SeeThruAnything: Learning to Remove Any Obstructions Across Distributions

We sincerely appreciate Reviewer tqQw's valuable feedback. To provide a comprehensive overview and visual evidence of the improvements, we have included the updated manuscript in the supplementary materials. We kindly invite you to review this revised version, which thoroughly documents all new comparisons and analyses.

Response-W1: Thank you for your comment. We believe there may be a misunderstanding regarding the source of our method’s generalization capability. The generalization performance does not stem primarily from the quality of the predicted masks but from our proposed formulation, which uniformly represents the complex task of obstacle removal. Even when provided with ground truth obstacle masks, many existing methods fail to remove obstructions, especially those unseen during training. This is evidenced by the visual comparison of rare obstacle removal in Figure 6 and the results of inpainting-based methods using masks, as detailed in Appendix C.3.

The effectiveness of our model lies in its ability to handle diverse obstructions through the introduction of soft and hard recovery strategies, which address both opaque and semi-transparent obstacles. Additionally, the use of multi-modal prompts allows for precise representation of obstacles, further enhancing the model’s capabilities. The mask detector we use is an off-the-shelf component and not the primary focus of this paper. However, our model is plug-and-play, meaning it can seamlessly integrate with more advanced mask detectors in the future to further improve performance.

Response-W2: Thank you for your comment. We initially included a comparison with inpainting-based methods in Figure 1 of the original draft. However, as discussed in the Introduction, inpainting-based methods, while capable of producing visually plausible results, often fail to faithfully reconstruct the original image content. This lack of fidelity was the primary reason we did not include a detailed comparison with inpainting methods in the initial draft, as our focus is on methods that preserve the authenticity of the original scene.

To address this concern and provide a more comprehensive evaluation, we expanded the comparison in Appendix C.2, focusing on zero-shot recovery performance for unseen obstacles. Since the code for GradPaint [1] is not publicly available, we included LaMa [3] and Repaint [2] as competitors. The results, presented in Table 5 and Figure 12, demonstrate that our method outperforms these inpainting-based methods both quantitatively—across three classic obstacle removal tasks—and visually, particularly for less common obstacles.

Table 5: PSNR and SSIM comparisons of our method with inpainting-based methods on unseen obstructions. The best results are highlighted in bold.

• [1] Grechka A, Couairon G, Cord M. GradPaint: Gradient-guided inpainting with diffusion models, CVIU, 2024.
• [2] Lugmayr A, Danelljan M, Romero A, et al. Repaint: Inpainting using denoising diffusion probabilistic models, CVPR 2022.
• [3] Suvorov, Roman, et al. Resolution-robust large mask inpainting with fourier convolutions,WACV, 2022.

Response-W3: Thank you for your observation. Our method is specifically designed to address the challenges posed by unseen obstructions, which may explain why its advantages are less pronounced for known obstructions. Nonetheless, comprehensive comparisons across the three datasets demonstrate that our method achieves consistently superior overall performance.

Moreover, the comparable performance on both seen and unseen categories highlights a key strength of our approach: the distribution-agnostic obstruction formulation. By treating all obstructions uniformly as a transparency recovery problem, our method ensures consistent and robust performance across diverse scenarios, regardless of whether the obstructions were encountered during training.

We sincerely thank Reviewer AeWH for their valuable comments. To provide a comprehensive overview and visual evidence of the improvements, we have included the updated manuscript in the supplementary materials. We kindly invite you to review this revised version, which thoroughly documents all new comparisons and analyses.

Response-W1: Thank you for your observation. To address this, we have included a detailed comparison of model parameters, FLOPs, and runtime in Appendix F. While our model has a slightly higher number of parameters compared to the baselines, the FLOPs and runtime—which are more reflective of computational efficiency—remain within a practical range. Notably, the increase in parameters primarily arises from the introduction of the cross-attention module for integrating multi-modal prompts, not from the tunable adapter itself. Specifically, the adapter contributes only 0.2176 million additional parameters and 11.08 giga FLOPs, ensuring that its impact on computational cost is minimal.

Table 7: Comparisons of parameters, FLOPs, and runtime between.

Response-W2: Thank you for your comment. We would like to clarify that we conducted this ablation study in our original submission, and the results are presented in Table 3. To make this clearer, we have highlighted this section in the revised manuscript and added explanatory footnotes to improve readability and ensure the results of our ablation study are easily understood.

Table 3: PSNR and SSIM comparisons of integrating different modules.

Response-W3: Thank you for your suggestion, we have reorganized Figure 1 for better comparison.

Response-Q1: Thank you for your insightful question. We have conducted experiments on images with multiple types of obstructions, as detailed in Appendix C.2 and illustrated in Figure 11. By leveraging multi-modal prompts to represent various obstructions, SeeThruAnything effectively removes the specified obstacles with high accuracy. Additional results and in-depth analysis are provided in the revised manuscript for your review.

Response-Q2: Thank you for your question. The mask detector used in our method is described in detail in Appendix A.1 of the initial draft. In the revised version, we have highlighted this section in blue for clarity. To ensure a fair comparison, the same mask inputs are used across all experiments for both our method and the comparative methods.

Response-Q3: Thank you for your question. Examples of the text commands we use are provided in Figure 8 and Figure 14, ranging from simple descriptive phrases to more complex sentences. We typically use different text commands to specify whether an obstacle is opaque or semi-transparent. A generic command like "remove obstruction" may not yield optimal results, as our method interprets it as referring to opaque obstacles and performs hard mask recovery accordingly. However, by including transparency-specific descriptions in the input text, such as "opaque yarn" or "semi-transparent obstacles," our method achieves significantly better results.

We sincerely appreciate Reviewer UqLU's valuable feedback. To provide a comprehensive overview and visual evidence of the improvements, we have included the updated manuscript in the supplementary materials. We kindly invite you to review this revised version, which thoroughly documents all new comparisons and analyses.

Response-W1: Thank you for your feedback. We believe this concern may stem from a partial or shallow understanding of some aspects of our techniques. Our method introduces a novel distribution-agnostic obstruction formulation that unifies obstruction removal as a single framework, treating all obstructions as a soft masking problem. This formulation focuses on understanding the underlying context of transparent obstructions, which is critical for effective removal.

We demonstrate that even with a ground truth mask, the inherent distribution of different obstructions cannot be easily handled using traditional approaches. Our approach addresses this limitation by enabling zero-shot removal through a flexible soft mask recovery strategy, which generalizes across diverse obstruction types, including unseen scenarios. This perspective has not been explored in prior work and significantly advances both the theoretical foundation and practical application of obstruction removal.

Response-W2&Q1: Thank you for your comments and questions. We believe some aspects of your concerns might stem from a misunderstanding. The original versions of WGWSNet and PromptIR are not designed for zero-shot tasks and are largely ineffective for unseen obstructions. To ensure a fair comparison, we standardized the inputs for all methods, including the same degraded images and masks, effectively giving these methods some degree of zero-shot removal capability. While their complex designs allow them to perform comparably to ours in specific unseen scenarios with distributions similar to their training data, our method consistently outperforms them. The PSNR and SSIM comparisons across three unseen scenarios (Table 2) highlight our model's advantages, with a PSNR lead of 1.88 dB over the second-best method.

Regarding the similarity between fences and yarn, this appears to be a misconception. Fences exhibit regular patterns and predictable distributions in the image domain, while rare obstructions like yarn have random distributions and highly irregular patterns, making them much harder to remove. This difference is clearly demonstrated in the results shown in Figure 6.

Our approach addresses these challenges through a novel soft and hard mask strategy that understands the underlying context of diverse obstructions, enabling effective removal even in complex or unseen scenarios. Competing methods struggle in these cases, which underscores the unique generalization capability of our model. We hope this clarification resolves any misunderstanding and highlights the strengths of our proposed method.

We sincerely appreciate the insightful comments provided by Reviewer Ms8C. To offer a detailed account and visual evidence of the improvements, we have included the updated manuscript as part of the supplementary materials. We kindly invite you to review this revised version, which thoroughly documents all new comparisons and analyses.

Response-W1: Thank you for raising this important point. Prior to training the SeeThruAnything model, we fine-tuned the CLIP text encoder to adapt its embedding space to our specific task. The fine-tuning process aimed to optimize the embedding space by reducing intra-class distances (i.e., decreasing the cosine distance between embeddings of different textual descriptions for the same task) while increasing inter-class distances (i.e., enhancing the separation between embeddings of textual descriptions for different tasks). This adaptation ensures that the CLIP text encoder can effectively handle task-specific instructions. As this is a relatively small adjustment, it was not included in the initial draft. However, we have provided a detailed explanation of this technique in Appendix B of the updated manuscript for your reference.

Response-W2: Thank you for your suggestion. To improve the model's reproducibility and make it easier for readers to understand, the usage details about CLIP have been added to the Appendix B of the updated version.

Response-W3: Thank you for your insightful suggestions. We have added the ablation experiment using only the visual prompt to Table 4 for completeness. Additionally, we explored replacing the CLIP text embedding model with the BLIP model, and the results show comparable performance to our final setup. Detailed metrics and analysis can be found in Appendix D.2 and Table 6.

Initially, we did not include the "visual prompt only" setting because using additional images as input deviates from the primary goal of leveraging text-based instructions, which are more direct, interpretable, and practical in real-world scenarios. Such a setting introduces complexity that may limit the applicability and flexibility of the model. However, we appreciate the opportunity to address this point in the revised manuscript.

Table 4: PSNR and SSIM comparisons of using different prompt strategies.

Table 6: PSNR and SSIM comparisons of using different prompt generation strategies.

Response-Q1: Thank you for raising this important concern. We believe this issue stems from a misunderstanding, which we have now clarified in Appendix B. Specifically, the CLIP text encoder was fine-tuned to better adapt to task-specific instructions, and its role within our framework is clearly outlined. To address the question of generalizability, we conducted additional experiments replacing the CLIP text encoder with the BLIP model, as detailed in Appendix D.2. The results confirm that our framework is not restricted to CLIP and can adapt to other commonly used pretrained text encoders.

Regarding textual prompts, their primary purpose in our approach is to specify the type of obstacle to be removed, rather than providing complex or detailed scene descriptions. This allows for robust performance without requiring intricate prompts. Examples illustrating this simplicity and effectiveness can be found in Figure 8 and Figure 14.

We hope this clarification resolves the misunderstanding and highlights the robustness and flexibility of our approach. In light of this additional evidence and explanation, we kindly request the reviewer to reconsider their rating.