To Reviewer 7rxu

Thanks very much for your valuable comments that helped us significantly improve this submission. Below are detailed point-to-point responses about how we address your questions.

---

[W1]: Question on Computational Efficiency

Response:Thanks, we kindly suggest you to refer to [Q1] from Reviewer hGay, to facilitate your reading, we made up the inference time and VRAM consumption of our model, comparing it with other baseline models, along with the detailed explanations below:

• Analysis of computational efficiency

"Our inference time is comparable to BrushNet [3] and faster than recent methods such as HDP [4] and PP [2]. Moreover, our model consumes less VRAM than other text-guided inpainting models, owing to the fact that our method performs low-and-mid frequency substitution across different branches after each denoising step. The basic idea is to be decoupled from the resource-intensive denoising process itself, it incurs only negligible additional memory cost than the single denosing process BLD [1]."

---

[W2]: Results on More T2I Models

Response:Thanks for the valuable comments. We kindly clarify that we adopt Stable Diffusion v1.5 (SD v1.5) since most existing text-guided image inpainting methods fine-tune their models based on SD v1.5. To ensure a fair comparison, we mainly present the inpainted results based on SD v1.5 in the mainbody.

Following your suggestion, we additionally conducted the inpainted results of other representative text-to-image (T2I) models, including the DiT-based method HunyuanDiT and Stable Diffusion XL (SDXL), which is an upgraded version of SD v1.5. We report the quantitative results on BrushBench, covering both inside and outside inpainting tasks, along with detailed explanations, as shown below:

• Analysis of the improvements achieved on powerful T2I models

"We adopt the IR and HPS to evaluate the image generation quality, and use the CLIP score to assess the text-image alignment, which is in line with Lines 270–277, Sec. 3.1 of the mainbody. The results show that both HunyuanDiT and SDXL outperform SD v1.5 on BrushBench. This improvement can be attributed to their significantly larger parameter scales and training on more diverse and extensive datasets. As a result, in the latent space, the features of the unmasked regions retain more information after downsampling. This enables our frequency-aware null-text-null diffusion framework to exchange richer mid-and-low frequency components during the early stage of the denoising process, which is consistent to the motivation in Sec. 2.3 of the mainbody."

---

[W3]: Writing and Presentation

Response: Thanks for the comments, we will fix all your suggestions regarding writings issues, such as long sentences, small figure and typos etc.

---

References

• [1] Avrahami O, Fried O, Lischinski D. Blended latent diffusion[J]. ACM transactions on graphics (TOG), 2023, 42(4): 1-11.

• [2] Zhuang J, Zeng Y, Liu W, et al. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 195-211.

• [3] Ju X, Liu X, Wang X, et al. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 150-168.

• [4] Manukyan H, Sargsyan A, Atanyan B, et al. Hd-painter: high-resolution and prompt-faithful text-guided image inpainting with diffusion models[C]//The Thirteenth International Conference on Learning Representations. 2025.

Dear Reviewer 7rxu:

We are sincerely grateful to the updated rating as you raised up and positive confirmation to our work, while largely benefited from your intuitive comments. Many thanks !

To Reviewer zpgQ

Thanks very much for your valuable comments that helped us significantly improve this submission. Below are detailed point-to-point responses about how we address your questions.

---

[Q1]: Discussion on the High-Frequency Band

Response: Thanks. As elaborated in Remark 1 (Lines 202–207, Sec. 2.3.2) of the mainbody, to ensure consistency between masked and unmasked regions, we introduce a concurrent null-text denoising process guided by the previously denoised mid-frequency components. This process helps to denoise the low-frequency band across both masked and unmasked regions. Meanwhile, the high-frequency band is also addressed in the final null-text denoising step of the early stage.

Following your suggestions, we further conduct the ablation studies on exploiting the denoised high-frequency result from the second text-guided denoising process, to substitute the denoised high-frequency band from the last null-text denoising process, along with the detailed explanations, as shown below:

• Analysis of the impact of substituting the denoised high-frequency components

"The results show that $Case_{HF}$ exhibits a substantial performance degradation (at most 23.87% for IR), confirming that the high-frequency band is inherently sparser and more susceptible to being overwhelmed by high-level noise during early denoising, which significantly limits its effectiveness in guiding low-frequency reconstruction, which is in line with footnote in Sec. 1, as well as the Line 53 - 61 of mainbody."

• We experimentally validate the denoising process as per several metrics to measure changes in the low-and-high frequency bands at different timesteps, we adopt two groups of metrics to analyze the frequency components of the denoising process below:

• Analysis of the high-and-low frequency in the denoising process

The results show that low-frequency metrics (Mean Gray, HSV-V, Luma) remain nearly unchanged from the beginning, indicating early stabilization. In contrast, high-frequency metrics (Avg Gradient, Variance, LBP Variance) increase steadily in the late timesteps, confirming that high-frequency details are refined gradually during denoising, which is in line with Lines 54–57, Sec. 1 of the mainbody."

---

[Q2]: Discussion on Adaptively Extracting the Low-and-Mid Frequency Bands

Response: Thanks for your valuable comments. We kindly remind that we truly considered to adaptively extract the low-and-mid frequency bands from masked images based on Eq.(7) and Eq.(10) of the mainbody, we copy that as follow:

• Line 197 - 198, Eq.(7) of the mainbody

$$th_{lp} = λ_{lp}^{f} + λ^{r}_{lp} \cdot \frac{\|M\|_1}{HW}$$

where $λ_{lp}^{f}$ and $λ_{lp}^{r}$ are hyper-parameters. We set $th_{lp}$ to adaptively correlate with the ratio of unmasked regions, since the larger unmasked regions require more low-frequency band from the null-text denoising process to substitute the low-frequency band within the text-guided denoising process.

• Line 219 - 220, Eq.(10) of the mainbody

$$th_{mp1} = λ_{mp1}^{f} - λ_{mp1}^{r} \cdot \frac{\|M\|_1}{HW}$$ $$th_{mp2} = λ_{mp2}^{f} + λ_{mp2}^{r} \cdot \frac{\|M\|_1}{HW}$$

where $\lambda_{mp1}^{f}$, $\lambda_{mp1}^{r}$, $\lambda_{mp2}^{f}$, and $\lambda_{mp2}^{r}$ are hyper-parameters. We also set $th_{mp1}$ and $th_{mp2}$ to correlate with the ratio of unmasked regions. This is because larger nmasked regions require more mid-frequency band to encode the information that originates from the low-frequency band substituted during the first null-text denoising process.

We provide further ablation studies on the impact of adaptively extracting low-and-mid frequency bands in the Line 70 - 89, Sec.C of the Appendix, i.e., Table. 2, we copy that below:

• Table. 2 of the Appendix

• Analysis of adaptively extracting the low-and-mid frequency bands

"As mentioned in Sec.2.3.2 and Sec.2.3.3 of the mainbody, we set $th_{lp}$ in Eq.(7), and $th_{mp1}$ and $th_{mp2}$ in Eq.(10) to correlate with the ratio of unmasked regions. Due to page limitations, we provide further ablation studies on the impact of adaptively extracting low-and-mid frequency bands ... Case III shows substantial performance degradation, confirming that larger unmasked regions demand more low-frequency information from the null-text denoising process to replace the low-frequency bands in the text-guided denoising process, which is consistent with the findings in Sec.2.3.2 of the mainbody. These observations are also illustrated in Table 2."

---

[Q3]: Discussion on the Performance and Computational Efficiency

Response: Following your suggestions, we extend the experiments in Table 2 and Line 300 - 313, Sec. 3.3.1 of the mainbody to more results regarding the inference time and VRAM consumption during the early-stage denoising process, as shown below:

• Table. 2 of the mainbody

• Analysis of the balance of performance and computational efficiency

"The results demonstrates that our method significantly improves key performance metrics across both EditBench and BrushBench datasets. For example, IR increases by up to 44.86% and 10.99%, while LPIPS and CLIP Score also show consistent improvements, indicating better image quality and semantic alignment. Although inference time increases by 2.18s, this trade-off is fully justified given the substantial gains in generation quality and consistency. Overall, the performance boost far outweighs the moderate increase in computational cost. "

---

[Q4]: Discussion on the Unmasked Region Preservation

Response: Thanks. We kindly clarify that the CLIP-based metric shown in Fig.1 of the mainbody is used as a complementary metric, for measuring the preservation of unmasked regions. While CLIP distance primarily captures semantic-level similarity, it can help reveal whether the denoised image has undergone undesirable semantic shifts—including in the unmasked areas.

To ensure a more comprehensive evaluation, we have included multiple metrics (e.g., LPIPS, PSNR, SSIM and MSE). These jointly verify that our method better preserves the unmasked regions, which demonstrates the ability to tame the hybrid frequency issue, the results further verify the intuition in Sec.1 of the mainbody - NTN-Diff achieves the semantics consistency between mid-and-low frequency bands across masked and unmasked regions, while preserving unmasked regions. To facilitate the validation, we copy the quantitative results of unmasked region preservation between BrushNet [1] and ours in Table 1 of the mainbody and Appendix below:

• Table. 1 of the mainbody

• Table. 1 of the Appendix

---

References

• [1] Ju X, Liu X, et al. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 150-168.

Dear Reviewer zpgQ:

Many Thanks for your positive confirmation for our rebuttal and our work; undoubtedly, your constructive comments significantly help improve our work, where we are also deeply grateful to the updated rating as you raised up, along with your best wishes.

To Reviewer uzVF

Thank you very much for your valuable comments that helped us significantly improve this submission. Below are detailed point-to-point responses about how we address your questions.

---

[W1]: The Meaning of “One stone with two birds”

Response: Thanks for your valuable comments. We would like to clarify that, as mentioned in Line 32 - 52, Sec.1 of the mainbody, the longstanding challenges of text-guided image inpainting lie in the preservation for unmasked regions, while achieving the semantics consistency between unmasked and masked regions as inpainted. Previous arts failed to address both of them, always with either of them to be remedied. In this paper, we propose a null-text-null frequency-aware diffusion models, named NTN-Diff, for text-guided image inpainting, to simultaneously address two challenges of unmasked regions preservation, along with its semantics consistency with inpainted masked regions. Hence, we title one stone with two birds.

---

[W2]: Pseudo Code of the Algorithm

Response: As you suggested, we summarize the pseudo code of the algorithm below, which is consistent to our code in Sec.G of the Appendix for reproducibility.

Algorithm: NTN-Diff

---

Input: masked image $I$, binary mask $M$, text prompt $C$, null-text prompt $C_{∅}$
Output: inpainted image $I'$

---

I. Diffusion Process for Unmasked Regions

01: Extract the initial latent feature of the unmasked region $z_{0}^{gt} = E(I)$.
02: for $t = 0$ to $T_{inv} - 1$ do
03:  Compute $z_{t+1}^{gt}$ from $z_{t}^{gt}$;
04: end for {DDIM inversion}

---

II. Early Stage

05: Initialize $z_{T}^{un}$, $z_{T}^{text}$, $z_{T}^{mid} \sim N(0,1)$.
06: for $t = T$ to $\lambda T + 1$ do
07:  Substitute the unmasked regions of $z_{t}^{un}$ with the corresponding counterpart of $z_{t}^{gt}$ via Eq.3;
08:  Compute $z_{t-1}^{un}$ from $z_{t}^{un}$, conditioned on $C_{∅}$ via Eq.4;
09:  Substitute low-frequency band of $z_{t}^{text}$ with the corresponding counterpart of $z_{t}^{un}$ via Eq.5–7;
10:  Compute $z_{t-1}^{text}$ from $z_{t}^{text}$, conditioned on $C$;
11:  Substitute mid-frequency band of $z_{t}^{mid}$ with the corresponding counterpart of $z_{t}^{text}$ via Eq.8–10;
12:  Compute $z_{t-1}^{mid}$ from $z_{t}^{mid}$, conditioned on $C_{∅}$;
13: end for

---

III. Late Stage

14: for $t = \lambda T$ to $1$ do
15:  Substitute unmasked regions of $z_{t}^{mid}$ with the corresponding counterpart of $z_{t}^{gt}$ via Eq.11;
16:  Compute $z_{t-1}^{mid}$ from $z_{t}^{mid}$, conditioned on $C$;
17: end for
18: Output the final inpainted image $I' = D(z_{0}^{mid})$.

---

[W3]: Writing and Presentation

Response: Thanks for the comments, we will fix all your suggestions regarding writings issues, such as long sentences, small figure and typos etc.

---

[Q1]: The Explanation of Fig.1

Response: Thanks for the valuable comments. To better understand red dashed lines, we kindly clarify that t-SNE visualization of the CLIP latent space is adopted to illustrate the behavior of the text-guided denoising process in Fig.1 (a–c) of the mainbody, which aims to examine whether our method can preserve the unmasked regions and achieve semantic consistency between masked and unmasked regions across timesteps in the denoising process.

• Instead of computing the direct distance between the noise-free ground truth and the text prompt, red dashed line in the Fig.1 highlights the trace the denoising trajectory of the ground truth at different timesteps and connect them to serve as a reference path.

• Besides, we compare our NTN-Diff (Fig.1(c)) with state-of-the-arts for text-guided inpainting: The BLD result in Fig.1(d) corresponds to the type represented in Fig.1(a), while the BrushNet results in Fig.1(d) align more closely with the type in Fig.1(b), which provides qualitative support for our analysis.

---

[Q2]: The Explanation of Fig.2

Response: Thanks the comments. We visualize the low- and mid-frequency bands in the latent space during the denoising process to better illustrate their differences.

• At each timestep, we apply the 2D Discrete Cosine Transform to the latent representation and use the masks defined in Eq. (6) (Line 195 - 196) and Eq. (9) (Line 216 - 217) of the mainbody to extract the corresponding low- and mid-frequency components.

• We then apply the 2D Inverse Discrete Cosine Transform to project these frequency-filtered components back into the latent space, yielding the mid-frequency latent representation and the low-frequency latent representation.

• To further facilitate visual comparison, we compute the corresponding timestep-0 representations by inverting the noise addition step defined in Eq. (1) (Line 114 - 115) of the mainbody—i.e., applying the reverse operation of the noise injection without running the full denoising process, to facilitate the observation on the influence of each frequency band on the final output in a more isolated and interpretable manner.

• We then decode these latent representations into the image space via the pre-trained VAE decoder for visualization.

---

[Q3]: Changes in Low-and-High Frequency Components During Denoising Process

Response: As you suggested, we conducted the experiments to verify the conclusion in Lines 54–57, Sec. 1 of the mainbody: the low-frequency band fluctuates more significantly during the early stage of the denoising process with high-level noise than the mid-and-high frequency bands. This observation is consistent with the fact that diffusion models typically recover the low-frequency components first and progressively refine the mid- and high-frequency details.

Motivated by the above, we experimentally validate the denoising process as per several metrics to measure changes in the low-and-high frequency bands at different timesteps, as reported below:

• Analysis of the high-and-low frequency in the denoising process

"We adopt two groups of metrics to analyze the frequency components of the denoising process:

• For low-frequency information, we adopt Mean Gray, Mean HSV-V, and Mean Luma, which primarily capture the global brightness and luminance structure of the denoised image. These metrics help evaluate how well the low-frequency bands are maintained throughout the denoising process.

• For high-frequency information, we use Average Gradient, Variance, and LBP Variance, which are sensitive to edge sharpness, local contrast, and texture complexity, respectively. These metrics reflect the preservation and recovery of fine details during denoising.

The results show that low-frequency metrics (Mean Gray, HSV-V, Luma) remain nearly unchanged from the beginning, indicating early stabilization. In contrast, high-frequency metrics (Avg Gradient, Variance, LBP Variance) increase steadily in the late timesteps, confirming that high-frequency details are refined gradually during denoising, which is in line with Lines 54–57, Sec. 1 of the mainbody."

---

[Q4]: Discussion on the Second Null-Text Denoising Process of the Early Stage

Response: Thanks. as mentioned in Line 202 - 207, Sec.2.3.2 of the mainbody. We mildly clarify that the low-frequency band across both masked and unmasked regions are preserved even under text prompts owing to the substitutions from the first null-text denoising process, with only mid-frequency aligned with text prompts for both regions, hence still failed to achieve the consistency between masked and unmasked regions. To address this, we propose another concurrent null-text denoising process, guided by above denoised mid-level frequency, to help denoise the low-frequency band for both regions, especially for masked regions, To verify that, we also conduct the ablation studies in Line 301 - 313, Sec.3.3.1 of the mainbody, i.e., Table. 2, we copy that below:

• Table. 2 of the mainbody

• Analysis of the second null-text denoising process

“...Case C: removing the null-text denoising process (Sec.2.3.3); ... Case C exhibits a substantial performance degradation (at most 44.8%), confirming that only mid-frequency aligned with text prompts for masked and unmasked regions, still failed to achieve the consistency across the whole regions, which is in line with Sec.2.3.3 and illustrated in Fig.6.”

Dear Reviewer uzVF:

We truly agree with your thoughtful suggestions; while being grateful to your positive confirmation to our rebuttal, consistent positive rating and efforts for the improvement dropped to our work.

Many thanks!

To Reviewer hGay

Thank you very much for your valuable comments that helped us significantly improve this submission. Below are detailed point-to-point responses about how we address your questions.

---

[Q1]: Question on Computational Efficiency

Response: Thanks for your valuable comments. As per your suggestion, we made up the inference time and Peak VRAM consumption of our model, comparing it with other baseline models, along with the detailed explanations, as shown below:

• Analysis of computational efficiency

"Our inference time is comparable to BrushNet [3] and faster than recent methods such as HDP [4] and PP [2]. Moreover, our model consumes less VRAM than other text-guided inpainting models, owing to the fact that our method performs low-and-mid frequency substitution across different branches after each denoising step. The basic idea is to be decoupled from the resource-intensive denoising process itself, it incurs only negligible additional memory cost than the single denosing process BLD [1]."

---

[Q2]: Question on Robustness to Extreme Mask Ratios

Response: Thanks for your valuable comments. According to Line 246- 254, Remark 3 of the mainbody, We kindly clarify that due to the denoised mid-frequency band encodes information from the low-frequency band substituted from the null-text denoising process for ground truth including both masked and unmasked regions, hence the denoised mid-frequency band not only aligns with the text prompts for both regions, but also achieving the consistency to the substituted unmasked regions for ground truth,

Based on the above, large masked regions make it easier to keep consistency between masked and unmasked regions. Our method adaptively extracts low-frequency bands. So when the masked region is very large, only a small amount of low-frequency information is needed to maintain consistency.

We also provide quantitative comparisons on BrushBench to support this claim. As indicated in Lines 261–265, Sec. 3.1 of the mainbody, BrushBench refines the task by considering two specific scenarios: segmentation mask inside-inpainting and segmentation mask outside-inpainting. The outside-inpainting setting focuses on cases with large masked ratios, as shown in Table. 1 (Outside-Inpainting) of the mainbody, we copy that as follow:

• Table. 1 (Outside-Inpainting) of the mainbody

• Analysis on the quantitative results of the large mask

"The quantitative results summarize our findings below: NTN-Diff enjoys larger IR, HPS v2, PSNR and CLIP Score, together with smaller MSE and LPIPS than the competitors. Notably, BrushNet remains the large performance margins (at most 6.65% for IR, 0.71% for HPS, 2.27% for PSNR, 10.59% for MSE, 4.11% for LPIPS and 0.77% for CLIP Score) compared to NTN-Diff, which demonstrates the ability to tame the hybrid frequency issue, the results further verifies the intuition in Sec.1 of the mainbody -- NTN-Diff achieves the semantics consistency between mid-and-low frequency bands across masked and unmasked regions, while preserving the unmasked regions. We also illustrate the above intuitions in the last two rows of Fig.1 (d) and the last row of Fig. 5(a) in the mainbody "

---

[Q3]: Regarding the experiments on EditBench

Response: Thanks.We mildly remind you that we do present a comparison of text-guided inpainting results with state-of-the-art methods on EditBench in the right part of Fig. 5 in the mainbody. In addition, due to the page limitation for mainbody, we include further quantitative comparisons on EditBench with state-of-the-art approaches in the Line 45-54, Sec. B of the Appendix, i.e., Table. 1, as indicated by Line 298, Sec. 3.2 of the mainbody, we copy that as follow:

• Table. 1 of the Appendix

• Analysis of the quantitative results on EditBench

"In particular, we further exhibit additional quantitative results by performing the experiments on the EditBench; It is observed that NTN-Diff enjoys larger PSNR, SSIM and CLIP Score, together with smaller MSE and LPIPS than the competitors. Notably, BrushEdit remains the large performance margins (at most, 3.86% for PSNR, 0.42% for MSE, 1.85% for LPIPS, 0.86% for SSIM, and 0.1% for CLIP Score) compared to NTN-Diff in table. The results further verifies the intuition in Sec.1 of the mainbody -- NTN-Diff can achieve the semantics consistency between mid-and-low frequency bands across masked and unmasked regions, while preserving unmasked regions."

---

References

• [1] Avrahami O, Fried O, Lischinski D. Blended latent diffusion[J]. ACM transactions on graphics (TOG), 2023, 42(4): 1-11.

• [2] Zhuang J, Zeng Y, Liu W, et al. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 195-211.

• [3] Ju X, Liu X, Wang X, et al. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 150-168.

• [4] Manukyan H, Sargsyan A, Atanyan B, et al. Hd-painter: high-resolution and prompt-faithful text-guided image inpainting with diffusion models[C]//The Thirteenth International Conference on Learning Representations. 2025.

• [5] Rombach R, Blattmann A, Lorenz D, et al. High-resolution image synthesis with latent diffusion models[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022: 10684-10695.

• [6] Zhang L, Rao A, Agrawala M. Adding conditional control to text-to-image diffusion models[C]//Proceedings of the IEEE/CVF international conference on computer vision. 2023: 3836-3847.

• [7] Li Y, Bian Y, Ju X, et al. Brushedit: All-in-one image inpainting and editing[J]. arXiv preprint arXiv:2412.10316, 2024.

Dear Reviewer hGay,

We are sincerely grateful to your positive confirmation and promoted rating.

Many thanks for your great guidance for our work !