflict Prompts: Semantic mismatches between reference image and target image (e.g., reference: bicycle in cloud library; target: fish walking in city).

Quantitative results are summarized below, where Normal Test Dataset refers to the test set used in the paper:

For Ambiguous Style prompts, compared to the Normal Test Dataset, style consistency decreases slightly , while $S_{text}$ drops more noticeably—due to ambiguous style descriptions also impairing content clarity, leading to less coherent and focused outputs. For Long and Complex and Conflict Prompts, compared to the Normal Test Dataset, style consistency is somewhat reduced, yet performance remains strong, thanks to Flux’s robust generation capacity in handling complex or conflicting inputs.

Weakness 3 & Question 2

Our method allows users to provide reference image as style input for generation.

Our method inherently requires the Q, K and V features from the reference image during the generation process to produce style consistent images. When the reference image is generated by Flux, we naturally have access to its intermediate Q, K and V. For a user-specified reference image, we obtain approximate cached Q, K and V by applying noise at different timesteps $t$ and feeding the noisy latent into the diffusion model.

The algorithm proceeds as follows:

Algorithm for user-provided reference image
Input: user-provided reference image I, num of inference timesteps T
Output: Cached Q, K, V of I
Step1 --> Initialize a random noise: noise ~ N(0, 1)
Step2 --> latent = vae_encode(I)
Step3 --> cached_qkv = {}
Step4 --> for t = T, T-1, ..., 0 (T timesteps) do:
              noise_input = t * noise + (1 - t) * latent
              Q_t, K_t, V_t = Flux(noise_input, t, others(e.g., prompt embeds, guidance scale, ...))
              cached_qkv[t] = (Q_t, K_t, V_t)
Step5 --> return cached_qkv        

The above algorithm extends our method to support user-provided images as style references. Moreover, our approach demonstrates strong performance. We conduct an evaluation on StyleCrafter's test set (20 distinct style images and 20 prompts, forming 400 test cases) and compare with StyleCrafter, StyleShot, and CSGO. The results are summarized in the table below:

Experimental results show that our method significantly outperforms existing approaches in both text controllability ($S_{text}$) and style consistency ($S_{sty}$, $S_{dino}$). Moreover, the compared methods require large-scale training datasets and model training, and our method is training-free, demonstrating clear practical advantages.

Weakness 4 & Question 3

We clarify that $S_{{sty}}$ and $S_{{dino}}$ are not strictly "higher is better" for style similarity evaluation, but are meaningful only within reasonable ranges—typically $S_{{sty}} \in [0.4, 0.8]$ and $S_{{dino}} \in [0.3, 0.6]$ in our task. Within these ranges, higher values could indicate more consistent stylization. Values beyond this range do not effectively reflect improved style similarity: e.g., scores approaching 1 indicate near-identical content, suggesting content leakage rather than enhanced stylization.

This is because neither metric is specifically designed for style assessment:

• $S_{{sty}}$ measures semantic similarity and may yield high scores when the semantic content of two images are very similar, even under significant style differences (e.g., photo vs. cartoon).
• $S_{{dino}}$, while more sensitive to texture and style due to self-supervised training, still captures structural and semantic information, not pure style.

In a word, although neither metric is purely designed for style similarity evaluation, both can effectively assess style consistency between images to a certain extent when used within their appropriate ranges.