RoomEditor: High-Fidelity Furniture Synthesis with Parameter-Sharing U-Net

Mathematical derivation

Thank you for raising this question. We agree that, in general, diffusion models do not take ground truth images as input. However, for inpainting tasks, it is common practice to provide the masked image (in our case, the masked background $\boldsymbol{I}_{\text{bg}}^{\text{M}}$) as input, so that the model can preserve the content in the unmasked regions. Since the model generates content over the entire image, this implies that a portion of the ground truth (i.e., the unmasked area) is effectively included in the input. Therefore, a characteristic of inpainting diffusion models is that their predictions in the unmasked regions can be considered as ground truth.

Equation (5): Based on this understanding, we consider two different model inputs. (1) By masking the object in $\boldsymbol{I}\_{\text{gt}}$ to get $\boldsymbol{I}\_{\text{bg}}^{\text{M}}$ and applying the aforementioned property, we find the model's prediction for the background region is essentially identical to that in $\boldsymbol{I}\_{\text{gt}}$; (2) by masking the background in $\boldsymbol{I}\_{\text{gt}}$ to retain only the object (represented as $\mathcal{R}(\boldsymbol{I}\_{\text{ref}}^{M})$ in our paper) and applying the same property, we find the prediction for the object region is also essentially identical to that in $\boldsymbol{I}\_{\text{gt}}$. Combining the masked outputs of these two ideal predictions would yield the ground truth $\boldsymbol{I}\_{\text{gt}}$.

Equation (6): It generalizes the principle of Equation (5) from the pixel level to the feature level. This generalization is based on a key assumption regarding the U-Net's architecture, for which the rationale is two-fold: (a) convolutional and MLP layers operate locally, and (b) cross-attention mechanisms are primarily designed to inject object features into the masked region, thus having minimal impact on the unmasked background features. Based on this assumption, we posit that the features at locations corresponding to $\boldsymbol{I}\_{\text{bg}}^{\text{M}}$ (i.e., $f\_{l}(\boldsymbol{I}\_{\text{bg}}^{\text{M}})$ ) and those corresponding to $\mathcal{R}(\boldsymbol{I}\_{\text{ref}}^{M})$ (i.e., $f\_{l}(\mathcal{R}(\boldsymbol{I}\_{\text{ref}}^{M})$) can be combined to approximate the features of the complete ground-truth image $\boldsymbol{I}\_{\text{gt}}$ (i.e., $f\_{l}(\boldsymbol{I}\_{\text{gt}})$ ).

Equation (7): It provides a high-level conceptual model of the reference encoding process. It assumes the existence of an ideal transformation, $\mathcal{R}\_{l}$, that maps the features of the input reference image to the features that would have been generated if the ground-truth object were present. In practice, different models learn different approximations of this mapping (see the discussion in Section 4.2.2 Comparison with Previous Works, line 194-213). Our formulation with parameter sharing is simply a novel and efficient way to learn this transformation.

Trainable Reference U-Net

• Is the reference U-Net kept fixed or is it updated: The reference U-Net is kept frozen (to remain consistent with MimicBrush).
• Trainable reference U-Net: This is an insightful point, and we have actually explored it in our ablation study. As shown in Table 4, “Dual U-Net (Frozen g) + CLIP” corresponds to MimicBrush$^{\dagger}$, while “Dual U-Net (Trainable g) + CLIP” represents the variant suggested by the reviewer. The results show that training the encoder $g(\cdot)$ leads to some performance improvement. However, it still falls short of the performance achieved by using a shared encoder $f(\cdot)$ for both inputs. This underscores the effectiveness of our design choice to adopt a single encoder in terms of both simplicity and performance.
• Qualitative results of trainable reference U-Net: We have not included the relevant qualitative results in the current version of the paper, but we will add them along with corresponding analysis in the revision. In our observations, although the overall harmony between the background and object is relatively high, using two separate encoders often leads to noticeable inconsistencies between the generated object and the reference in terms of color, shape, and pattern.

The Motivation for Incorporating CLIP in the Ablation Study

Thank you for this question. As noted in lines 157–159, mainstream dual U-Net architectures (such as MimicBrush) typically incorporate an additional image encoder. In contrast, our position is that this additional encoder can be removed, as its functionality largely overlaps with the reference U-Net, making it redundant. We validate this claim through ablation studies demonstrating its dispensability (see Table 4).

The metric Fidelity Rank in Table 2

Thank you for pointing this out. Fidelity refers to the identity consistency between the reference object and the generated object in terms of color, shape, pattern, and other visual attributes. We will clarify this in Section 5.2.2 (User Study) of the revised version.

Number and Background of User Study

• Number of user study: Following the protocol in the MimicBrush paper [3], we initially conducted a user study with 10 participants for consistency. To enhance the reliability of our results, we later expanded the study to include 25 participants in total.
• Background of user study: The participants consisted of 14 undergraduate students, 7 parents of some of the students, and 4 university faculty or staff members, all of whom were informed of the specific evaluation criteria.
• Updated results: The findings remain consistent—our method outperforms existing approaches across all evaluated metrics. These updated results will be included in the revised version.

Response to the Comment on Virtual Try-on Methods

We appreciate the reviewer’s comment. Virtual try-on indeed involves modeling complex garment-body interactions. However, our task differs in that furniture synthesis requires handling diverse furniture shapes, whereas clothing in virtual try-on applications typically exhibits more consistent shapes. We will revise the relevant statement in the manuscript to better clarify these distinctions.

Response to Claim about inconsistency

Thank you for raising this point. Here, "inconsistency" refers to discrepancies between the object or color in the reference image and those in the background image—in other words, they do not depict the same object. For example, in Appendix Figure 5, the left image in the second row (background image) and the right image in the same row (reference image) clearly correspond to different objects. Such cases are quite common in practice, as a single product may have multiple versions or color variants.

References

[1] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381–18391, 2023.

[2] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593–6602, 2024.

[3] Xi Chen, Yutong Feng, Mengting Chen, Yiyang Wang, Shilong Zhang, Yu Liu, Yujun Shen, and Heng shuang Zhao. Zero-shot image editing with reference imitation. Advances in Neural Information Processing Systems, 37:84010–84032, 2025.

We sincerely thank the reviewer for the fast response, and we are very sorry for not providing a clearer explanation on the contribution earlier. We also appreciate the recognition of our benchmark contribution — we agree that a comprehensive benchmark is crucial for advancing the field.

Architectural Differences: Methods like MimicBrush [7] typically employ dual U-Nets and rely on additional components such as the CLIP visual encoder. In contrast, our architecture is more concise: (1) we use a single U-Net shared for both reference and background encoding; and (2) we eliminate the need for CLIP or other extra modules. We also validate through ablation studies (Table 4) that the contribution of CLIP in existing U-Net architectures is marginal, which further confirms the effectiveness of our extremely concise design. As a result, this simplification significantly reduces model complexity and computational overhead (see table below) , while greatly improving performance (see Table 1 and Table 3).

Insight and Contribution: Beyond the dataset itself, one of our key contribution is the theoretical modeling of semantic image synthesis, through which we identify a fundamental issue in current SOTA dual U-Net designs — feature inconsistency — that greatly reduces image fidelity. We address this by instantiating $\mathcal{R}\_l$ via cross-attention and $f\_l$ via $g\_l$ (section 4.2), leading to a shared-parameter U-Net architecture that effectively mitigates inconsistency. Experimental results show that our architecture design can generalize well to generic objects, as also recognized by the reviewer. We believe this finding provides valuable insights for the image editing field. We also believe our concise, efficient architecture design with good generalization ability and a certain theoretically support can approach an important trend in current AI algorithms.

Great thanks for your fast reply again.

Robustness to Mask Quality

We thank the reviewer for raising this important point. We fully agree that a model's performance with low-quality masks is a critical measure of its practical utility, and we have in fact discussed this specific issue in our work. Our methodology reflects this discussion in two key stages:

1. Robustness in Training: To better simulate real-world user scenarios, our training process (detailed in Appendix B, line 482-484) intentionally uses imperfect inputs. This includes using random perturbations and simple bounding boxes to make the model robust against imprecise segmentation (see Figure 6)."
2. Realistic Evaluation: Similarly, to simulate real-world scenarios, we instructed annotators to create intentionally coarse masks for the test set (as stated on lines 484-486). Consequently, all results reported in our paper are based on these coarse-grained annotations, not high-quality segmentations, with examples illustrated in Figure 7(b).

To further evaluate the robustness of our model, we add an experiment to our RoomBench using bounding boxes as masks, without any shape prior. The results are sufficiently illustrative, as shown below.

Despite the lack of shape guidance, our model still outperforms MimicBrush, which has been trained on large-scale data and is known for its robustness. This provides strong evidence that our method is highly robust to mask variations.

Insufficient Related Methods

We sincerely thank the reviewer for pointing out these advanced approaches. In our revised paper, we will add a detailed discussion of both In-Context LoRA [1] and HomeDiffusion [2] to the related work section, summarizing their core contributions as follows:

• In-Context LoRA leverages in-context learning to generate thematically consistent images from a few visual examples at inference time.
• HomeDiffusion employs a two-stage strategy, first learning generation from multi-view references and then focusing on object-background integration.

However, a direct quantitative comparison with these methods is unfortunately infeasible. In-Context LoRA's task formulation differs significantly from ours, as it does not take two images (background and reference) as input or support precise object placement. While HomeDiffusion is more related to our task, its code and benchmark are not publicly available, precluding a fair comparison.

To further strengthen our evaluation and address the reviewer's valid concern about comparing with SOTA models, we selected the commercial, closed-source model GPT-4o, which supports multi-image inputs and demonstrates impressive image editing capabilities. Due to cost and time constraints, we randomly sampled 200 examples from the RoomBench dataset and compared GPT-4o with our method and MimicBrush [3] on this subset.

Our qualitative analysis reveals that while GPT-4o often produces coherent images with realistic lighting, it struggles with preserving object fidelity, frequently altering the shape and design of the reference object. In contrast, our method achieves superior fidelity, which is reflected in its higher CLIP-score and DINO-score. We will include this comparison in the revised version.

On Generalization and Cross-Dataset Evaluation

Thank you for raising this important point. We address these concerns with existing clarifications and new experiments as follows:

• Strict Data Separation: To prevent data leakage between training and testing sets, we ensured a strict separation of data sources during the collection of our RoomBench dataset. We will explicitly state this protocol in the revised paper to ensure clarity on the validity of our results.

• Existing Generalization Evidence: We would like to gently point out that our original paper already demonstrated strong generalization to non-furniture objects through experiments on the DreamBooth dataset [4] (Table 3 and section 5.3, line 277). This experiment shows that our model, although not trained on generic data, outperforms models trained on large-scale generic datasets. This demonstrates the strong generalization capability of our approach.

• New Cross-Dataset Evaluation on 3D-FUTURE: In addition, in response to the reviewer’s suggestion of testing on alternative furniture composition datasets, we evaluated our model on the 3D-FUTURE [5] dataset. We selected 1,020 samples covering 34 furniture categories (randomly 30 images per category). Since the dataset does not provide explicit image pairs, we applied horizontal flipping to reference objects and blurred the masks to avoid trivial copy-paste solutions. The results are as follows:

Although our model was not trained on 3D-FUTURE, it achieves consistently better performance than Mimicbrush across multiple metrics, highlighting its strong generalization ability. We will incorporate all these clarifications and new results into our revised paper.

Computational Resources and Time Consumption

Thank you for raising this question. We report the total number of parameters and trainable parameters for different architectures in Table 4: Ablation Study. To enable a more comprehensive evaluation—including speed and memory consumption—we provide a comparative summary in the following table. All experiments were conducted using a single A6000 GPU in FP16 mode on our test set, with single-image inference and 50 denoising steps.

As shown, our method requires significantly fewer parameters and less GPU memory compared to Mimicbrush, while also achieving faster generation speed.

Multi-view images

Thank you for this valuable question. Our model currently operates on single-view images. While we certainly acknowledge the potential of multi-view inputs, this direction was not prioritized in the current scope due to our specific research focus and challenges in data availability. We will ensure this is highlighted as a promising avenue for future research in our revised conclusion.

Quality of reference images

This is a practical and important concern. Our experiments show that when the reference object includes watermarks or mosaic artifacts, the model tends to perceive them as integral parts of the object and thus fails to remove them during generation. Improving robustness to such artifacts is a valuable direction we plan to explore in future work.

References

[1] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers. arXiv preprint arXiv:2410.23775, 2024.

[2] Guoqiu Li, Jin Song, and Yiyun Fei. Homediffusion: Zero-shot object customization with multi-view representation learning for indoor scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4698–4706, 2025.

[3] Xi Chen, Yutong Feng, Mengting Chen, Yiyang Wang, Shilong Zhang, Yu Liu, Yujun Shen, and Heng shuang Zhao. Zero-shot image editing with reference imitation. Advances in Neural Information Processing Systems, 37:84010–84032, 2025.

[4] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500–22510, 2023.

[5] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313–3337, 2021.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with less than three days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Baseline Choice

Thank you for the question. The following are relevant explanations.

1. Regarding Text-to-Image Models and Diffuse-to-Choose: We sincerely appreciate the suggestion. However, text-to-image models are not directly applicable to our task, which involves generating a new object conditioned on both a masked background and a reference image—input modalities that such models are not inherently designed to handle. As for Diffuse-to-Choose [1], we acknowledge its relevance, but unfortunately, its code, model, and evaluation data are not publicly available at this time, which makes a fair and reproducible comparison currently infeasible.

2. Current Baseline Choice : We selected three methods that were the most relevant, representative, and state-of-the-art at the time of our study, each covering a distinct architectural paradigm in example-based image composition.

   • PBE [2] (line 97): A Stable Diffusion inpainting model that injects reference via CLIP visual encoder; suffers from information loss due to CLIP.
   • AnyDoor [3] (lines 99-100): Uses ControlNet for background consitency and a stronger DINO encoder for reference integration.
   • MimicBrush [4] (lines 101-102): Employs a dual U-Net architecture with explicit reference feature encoding.

3. Further evaluation with GPT-4o: Given the lack of directly comparable methods in existing work, we further evaluated GPT-4o on our dataset to benchmark against a more advanced model. We provided the reference image and the masked background as input, using the prompt: "Place the object from the first image into the masked area of the second image." Due to API costs and long generation time, we randomly sampled 200 examples and compared GPT-4o with our method and MimicBrush.

Our findings indicate that GPT-4o generates highly coherent images with realistic lighting. However, it sometimes struggled with object fidelity, where the generated object's shape or design deviated from the reference image. This is reflected in its lower CLIP-score and DINO-score compared to our method. We will include this new comparison in our revised paper. We believe benchmarking against a state-of-the-art model like GPT-4o provides a more comprehensive evaluation.

Mathematical Justification (Equations 5–7)

We are happy to clarify the reasoning behind our mathematical formulation. We acknowledge that our derivations are based on certain modeling assumptions. Below, we explain the rationale behind each equation in more detail.

• Equation (5): This equation is grounded in a key empirical property of inpainting models: they are trained to preserve unmasked regions with high fidelity. This arises from both the task design—where known regions are provided as input—and the loss function, which is applied over the entire image. We consider two different model inputs: (1) by masking the object in $\boldsymbol{I}\_{\text{gt}}$ to get $\boldsymbol{I}\_{\text{bg}}^{\text{M}}$ and applying the aforementioned property, we find the model's prediction for the background region is essentially identical to that in $\boldsymbol{I}\_{\text{gt}}$; (2) by masking the background in $\boldsymbol{I}\_{\text{gt}}$ to retain only the object (represented as $\mathcal{R}(\boldsymbol{I}\_{\text{ref}}^{M})$ in our paper) and applying the same property, we find the prediction for the object region is also essentially identical to that in $\boldsymbol{I}\_{\text{gt}}$. Combining the masked outputs of these two ideal predictions would yield the ground truth $\boldsymbol{I}\_{\text{gt}}$.
• Equation (6): Equation (6) generalizes the principle of Equation (5) from the pixel level to the feature level. This generalization is based on a key assumption regarding the U-Net's architecture, for which the rationale is two-fold: (a) convolutional and MLP layers operate locally, and (b) cross-attention mechanisms are primarily designed to inject object features into the masked region, thus having minimal impact on the unmasked background features. Based on this assumption, we posit that the features at locations corresponding to $\boldsymbol{I}\_{\text{bg}}^{\text{M}}$ (i.e., $f\_{l}(\boldsymbol{I}\_{\text{bg}}^{\text{M}})$ ) and those corresponding to $\mathcal{R}(\boldsymbol{I}\_{\text{ref}}^{M})$ (i.e., $f\_{l}(\mathcal{R}(\boldsymbol{I}\_{\text{ref}}^{M})$) can be combined to approximate the features of the complete ground-truth image $\boldsymbol{I}\_{\text{gt}}$ (i.e., $f\_{l}(\boldsymbol{I}\_{\text{gt}})$ ).
• Equation (7): Equation (7) provides a high-level conceptual model of the reference encoding process. It assumes the existence of an ideal transformation, $\mathcal{R}\_{l}$, that maps the features of the input reference image to the features that would have been generated if the ground-truth object were present. In practice, different models learn different approximations of this mapping (see the discussion in Section 4.2.2 Comparison with Previous Works, line 194-213). Our formulation with parameter sharing is simply a novel and efficient way to learn this transformation.

We will update the explanation accordingly in the revised version.

Originality in the model architecture

We understand the reviewer’s concern about architectural novelty. However, we would like to highlight that our contributions extend beyond architecture and include data and in-depth analysis. We introduce a new modeling perspective by identifying a feature inconsistency issue commonly present during training in existing methods (validated in Line 214 and Figure 3), and address it through a parameter-sharing strategy. While this architectural modification is concise, it yields significant improvements in both performance and efficiency. In addition, to the best of our knowledge, we are the first to provide a mathematical analysis and corresponding empirical validation for this approach, which we believe offers the community concrete insights into this problem space.

References

[1] Mehmet Saygin Seyfioglu, Karim Bouyarmane, Suren Kumar, Amir Tavanaei, and Ismail B Tutar. Diffuse to choose: Enriching image conditioned inpainting in latent diffusion models for virtual try-all. arXiv preprint arXiv:2401.13795, 2024.

[2] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381–18391, 2023.

[3] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593–6602, 2024.

[4] Xi Chen, Yutong Feng, Mengting Chen, Yiyang Wang, Shilong Zhang, Yu Liu, Yujun Shen, and Heng shuang Zhao. Zero-shot image editing with reference imitation. Advances in Neural Information Processing Systems, 37:84010–84032, 2025.

Further mathematical justification and verification experiments

Thank you for the comment. We acknowledge that some of the underlying assumptions may lack sufficient rigor. Empirical validation of these assumptions is a valuable suggestion. In the following, we conduct experiments to verify the core assumptions proposed in Equations (5)–(7).

Equations (5): the core assumption is that the inpainting diffusion model has the ability to preserve unmasked regions with high fidelity. Therefore, we aim to demonstrate that the expression $\boldsymbol{M} _{\text{bg}} \odot \epsilon _{\theta}(\boldsymbol{I} _{\text{bg}}^{\text{M}}) + \overline{\boldsymbol{M} _{\text{bg}}} \odot \epsilon _{\theta}(\mathcal{R}(\boldsymbol{I} _{\text{ref}}^{\text{M}} \mid \boldsymbol{I} _{\text{bg}}^{\text{M}}))$ can effectively approximate the ground truth (GT) $\epsilon\_\theta(\boldsymbol{I\_\mathrm{gt}})$. To verify this, we conducted experiments using a pretrained diffusion inpainting model on our full test set. During the experiments, we added random noise to the inputs and sampled random timesteps t, then fed the model with partial content from the full image (background or object). We extracted predictions from the corresponding regions and used $L\_2$ loss as the evaluation metric, i.e.,

As shown in the table below:

Here, the training loss refers to the average loss during model training (as can be verified from Figure 3 of the paper). Notably, the merged result—obtained by feeding the model with separate parts and combining the corresponding outputs—yields a significantly lower loss than the training loss, supporting the validity of the assumption made in Equation (5).

Equations (6): the assumption in Equation (6) is that a property similar to that in Equation (5) also holds at the feature level. To better assess this from a similarity perspective, we adopt cosine similarity as the evaluation metric. Specifically, we compute the cosine similarity between feature representations from different attention layers on the test set, i.e.,

The results are presented below, where 'd', 'm', and 'u' denote down, mid, and up layers, respectively.

It can be observed that the cosine similarities are consistently high (> 0.9) across many layers, which supports the validity of our assumption in Equation (6).

Equation (7): It presents an existence assumption—namely, that it is possible to learn a transformation between features such that the fused features approximate the ground-truth (GT) features. We provide empirical support for this assumption in Figure 3 of the paper: (1) The feature similarity of different methods consistently decreases throughout the training process, indicating that the model is indeed learning to align the fused features with the GT features; (2) Low feature losses across multiple layers further demonstrate a high degree of feature consistency.

Regarding the sample size of GPT-4o

Thank you for the comment. The reason we evaluated GPT-4o on a subset of only 200 images is primarily due to practical constraints. GPT-4o's API inference is extremely slow, averaging about one minute per image, and incurs significant costs about \$0.3 per image. Running the evaluation on the entire validation set would require approximately \\$600, a substantial expense that is challenging for smaller research labs with limited budgets. However, if this evaluation approach is demonstrated to be broadly applicable and beneficial, we will consider conducting experiments on the full test set in future work.

Thank you very much for your detailed response, and we are very pleasure to hear that our response address your concerns. We sincerely appreciate your recognition of our work and continuing strong support. Particularly, your thoughtful comments greatly improve our work, which will be definitely incorporated into our revised manuscript. Great thanks again for your time and valuable insights.

Sample Size of User Study

We thank the reviewer for pointing out this issue. Following the protocol in the MimicBrush paper [1], we initially conducted a user study with 10 participants. To strengthen the reliability of our results, we conducted an additional study with 15 more participants (total 25). The participants comprised 14 undergraduate students, 7 of their parents (to provide a non-technical perspective), and 4 university faculty or staff members, all of whom were explicitly informed of the evaluation criteria.

The findings remain consistent—our method outperforms existing approaches across all evaluated metrics. These updated results will be included in the revised version.

Failure Case and Cause Analysis

Thank you for the feedback. An initial failure case analysis was provided in the appendix. Here, we offer a more detailed explanation, as figures cannot be included.

1. Our model may occasionally struggle with capturing complex furniture shapes, as the large intra-category variation in geometry makes it difficult for the dataset to fully represent all possible shape patterns, thereby affecting the learned shape priors (e.g., Figure 11 shows distorted shapes).
2. Since our method edits only the masked region, shadows outside the mask cannot be generated. This is a common limitation shared by PBE [2], AnyDoor [3], and MimicBrush [1].

In summary, the main causes are the diversity of real-world furniture types and inherent constraints of the masked-editing paradigm. We will include more visualizations and in-depth analysis in the revised version.

Robustness to Lighting Changes

Thank you for raising this important point. Since our original test set does not account for lighting variation, we conducted additional experiments to evaluate the robustness of our model under different illumination conditions by adjusting the image brightness. Specifically, we applied lighting intensity adjustments to the background images with scaling factors of 0.6 and 0.8 (darker), as well as 1.2 and 1.4 (brighter), and conducted evaluations on the full test set. We compared our method with MimicBrush and computed the average scores (all metrics were rescaled to a 0–100 range, with FID and LPIPS inverted), as shown below.

Since our model design and data construction did not explicitly consider extremely dark or bright lighting conditions, there was a slight performance drop. However, it is worth noting that the degree of performance drop is comparable to that of MimicBrush, so our method still maintains a significant lead under the same settings. We will include lighting-related experiments, analysis, and visual results in the revised version.

References

[1] Xi Chen, Yutong Feng, Mengting Chen, Yiyang Wang, Shilong Zhang, Yu Liu, Yujun Shen, and Heng shuang Zhao. Zero-shot image editing with reference imitation. Advances in Neural Information Processing Systems, 37:84010–84032, 2025.

[2] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381–18391, 2023.

[3] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593–6602, 2024.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with less than three days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.