SeaS: Few-shot Industrial Anomaly Image Generation with Separation and Sharing Fine-tuning

Q1: The design of "Unbalanced Abnormal Text Prompt" is similar with previous works [1, 2], which compromises the novelty.

A1: Thanks for your comment! Here, we clarify the fundamental difference between the Unbalanced Abnormal (UA) Text Prompts in SeaS and existing prompts. This difference is one of the key reasons behind the improvement of SeaS. Below, the existing prompts are briefly outlined:

Textual Inversion [1] leverages the following prompt: "A photo of $S_*$" where $S_*$ is a placeholder string, representing the new concept we wish to learn. This approach only learns one concept at a time, which is not sufficient for AIG tasks that require differentiation between anomalies and products. We added the experiment of only using the Textual Inversion (TI) method to learn the product, which can be shown in Fig. 30. Due to the limited number of learnable parameters, the TI method struggles to generate images similar to the real product.

PromptAD [2] leverages the following prompt: \begin{eqnarray}s^n & = & [P_1][P_2] . . . [P_{E_N} ][obj.] \end{eqnarray} \begin{eqnarray}s^m & = & [P_1][P_2] . . . [P_{E_N}][obj.][with][color][stain] \end{eqnarray} \begin{eqnarray}s^l & = & [P_1][P_2] . . . [P_{E_N} ][obj.][A_1] . . . [A_{E_A}]\end{eqnarray} where $E_N$ denotes the length of the learnable normal prefix in the learnable normal prompt and manual anomaly prompt, "[with][color][stain]" denotes anomaly suffixes from the anomaly labels of the datasets, and $E_A$ denotes the length of learnable anomaly suffix in learnable anomaly prompt. PromptAD incorporates predefined semantic words such as “bottle” and “color stain”. While this method incorporates anomaly labels, it relies on fixed, generic semantic terms that may not align well with a few training images that contain specific anomaly types.

In contrast, our UA Text Prompts approach allows the product and anomaly to be independently learned through embeddings, making it more flexible and capable of handling a wider variety of anomalies effectively. The unbalanced design is not just a structural difference, but is crucial for separately learning common characteristics of products and diverse characteristics of anomalies.

We hope this explanation clarifies the novelty and advantages of our approach. Thank you again for your feedback!

Q2: While the paper demonstrates strong performance on MVTec datasets, it is unclear how well the SeaS method generalizes to other datasets or different types of unseen anomalies during training.

A2: Thank you for the comment! As we mentioned in Lines 22-24 in the paper. "Recombining the decoupled attributes may produce anomalies that have never been seen in the training dataset". Such a conclusion can be observed in Fig. 3(c) in the paper, and unseen samples in the same category can be handled efficiently.

For the generalization to other datasets, as we mentioned in lines 407-409 in the paper, SeaS demonstrates strong performance on the RGB images of the MVTec 3D AD dataset. The MVTec 3D AD dataset contains some common challenges, such as lighting condition variations, product pose variations, and so on, which are very important to validate the generalizability and robustness of our method. Additionally, we have added experimental evaluations on the VisA dataset in the General Response.

Q3: The paper should provide more details on the computational complexity, as the two training processes.Please report the time cost of the proposed method.

A3: Thanks for your comment. As we mentioned in Appendix A.3 in the paper, SeaS trains a generation model on a single NVIDIA Tesla A100 40G GPU, which only requires about 20G GPU memory. 

Comparisons of computational resource costs are given in Tab. 6 and Tab. 7 of the appendix. For the entire MVTecAD datasets with 73 anomaly types, our training takes 73 hours totally, which is only 30% of the time of the AnomalyDiffusion (249 hours) and only 18% of the time of DFMGAN (414 hours). In terms of inference time, SeaS costs 720 ms per image, which is only 19% of the 3830 ms per image required by the diffusion-based method AnomalyDiffusion.

Thank you again for your detailed response. Although I still have doubts about this type of work, considering the author's positive response and extensive experimentation, I will increase my score to 6.

Q1: The examples provided in the article primarily focus on straightforward instances of image generation. They do not explore more complex scenarios, such as the cable swap and cable missing categories found in the MVTec dataset. Are there specific challenges or limitations when generating these complex anomalies? If the method has been tested with these examples, presenting those results could significantly strengthen the paper. If not, outlining potential strategies for addressing these complexities would be advantageous.

A1: Thanks for your insightful comment! There are no specific challenges or limitations for SeaS to generate these complex anomalies. The results on cable_swap and missing_cable were given in Fig. 13 of Appendix A.6 in the paper and Tab. R8 in the rebuttal. SeaS performed well on them.

The reason is that the self-attention mechanism in U-Net enables the model to effectively capture the structural information of products. During fine-tuning with abnormal images, the model leverages self-attention to learn the relationship between anomaly regions and normal regions, as well as their impact on the overall product appearance. Complex anomalies, such as logic anomalies (e.g., "cable swap" or "cable missing"), present challenges due to masks that often mark empty areas or lack valid texture information. U-Net's self-attention is essential for achieving a global receptive field, facilitating the generation of these complex anomalies. Relying solely on anomaly tokens makes it difficult to associate tokens with anomalies, reducing generation effectiveness.

Table R8. The generation results on the complex anomalies on MVTec AD. ||IS|IC-LPIPS| |:---:|:---:|:---:| |cable_swap|2.21|0.43| |missing_cable|2.60|0.38|

Q2: The SeaS framework, while innovative in its application of few-shot learning for industrial anomaly image generation, does not fundamentally differ from the AnomalyDiffusion framework. Both aim to tackle the challenge of generating diverse and realistic anomaly images from limited data. However, SeaS introduces distinctive mechanisms, such as the Unbalanced Abnormal Text Prompt and the Separation and Sharing Fine-tuning strategy. These innovations give SeaS an edge in generating anomalies with greater fidelity and diversity. Nonetheless, both frameworks share a common goal of enhancing anomaly generation capabilities, indicating a gradual evolution in the field of generative models for industrial applications rather than a revolutionary shift.

A2: Thanks for your comment! Despite their shared goal of generating diverse and realistic anomaly images from limited data, SeaS introduces several key innovations that differentiate it from AnomalyDiffusion, achieving several advantages.

Generating products and anomalies at the same time. AnomalyDiffusion is essentially an image editing algorithm that does not require generating products. SeaS generates products and anomalies at the same time, which additionally ensures the uniqueness of the product in each output.

No mask input. SeaS does not require mask input, which reduces the cost of obtaining input data. However, AnomalyDiffusion requires a mask along with the RGB image, which may cause concerns about the misalignment between products and masks.

Q3: The paper positions SeaS as a superior approach compared to AnomalyDiffusion. What are the key factors that make SeaS more effective in certain aspects of industrial anomaly image generation? Are there scenarios in which AnomalyDiffusion may still offer advantages? A detailed comparative analysis or a side-by-side case study featuring both frameworks could clarify their respective strengths and weaknesses.

A3: Thanks for your comment! We clarify the key differences between SeaS and AnomalyDiffusion in terms of their overall framework, and application scenarios.

Overall Framework. As highlighted in Fig. 1 of the paper, the key difference between AnomalyDiffusion and SeaS is that AnomalyDiffusion is an editing approach, that needs the mask as input, and edits the anomalies onto a given normal product image based on the given mask, the SD is fixed. While SeaS is a fine-tuning approach, during training, it fine-tunes the U-Net in SD by using a few anomaly images. The ability to generate diverse anomalies and accurate alignment of anomalies and masks makes SeaS more effective for industrial anomaly image generation.

Application Scenarios. While SeaS is more versatile for generating a wide variety of anomaly images and is well-suited for training downstream segmentation models, AnomalyDiffusion may still be the preferred choice when the goal is precise anomaly editing. For example, if a user needs to insert a specific defect at a certain location on a product image, AnomalyDiffusion is better suited for this task due to its image editing framework.

Q1: The current fixed anomaly token setup (N=4) does not allow explicit control over specific types of anomalies, particularly in mixed anomaly cases (e.g., mixed anomalies in cable). This limitation reduces the flexibility of the model in generating targeted anomaly types. Is this interpretation correct, or is there a way to dynamically adjust or control anomaly types within this setup? Further clarification on this point would be very helpful.

A1: Thanks for your meticulous comment! This interpretation needs to be recorrected as below:

Our method allows for explicit control over specific types of anomalies through the use of different condition text prompts, denoted as $\mathcal{P_n}$. As we mentioned in lines 292-293 in the paper, these prompts are designed with different sets of anomaly tokens, which guide the training and generation of each anomaly type. The prompt is defined as:

$\mathcal{P_n} = \text{a <ob> with } \text{<}\text{df}_{4 × n-3}\text{>,}\text{<}\text{df}_{4 × n-2}\text{>,} \text{<}\text{df}_{4 × n-1}\text{>,} \text{<}\text{df}_{4 × n}\text{>}$

where $n$ represents the index of the anomaly types in the product. For instance, for the "cable" product, the prompts would be $\mathcal{P_1} = \text{a <ob> with } \text{<}\text{df}_{1}\text{>,}\text{<}\text{df}_{2}\text{>,} \text{<}\text{df}_{3}\text{>,} \text{<}\text{df}_{4}\text{>}$ for the training and generation of the first anomaly type "bent_wire", and $\mathcal{P_2} = \text{a <ob> with } \text{<}\text{df}_{5}\text{>,}\text{<}\text{df}_{6}\text{>,} \text{<}\text{df}_{7}\text{>,} \text{<}\text{df}_{8}\text{>}$ for the second anomaly type "cable_swap". For the "combined" anomaly category of "cable", the training and inference methods are the same as for other types of anomalies. Thus, our method provides flexibility in adjusting and controlling the specific types of anomalies by tailoring the prompt set accordingly.

Q2:Unlike AnomalyDiffusion [1], which leverages the text inversion technique [2] without needing to fine-tune the U-Net, this method requires U-Net fine-tuning. This approach might reduce the domain gap for anomaly data, potentially contributing to the improved quality of generated images. However, it would be helpful to clarify the effect of freezing versus fine-tuning the U-Net on image quality. Since this method requires fine-tuning the U-Net while AnomalyDiffusion [1] uses text inversion [2] without U-Net adjustments, could the authors clarify whether the observed quality improvement in anomaly generation stems from this fine-tuning? A comparison between freezing and fine-tuning the U-Net would be insightful.

A2: Thanks for your insightful comment! We agree with your opinion that "Fine-tuning the U-Net contributes to the improved quality of generated images". Fine-tuning the U-Net is essential for the AIG method.

To investigate the role of fine-tuning, we conducted the experiment of only using the Textual Inversion (TI) method to learn the product without fine-tuning the U-Net. The results, shown in Fig. 30 in the appendix, demonstrate that the TI method struggles to generate images similar to the real product due to the limited number of learnable parameters. In contrast, when fine-tuning the U-Net as part of the AIG method, the generated images exhibit significant improvements in both quality and global consistency. This suggests that the observed improvement in image quality and anomaly generation can largely be attributed to the fine-tuning of the U-Net.

As we mentioned in Lines 83-84 of the paper, "the products satisfy global consistency with minor variations in local details, while the anomalies hold randomness," the generated products should be globally consistent with the real products. Therefore, unlike the AG method Anomalydiffusion, which does not need to consider the generation of the product, we fine-tune the U-Net.

Q3: The inference process in this paper lacks clarity, making it difficult to understand how to specify a particular anomaly type. Could the authors provide additional details on the inference process, specifically on how one could specify or target a particular anomaly type during generation?

A3:  Thanks for your comment! During inference, we use different condition text prompts $\mathcal{P_n}$ to specify the particular anomaly type of the product. We explained the design of different text prompts in Response A1. For example, aiming to generate "cable_swap", which is the second anomaly type for the "cable" product, we have to set the condition prompt as  $\mathcal{P_2} = \text{a <ob> with } \text{<}\text{df}_{5}\text{>,}\text{<}\text{df}_{6}\text{>,} \text{<}\text{df}_{7}\text{>,} \text{<}\text{df}_{8}\text{>}$ during inference.

Q1: The word “Unbalanced” in Unbalanced Abnormal Text Prompt is inappropriate. The author believes fixed generic semantic words may fail to align with a few training images that contain specific defect types. Therefore, the text prompt should be expressed as dynamic or learnable, etc.

A1: Thanks for your comment! As we mentioned in lines 214-215 of the paper, "unbalanced" means using multiple anomaly tokens while only employing one normal token, rather than referring to the learnability of the prompt.

Q2: The author should add the results of Baseline in Table 4, such as generation of typical text prompt with fixed generic semantic words.

A2: Thanks for your comment! We included the baseline result (short for with TP ) in Tab. 4 and Tab. R5 in the rebuttal. The results demonstrate that using a typical text prompt with fixed generic semantics words leads to a decrease in the fidelity and diversity of the generated images, as well as in segmentation performance. This outcome proves the effectiveness of the text prompt we proposed.

Table R5. More ablation on the generation model. ||IS|IC-L|AUROC|AP|F1-max|IoU| |:---:|:---:|:---:|:---:|:---:|:---:|:---:| |with TP|1.72|0.33|94.72|57.16|55.67|50.46| |Ours|1.88|0.34|97.21|69.21|66.37|55.28|

Q3: Moreover, the result with different layers in UNet in RMP branch should be discussed.

A3: Thanks for your comment! We have provided the results with different layers of U-Net in the RMP branch in Table 25 of Appendix A.8. Our experiments show that "employing a combination of {F2,F3} for coarse feature extraction, achieves the best performance in downstream segmentation task", which explains our selection of these layers for the RMP branch.

Q4: The authors should replace the abnormal synthesis strategy in the existing unsupervised methods with the proposed generation strategy to prove the effectiveness of the proposed method, such as DRAEM, because the methods in the table 2 are not specifically designed for anomaly detection.

A4: Thanks for your comment! We follow your instructions to replace the abnormal synthesis strategy in DRAEM with the proposed generation strategy, the results are given in Tab. R6. The segmentation result demonstrates that our method outperforms DRAEM, with improvements of +9.46 % in AP, +7.23% in F1-max and +3.69% in IoU.

Table R6. Comparison with unsupervised method on MVTec AD. ||image-level|||pixel-level||||| |:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:| ||AUROC|AP|F1-max|AUROC|AP|F1-max|PRO|IoU| |DRAEM|98.00|98.45|96.34|97.90|67.89|66.04|91.49|60.30| |SeaS|99.25|99.66|98.35|97.98|77.35|73.27|93.79|63.99|

Q5: Implementation details of inference should be described, such as inference step and guidance scale, which is critical to the quality of the generation.

A5: Thanks for your comment! For all experiments, we use t = 1500 to perform diffusion forward on normal images to get the initial noise. We employ T = 25 steps of sampling with a classifier-free guidance strength set to 8. To generate anomalies of a specific category, we use the same condition corresponding to that specific category of anomalies during training.

Q6: The latest generation method [1] should be compared.

A6: Thanks for your comment! Following your instructions, we substituted the abnormal synthesis strategy in AnoGen [1] with the proposed generation strategy. The results, which are presented in Tab. R7, show that our method surpasses AnoGen, achieving improvements of +5.75% in AP, +4.09% in F1-max, and +3.07% in IoU.

Table R7. Comparison with the latest generation method on MVTec AD. ||image-level|||pixel-level||||| |:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:| ||AUROC|AP|F1-max|AUROC|AP|F1-max|PRO|IoU| |AnoGen|98.28|99.31|97.70|97.79|71.60|69.18|91.92|60.92| |SeaS|99.25|99.66|98.35|97.98|77.35|73.27|93.79|63.99|

Q7: More datasets, such as VisA or RealIAD, are suggested to used for further evaluation of the proposed method.

A7: Thanks for your suggestion! The new results on the VisA dataset are given in the General Response. Our method outperforms others across all the segmentation models, with an 11.71% average improvement on IoU.

Q1: Some SOTA anomaly detection methods should also be compared besides generative model-based anomaly detection methods, e.g., DiAD [1]. Although RealNet [2] is compared in appendix A.5, the proposed method does not significantly outperform RealNet.

A1: Thanks for your comment!

First, with the training samples generated by SeaS, the conventional lightweight segmentation models surpass existing SOTA anomaly detection methods, as shown in Tab. 11 and Tab. 12. Moreover, we added the comparison with the non-generative model-based anomaly detection method in Tab. 12, the result of "SeaS + UperNet" outperforms the DiAD [1] method on all metrics. For RealNet [2], as we mentioned in Lines 1177-1179 in Appendix A.5, "Real-Net contains 591M parameters, around 177 times larger than BiSeNet V2, and 631 times larger than LFD, while the pixel-level AP and IoU measures of Real-Net are even worse than those of BiSeNet V2 and LFD. Although the pixel-level AUROC metric, which is more sensitive to false negatives than to false positives, is slightly higher for Real-Net, we observe that it generates a high number of false positives, substantially reducing pixel-level AP, F1-max, and IoU scores. " For effective industrial anomaly detection, a method should be small-scale, and balance false positives and false negatives.

Additionally, in Tab. R10, we compare the results of training BiSeNet V2 using only the original data from MVTec AD with using the generated anomaly image-mask pairs from SeaS. The results show significant improvements across all metrics when using the proposed generation approach, further demonstrating that our method greatly enhances the performance of downstream tasks.

Table R10. Comparison with only using original training data on MVTec AD. ||image-level|||pixel-level|||| |:---:|:---:|:----:|:---:|:---:|:---:|:---:|:---:| ||AUROC|AP|F1-max|AUROC|AP|F1-max|IoU| |BiSeNet V2|68.89|80.51|82.22|70.35|23.02|28.72|15.96| |SeaS + BiSeNet V2|96.00|98.14|95.43|97.21|69.21|66.37|55.28|

Q2: Since the proposed method is not specific to the multi-class anomaly detection setting, I would also wonder about the comparison with SOTA methods in the single-class anomaly detection setting, especially supervised/semi-supervised methods, e.g., PRN [3] or BGAD [4], because anomaly samples are used during training of the proposed method.

A2: Thanks for your comment! In terms of the downstream supervised segmentation task, as we mentioned in Lines 475-476 in the paper, we train a single unified model for the segmentation task in a multi-class setting. While our method isn't specifically tailored for single-class anomaly detection, we have conducted comparisons with existing state-of-the-art methods and demonstrated that our model outperforms them.

To elaborate, since the official source code for PRN [3] is unavailable, we compared our method using the reported metrics from Tab. 2 and Tab. 3 of AnomalyDiffusion. We included this comparison in Tab. 12 of our paper, where our method shows superior performance compared to PRN. As for BGAD [4], we replaced its pseudo-anomaly generation strategy with our proposed generation approach. Experimental results, presented in Tab. R4, show that our approach outperforms BGAD as well.

Table R4. Comparison with supervised method on MVTec AD. ||image-level|||pixel-level|||| |:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:| ||AUROC|AP|F1-max|AUROC|AP|F1-max|IoU| |BGAD|98.31|98.05|98.97|99.26|73.85|77.89|60.60| |SeaS+BGAD|98.44|98.18|99.08|99.26|73.85|77.93|60.81|

Q3: More datasets are required to evaluate the proposed anomaly generation method on image AUROC, pixel AUROC and PRO, such as VisA dataset containing more diverse and tiny anomalies. It is more challenging to generate these tiny anomalies.

A3: Thanks for your suggestion! The new results on the VisA dataset are given in the General Response. Our method outperforms others across all the segmentation models, with an 11.71% average improvement on IoU.

Q1: The relationship between anomaly tokens and training different types of anomalies is not clear. I want to clarify the relationship between exception marking and training different types of exceptions.

A1: Thanks for your comment! As we mentioned in Lines 1433-1436 in Appendix A.8, we use 4 $<\text{df}>$ as a set for each anomaly type, and 1 $<{\text{ob}}>$ for each product class. Taking the product "cable" with 8 anomaly types as an example, there are 1 <ob> and 8 different token sets for $<\text{df} _ n>$ , i.e., {<\text{df} _ 1>$$<\text{df} _ 2>,$<\text{df} _ 3>$,$<\text{df} _ 4>$, {$<\text{df} _ 5>$,$<\text{df} _ 6>$,$<\text{df} _ 7>$,$<\text{df} _ 8>$},...,{$<\text{df} _ {29}>$,$<\text{df} _ {30}>$,$<\text{df} _ {31}>$,$<\text{df} _ {32}>$}. Each token set contains 4 tokens corresponding to one anomaly type of "cable".

Q2: The paper has not discussed how to control the type of exceptions generated during inference.

A2: Thanks for your insightful comment! During inference, for a particular type of anomaly, we use the Unbalanced Abnormal (UA) Text Prompt $\mathcal{P _ n}$ with different sets of anomaly tokens as the condition to generate the specified type of anomaly.

$\mathcal{P_n} = \text{a <ob> with } \text{<}\text{df}_{4 × n-3}\text{>,}\text{<}\text{df}_{4 × n-2}\text{>,} \text{<}\text{df}_{4 × n-1}\text{>,} \text{<}\text{df}_{4 × n}\text{>}$

where $n$ represents the index of the anomaly types in the product. For example, for the "cable" product, the prompts would be $\mathcal{P_1} = \text{a <ob> with } \text{<}\text{df}_{1}\text{>,}\text{<}\text{df}_{2}\text{>,} \text{<}\text{df}_{3}\text{>,} \text{<}\text{df}_{4}\text{>}$ for the training and generation of the first anomaly type "bent_wire", $\mathcal{P_2} = \text{a <ob> with } \text{<}\text{df}_{5}\text{>,}\text{<}\text{df}_{6}\text{>,} \text{<}\text{df}_{7}\text{>,} \text{<}\text{df}_{8}\text{>}$ for the second anomaly type "cable_swap".

Q3: The paper does not explain why the U-net used to predict noise in the Refined Mask Prediction branch has a highly discriminative feature.

A3: Thanks for your comment! During the fine-tuning, we use Decoupled Anomaly Alignment loss to align the anomaly embeddings with anomaly region features, while simultaneously preventing the normal region features from aligning with these anomaly embeddings. This approach ensures that the features of the anomaly region are distinguishable from those of the normal region. This can be observed in Fig. 29 of the appendix, we use the output features of the “up-2” and “up-3” layers in U-Net. After applying convolutional blocks and concatenation operations, these output features exhibit distinct responses between anomaly and normal regions, thereby demonstrating the high discriminability of the U-Net features.

Thanks to the author for your response! For Q1 I suggest putting them in the main text to make it easier for readers to understand. I have no more questions about Q2.

Also in Q3 I would like you to use more words to explain why such an approach does lead to "more discrimination" than before. In other words, I understand that you want to create "more discriminative feature representations, especially between normal and abnormal features". However, I would like you to re-emphasize the motivation of your model in this regard, especially why your model does better in this regard, to help the reader understand.

A few others that don't matter:

1. Some of the images in the appendix may affect the visual impression, for example, the sample selection in Fig.7 is not correct, perhaps we would prefer to see different results of each model on the same sample.
2. We know that the depth map is a single-channel grayscale image, but we can actually stack it into a 3-channel image. I am curious what effect will be produced on the depth map of MVTec3D-AD? Are RGB and depth maps similar for the same sample? This is important for multimodal anomaly detection. Maybe you can show a few examples in the appendix to help people understand.

Good luck again!

I apologize for delaying the discussion due to my personal reasons, and I hope you will reply as soon as possible.

Q6: In addition to the concerns we mentioned earlier, you may need to consider more about the control of the type of anomalies generated by the model. For example, the idea of "generating unprecedented anomalies by recombining" is not enough just through some visualization, we would like to know where the bottom line of this model is. Because even though the variety of existing anomalies to synthesize is often limited, we may need more examples to support your idea of generating more "invisible new anomalies".

A6: Thank you for your meticulous comment! We fully acknowledge the importance of controlling the types of anomalies generated by the model. In future work, we plan to explore attribute decoupling and recombination principles more deeply, with the goal of enabling precise control over individual anomaly attributes such as color, shape, and texture.

To clarify the potential of generating "invisible new anomalies", we have provided additional examples in Fig. 32 of the appendix, where the model generates anomalies that differ significantly from the training data in terms of both color and shape. For instance, we present new examples such as bottle_contamination, hazelnut_print, and tile_gray_stroke, which introduce novel shapes. We also show wood_color and metal_nut_scratch, which differ in color, and pill_crack, which introduces a new shape with multiple cracks, whereas the training samples contained only single cracks. These examples illustrate the model’s capacity to generate unseen anomalies by recombining learned attributes in innovative ways, going beyond the limitations of the original anomaly diversity in the training set.

Thank you for your answer. I'm glad you answered or practiced Q1 and Q4. I still recommend writing more clearly and explaining the motivation, including putting Q3 into the body of the text and describing it in more detail (although this is a suggestion that does not affect the rating). On the other hand, for Q5 I am just a suggestion, maybe the model can adapt to the two modalities of RGBD, and the generated exceptions can correspond to one of the challenges that our field needs to face.

I may have some objections to Q6, just as you said, "recombining way" to generate exceptions, which may cause some misunderstandings. For example, this way is easy to be misunderstood as the weighted generation of existing exceptions, I hope to explain further in the appendix.

There are also some simple questions: As other reviewers have said, although you perform well on some anomaly generation metrics, your main task still needs to focus on anomaly detection. Your performance on some anomaly detection metrics is not so impressive. Maybe you need to balance the two points more.

Based on the 8 score I've improved, I stay the same.