Improving Art Style Classification Through Data Augmentation Using Diffusion Models

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors apply StyleGAN, an already existing and widely used GAN model for synthetic paintings generation. However, they do not test the GAN's convergence and they do not get into details about the GAN's training. In addition, they do not consider several studies in this area, that utilize GAN architectures for painting generation. It is not clear why ResNet-50 is the only model that is trained and why the authors did not consider other models too. As minor comments, equations 12-15 should be rewritten for the multiclass classification and some labels in the confusion matrices are trimmed.

Author Response

Manuscript ID: electronics-3341862

Type of manuscript: Article

Title: Improving Art Style Classification through Data Augmentation Using Generative Adversarial Networks

Authors: Miguel Ángel Martín-Moyano, Iván García-Aguilar, Ezequiel López-Rubio, Rafael M. Luque-Baena *

 

ANSWERS TO REVIEWERS

We thankfully acknowledge the reviewers for their constructive comments, which have indeed helped us to improve the paper. The reviewer’s concerns have been addressed and the paper has been deeply reviewed.

Yours faithfully,

The authors

 

1. The authors apply StyleGAN, an already existing and widely used GAN model for synthetic paintings generation. However, they do not test the GAN's convergence and they do not get into details about the GAN's training.

 

Answer:

 

To address this, a new subsection (4.3.2) has been included in the manuscript to provide a detailed explanation of the training process and convergence assessment for the StyleGAN model. The following paragraphs have been added:

 

"The training process of the StyleGAN network focused on ensuring the model's convergence and generating images of sufficient quality for the intended classification task. The training process was carefully monitored to achieve this, paying particular attention to the adversarial dynamics between the generator and the discriminator. The generator was adjusted iteratively to minimize its loss based on feedback from the discriminator, while the discriminator improved its ability to distinguish real from synthetic images.

 

Convergence was assessed primarily through qualitative analysis of the generated images. Throughout the training, periodically generated images were evaluated to identify visual artifacts and ensure alignment with the intended styles. This process included checking for consistency in texture, color, and composition and the absence of mode collapse or repetitive patterns, which are common challenges in GAN training.

 

The generator and discriminator were optimized jointly under the adversarial framework, with the discriminator providing iterative feedback to improve the generator’s outputs. A careful adjustment of hyperparameters, such as learning rate and batch size, was performed to stabilize the training further and prevent divergence. Specifically, an initial learning rate of $10^{-4}$ was used to maintain stability during early training stages.

 

The stylistic coherence and visual diversity of the generated images were key indicators used to confirm the adequacy of the trained model. This qualitative assessment ensured that the generated images were sufficiently representative for their use in augmenting the dataset and supporting the classification task."

2. In addition, they do not consider several studies in this area, that utilize GAN architectures for painting generation.

 

Answer: 

 

The introduction has been expanded to include a review of relevant studies on GAN-based painting generation. These references provide a broader context for the application of GANs in artistic domains. This expanded discussion situates the study within the broader literature on GAN-based artistic style generation. The following paragraphs have been added:

 

"The application of Generative Adversarial Networks (GANs) in painting generation has marked a significant milestone in artificial intelligence and art. These architectures enable the creation of synthetic artworks that emulate recognized artistic styles and explore new forms of digital creativity. Several authors have explored the use of GANs in synthesizing traditional paintings in various domains. For example, CA-GAN introduces an architecture that merges attributes and content to generate high-quality works, verifying consistency through a discriminator based on a cross-cycle consistency constraint \cite{Chen2023}. Similarly, SAPGAN proposes an end-to-end model that uses two networks (SketchGAN and PaintGAN) to generate Chinese landscapes, from edge maps to complete paintings \cite{xue2020endtoendchineselandscapepainting}. Likewise, the LMGAN model implements progressive generation with memory modules to capture essential features and create high-quality Chinese landscapes \cite{LMGAN}.

 

Regarding style transfer and abstraction, DLP-GAN and RPD-GAN are examples of architectures designed to translate images from one domain to another, whether transforming classical landscapes into modern styles or generating realistic paintings through unsupervised approaches \cite{Gui2023, 9179994}. Seg-CycleGAN presents improvements to the CycleGAN model for abstract art generation, optimizing the creation of abstract paintings from initial data \cite{10167194}.

 

Additionally, some studies have focused on generating specific styles, such as Stroke-GAN, which learns stroke styles from various artists to emulate human painting techniques with high fidelity \cite{Wang2023}. Similarly, methods have been proposed to translate hand-drawn sketches into abstract or realistic art using architectures such as Pix2Pix and CycleGAN, expanding creative possibilities in this field \cite{gao}. Moreover, more conceptual studies, such as Creativity and Style in GAN and AI Art, address the historical and cultural impact of GANs on art, exploring questions of originality and creativity in AI-assisted production \cite{Berryman2024}."

3. It is not clear why ResNet-50 is the only model that is trained and why the authors did not consider other models too.

 

Answer: 

 

The manuscript has been updated to clarify the rationale for selecting ResNet-50 and to emphasize that the proposed methodology is not limited to this architecture. This update underscores the flexibility of the methodology and the experimental rationale behind the use of ResNet-50.

 

“The selection of ResNet-50 as the classification model in this study to evaluate the proposed methodology is based on its proven effectiveness in image classification tasks, including artistic style classification. ResNet-50 is a widely used architecture in the literature due to its ability to capture complex visual features, facilitated by its residual block design, which allows for the efficient training of deep networks. This model combines precision and computational efficiency, making it an ideal choice for exploring the impact of GAN-based data augmentation in the proposed context.

 

However, it is important to highlight that the methodology presented in this work is not exclusively tied to ResNet-50 but constitutes a meta-method applicable to other classification architectures. The proposed approach is designed to enhance the performance of any model reliant on a large and diverse dataset. This implies that other architectures could similarly benefit from the GAN-based data augmentation implemented in this study. The selection of ResNet-50 was made to establish a clear and consistent experimental baseline for evaluating the effectiveness of the data augmentation method."

4. As minor comments, equations 12-15 should be rewritten for the multiclass classification and some labels in the confusion matrices are trimmed.

 

Answer: 

The equations (12–15) have been revised to align with the multiclass classification framework. Additionally, the confusion matrices have been updated to ensure that all labels are fully displayed, addressing the reported issue.

 

- Several improvements have been made throughout the manuscript to enhance its clarity and depth. New references have been added to situate the study within existing literature. Additionally, the description of the methodology has been expanded, and the discussion section has been enriched with more detailed analyses and comparisons. These updates provide a more comprehensive understanding of the study's contributions and findings.

Reviewer 2 Report

Comments and Suggestions for Authors

The paper explores the application of Generative Adversarial Networks for data augmentation in the context of art style classification, presenting a novel methodology to address the challenges posed by limited and imbalanced datasets. By generating synthetic artworks with StyleGAN and integrating them with real datasets, the authors aim to enhance the generalization capabilities of deep learning models, specifically ResNet-50. The results demonstrate that combining real and synthetic data improves classification accuracy and robustness, particularly for certain artistic styles. The paper provides a clear framework for integrating GANs into art style classification and highlights the potential for such methods to overcome data limitations in visual domains. Overall, this study represents a significant step forward in applying advanced machine learning techniques to art classification and is recommended for publication, with consideration of the noted areas for refinement.

1. The style names in the confusion matrix are incomplete, please fix it.

2. Did the generated synthetic images successfully introduce unique stylistic features that were not present in the original dataset? If so, what methods or metrics were used to quantify these differences, such as measures of image diversity, style variation, or distribution shifts in feature space? Was there an evaluation of whether these new features meaningfully expanded the representation of underrepresented styles in the dataset?

3. How does the balance between the real and synthetic data for each artistic style affect the classification model's performance? Were analyses performed to determine whether certain styles were overrepresented or underrepresented in the augmented dataset? If imbalances were present, how did they influence model predictions, and were steps taken to mitigate potential biases introduced by synthetic data generation?

4. Did the inclusion of synthetic data pose risks of overfitting, especially if the generated images were overly similar to the real data or exhibited repetitive patterns? Were any diagnostic tests or visual inspections conducted to detect signs of overfitting, such as reduced generalization performance on unseen data or the model’s reliance on synthetic-specific features? How were these risks addressed, if identified?

Author Response

Manuscript ID: electronics-3341862

Type of manuscript: Article

Title: Improving Art Style Classification through Data Augmentation Using Generative Adversarial Networks

Authors: Miguel Ángel Martín-Moyano, Iván García-Aguilar, Ezequiel López-Rubio, Rafael M. Luque-Baena *

 

ANSWERS TO REVIEWERS

We thankfully acknowledge the reviewers for their constructive comments, which have indeed helped us to improve the paper. The reviewer’s concerns have been addressed and the paper has been deeply reviewed.

Yours faithfully,

The authors

 

1. The style names in the confusion matrix are incomplete, please fix it.

 

Answer:

 

The confusion matrices have been revised to ensure that all style names are fully displayed. 

 

2. Did the generated synthetic images successfully introduce unique stylistic features that were not present in the original dataset? If so, what methods or metrics were used to quantify these differences, such as measures of image diversity, style variation, or distribution shifts in feature space? Was there an evaluation of whether these new features meaningfully expanded the representation of underrepresented styles in the dataset?

Answer: 

 

To address this point, the manuscript has been expanded to include an analysis of the impact of synthetic data on dataset diversity and classification performance. This addition clarifies how the generated synthetic images contributed unique stylistic features and describes the methods used to evaluate their impact on dataset diversity and classification performance. The following content was added in the section 4.3.4:

 

"To evaluate the impact of the synthetic images, performance metrics such as accuracy, adjusted Rand Score, and the confusion matrix were analyzed. These metrics provided indirect insights into the contributions of the synthetic data, revealing improvements for underrepresented styles and challenges in distinguishing visually similar styles such as Impressionism and Expressionism.

 

To address potential risks of overfitting due to the inclusion of synthetic data, a balanced dataset approach was adopted. Synthetic images were generated uniformly across styles (300 per style) to prevent any single style from dominating the training set. The model’s generalization ability was also assessed using validation and test datasets of real images. The results indicated no significant discrepancies between training and validation performance, suggesting that the model maintained its generalization ability.

 

These measures ensured that the synthetic data complemented the real dataset without introducing significant biases or overfitting risks, supporting improved classification performance for underrepresented styles while maintaining diversity and stylistic coherence."

3. How does the balance between the real and synthetic data for each artistic style affect the classification model's performance? Were analyses performed to determine whether certain styles were overrepresented or underrepresented in the augmented dataset? If imbalances were present, how did they influence model predictions, and were steps taken to mitigate potential biases introduced by synthetic data generation?

 

Answer: 

 

To address this point, the manuscript now explicitly discusses the balance between real and synthetic data for each artistic style and its effect on classification performance. This addition highlights the measures taken to ensure dataset balance and their effect on mitigating potential biases in the augmented dataset. The following content was added in the section 4.3.5:

 

"The balance between real and synthetic data for each artistic style significantly affected the classification model's performance. Mixing synthetic and real data ensured that the training dataset maintained uniformity across styles, reducing the risk of overrepresentation or underrepresentation for any specific style. This strategy yielded the best results, indicating that a balanced combination of real and synthetic data positively influenced the model's generalization ability."

 

4. Did the inclusion of synthetic data pose risks of overfitting, especially if the generated images were overly similar to the real data or exhibited repetitive patterns? Were any diagnostic tests or visual inspections conducted to detect signs of overfitting, such as reduced generalization performance on unseen data or the model’s reliance on synthetic-specific features? How were these risks addressed, if identified?

Answer: 

The manuscript now addresses potential overfitting risks associated with synthetic data. This explanation demonstrates the precautions taken to minimize overfitting risks and highlights the validation strategies employed. The following content was added in the section 4.3.4:

"To evaluate the impact of the synthetic images, performance metrics such as accuracy, adjusted Rand Score, and the confusion matrix were analyzed. These metrics provided indirect insights into the contributions of the synthetic data, revealing improvements for underrepresented styles and challenges in distinguishing visually similar styles such as Impressionism and Expressionism.

To address potential risks of overfitting due to the inclusion of synthetic data, a balanced dataset approach was adopted. Synthetic images were generated uniformly across styles (300 per style) to prevent any single style from dominating the training set. The model’s generalization ability was also assessed using validation and test datasets of real images. The results indicated no significant discrepancies between training and validation performance, suggesting that the model maintained its generalization ability.

These measures ensured that the synthetic data complemented the real dataset without introducing significant biases or overfitting risks, supporting improved classification performance for underrepresented styles while maintaining diversity and stylistic coherence."

Reviewer 3 Report

Comments and Suggestions for Authors

In this work, the authors propose a novel data augmentation approach using Generative Adversarial Networks (GANs) in contrast to traditional augmentation techniques. The method generates new samples based on the existing data, expanding the available dataset and enhancing the generalization capability of classification models. They evaluate the effectiveness of this data augmentation technique by training deep learning models with varying proportions of augmented and real data and assessing their performance in pictorial style classification. Their results demonstrate that the proposed GAN-based augmentation significantly improves classification accuracy, suggesting that it can be a viable solution for overcoming data limitations in similar applications. While the study is promising and demonstrates some advancements, it requires some revisions to improve clarity, rigor, and impact.

·       The paper lacks detailed explanation for several steps, like the details of the existing data augmentation techniques. Authors are suggested to provide the details.

·       It would be beneficial to elaborate further on why StyleGAN was specifically chosen over other GAN models. Including a brief comparison with alternative models could provide more justification for this choice.

·       The authors should add a brief statement on how Wasserstein loss function impacts the results or improves the image quality.

·       Extend evaluations to include more diverse datasets to demonstrate robustness and generalizability.

·       Add the discussion in the introduction section that how the proposed data augmentation method will work on the other dataset like medical imaging etc. For instance, discuss the use of the proposed method for the following works: A deep learning and handcrafted based computationally intelligent technique for effective COVID-19 detection from X-ray/CT-scan imaging,  Localization and classification of gastrointestinal tract disorders using explainable AI from endoscopic images, ExpressionHash: Securing Telecare Medical Information Systems Using BioHashing

·       The discussion section is a critical part of the manuscript where the significance of the findings is thoroughly examined in the context of existing literature. I suggest extending this section to provide a deeper analysis and interpretation of the results.

 

 

 

 

 

Author Response

Manuscript ID: electronics-3341862

Type of manuscript: Article

Title: Improving Art Style Classification through Data Augmentation Using Generative Adversarial Networks

Authors: Miguel Ángel Martín-Moyano, Iván García-Aguilar, Ezequiel López-Rubio, Rafael M. Luque-Baena *

 

ANSWERS TO REVIEWERS

We thankfully acknowledge the reviewers for their constructive comments, which have indeed helped us to improve the paper. The reviewer’s concerns have been addressed and the paper has been deeply reviewed.

Yours faithfully,

The authors

 

1. The paper lacks detailed explanation for several steps, like the details of the existing data augmentation techniques. Authors are suggested to provide the details.

 

Answer:

 

The manuscript initially contained an error suggesting the use of data augmentation techniques beyond real and GAN-generated images. This has been corrected in the methodology section, clarifying that only real and GAN-generated images were used for training the model. No additional data augmentation techniques were applied in this study.

2. It would be beneficial to elaborate further on why StyleGAN was specifically chosen over other GAN models. Including a brief comparison with alternative models could provide more justification for this choice.

Answer: 

 

The manuscript has been revised to provide additional clarity and detail regarding the selection of StyleGAN over other GAN frameworks such as DCGAN, Pix2Pix, and CycleGAN. A new paragraph has been added to explain the rationale for this choice, emphasizing the advantages of StyleGAN for this specific application. This addition provides a clear comparison and justification for the use of StyleGAN, enhancing the clarity of the methodology. The following content was added in the section 3.1:

 

"Generative Adversarial Networks (GANs) constitute a fundamental process for creating an expanded dataset that enhances the performance of the pictorial-style classifier. For this purpose, the StyleGAN model \cite{9156570} was selected due to \textcolor{blue}{its advanced capabilities for generating high-quality, diverse, and stylistically consistent images compared to other generative models such as DCGAN \cite{dcgan}, Pix2Pix \cite{pix2pix}, and CycleGAN \cite{8237506}. Unlike DCGAN, which primarily focuses on generating general-purpose images and lacks explicit control over the visual styles of the generated images, StyleGAN incorporates a style-based architecture that allows detailed control over visual attributes. This made it advantageous in this context, where preserving specific stylistic elements such as texture, composition, and color palette was essential. Compared to Pix2Pix, designed for paired image-to-image translation tasks and requires a dataset of aligned image pairs, StyleGAN operates in an unpaired setting, making it more suitable for this application. The absence of aligned datasets for most artistic styles renders Pix2Pix impractical for generating synthetic images in this context. CycleGAN excels in unpaired image-to-image translation, allowing the conversion of images between domains without requiring aligned pairs. However, it focuses on domain translation rather than generating new, diverse samples within a specific domain. This limits its applicability when the objective is to expand a dataset with novel samples rather than transforming existing ones. Moreover, CycleGAN often struggles with generating highly diverse and stylistically consistent images within a single domain.

 

For all these reasons, StyleGAN was the model selected, as it allows the generation of highly diverse images while maintaining coherence within a given style. This level of control and quality is critical for generating synthetic data that effectively complements the dataset without introducing artifacts or inconsistencies that could negatively impact the performance of the classification model, making it the most suitable option for this study."

3. The authors should add a brief statement on how Wasserstein loss function impacts the results or improves the image quality.

 

Answer: 

 

A detailed explanation of the Wasserstein loss function has been added to make the methodology accessible to readers unfamiliar with modern GAN techniques. This explanation clarifies the purpose and benefits of using the Wasserstein loss function, particularly in the context of GAN training stability and output quality. The following content was added in the section 3.1:

"The Wasserstein loss function was designed to address common challenges in training traditional GANs, such as instability and mode collapse. Unlike the cross-entropy-based loss, which classifies individual samples as real or generated, the Wasserstein loss is based on the Wasserstein distance. This metric measures how much "work" is required to transform the distribution of generated data into the distribution of real data. This approach quantifies the difference between real and generated data distributions by comparing their statistical properties rather than focusing on individual classifications. The key advantage of the Wasserstein loss lies in its ability to provide a smoother and more stable optimization process. Unlike traditional losses that can saturate when generated data deviates significantly from the real distribution, the Wasserstein loss varies continuously even in such cases. This stability mitigates training challenges and ensures that the generator receives meaningful gradients throughout the process. By stabilizing the training dynamics, the Wasserstein loss enables the GAN to converge to a more balanced state, where the generator produces images that closely resemble the target distribution. This leads to improved visual quality, with generated images exhibiting finer details, better texture consistency, and higher stylistic fidelity."

 

4. Extend evaluations to include more diverse datasets to demonstrate robustness and generalizability.

 

Answer: 

The inclusion of additional datasets was not feasible within the scope of this study due to the limited availability of suitable datasets for training GAN models on artistic styles and the short review timeframe. Expanding evaluations with more diverse datasets would require extensive preprocessing and alignment with the methodology, which was beyond the practical constraints of this revision process. However, this limitation is acknowledged in the manuscript, and future work will explore the application of the proposed methodology to other datasets.

 

5. Add the discussion in the introduction section that how the proposed data augmentation method will work on the other dataset like medical imaging etc. For instance, discuss the use of the proposed method for the following works: A deep learning and handcrafted based computationally intelligent technique for effective COVID-19 detection from X-ray/CT-scan imaging,  Localization and classification of gastrointestinal tract disorders using explainable AI from endoscopic images, ExpressionHash: Securing Telecare Medical Information Systems Using BioHashing

 

Answer: 

The introduction has been expanded to include a discussion of how the proposed methodology could benefit other fields, such as medical imaging and biometric security. The referenced studies were cited, and their relevance to the methodology was explained. The following paragraphs has been included:

"The proposed methodology could have significant potential for broader applications in other fields where data scarcity and variability are key challenges. In medical imaging, synthetic data generated through GANs can provide a robust solution to augment datasets, addressing critical limitations in disease detection, anomaly classification, and secure data management. In the context of disease detection, studies such as \cite{paper1} highlight the challenges posed by the limited availability of labeled medical imaging data. The proposed GAN-based approach could be adapted to generate synthetic X-ray or CT-scan images that mimic diverse pathological patterns, such as ground-glass opacities associated with COVID-19. This would increase the diversity of training datasets and improve model robustness and sensitivity to rare or complex cases.

Similarly, \cite{paper2} emphasizes the need for comprehensive datasets to train models capable of detecting subtle variations in gastrointestinal anomalies. The methodology proposed in this study could be leveraged to generate synthetic endoscopic images that reflect a wide range of textures, colors, and morphological features. These synthetic images could enhance the classifier's ability to generalize across patient populations and improve the interpretability of AI-driven diagnostic systems, particularly in capturing edge cases or rare conditions.

In the domain of secure medical systems, \cite{paper3} underscores the importance of robust biometric datasets for validating authentication mechanisms. GAN-generated biometric samples, such as synthetic facial expressions or physiological signals, could be valuable resources for testing and enhancing biohashing techniques. By introducing controlled variability into the dataset, the proposed approach could improve the resilience of telecare systems against adversarial attacks and ensure reliability in real-world applications."

6. The discussion section is a critical part of the manuscript where the significance of the findings is thoroughly examined in the context of existing literature. I suggest extending this section to provide a deeper analysis and interpretation of the results.

 

Answer: 

Section 5 (Discussion) has been significantly expanded to include a deeper analysis of the results.

Reviewer 4 Report

Comments and Suggestions for Authors

Review on “Improving Art Style Classification through Data Augmentation Using Generative Adversarial Networks”

Dear Authors,

After thorough review and evaluation of the study on “Improving Art Style Classification through Data Augmentation Using Generative Adversarial Networks”, here are the reviewer’s comments;

1.      The methodology is thoroughly documented, especially the use of GANs and StyleGAN for sample augmentation. However, the explanation regarding the GAN framework may be improved to improve clarity.

·         Provide further information on why StyleGAN was chosen over DCGAN or Pix2Pix.

·         Provide a simple description of the Wasserstein loss function for those who are inexperienced with modern GAN techniques.

·         While the dataset preprocessing methods are discussed, they might go into further detail about the issues (for example, balancing stylistic diversity or resolving data biases).

Author Response

Manuscript ID: electronics-3341862

Type of manuscript: Article

Title: Improving Art Style Classification through Data Augmentation Using Generative Adversarial Networks

Authors: Miguel Ángel Martín-Moyano, Iván García-Aguilar, Ezequiel López-Rubio, Rafael M. Luque-Baena *

 

ANSWERS TO REVIEWERS

We thankfully acknowledge the reviewers for their constructive comments, which have indeed helped us to improve the paper. The reviewer’s concerns have been addressed and the paper has been deeply reviewed.

Yours faithfully,

The authors

 

1. Provide further information on why StyleGAN was chosen over DCGAN or Pix2Pix.

 

Answer:

 

The manuscript has been revised to provide additional clarity and detail regarding the selection of StyleGAN over other GAN frameworks such as DCGAN, Pix2Pix, and CycleGAN. A new paragraph has been added to explain the rationale for this choice, emphasizing the advantages of StyleGAN for this specific application. This addition provides a clear comparison and justification for the use of StyleGAN, enhancing the clarity of the methodology. The following content was added in the section 3.1:

 

"Generative Adversarial Networks (GANs) constitute a fundamental process for creating an expanded dataset that enhances the performance of the pictorial-style classifier. For this purpose, the StyleGAN model \cite{9156570} was selected due to \textcolor{blue}{its advanced capabilities for generating high-quality, diverse, and stylistically consistent images compared to other generative models such as DCGAN \cite{dcgan}, Pix2Pix \cite{pix2pix}, and CycleGAN \cite{8237506}. Unlike DCGAN, which primarily focuses on generating general-purpose images and lacks explicit control over the visual styles of the generated images, StyleGAN incorporates a style-based architecture that allows detailed control over visual attributes. This made it advantageous in this context, where preserving specific stylistic elements such as texture, composition, and color palette was essential. Compared to Pix2Pix, designed for paired image-to-image translation tasks and requires a dataset of aligned image pairs, StyleGAN operates in an unpaired setting, making it more suitable for this application. The absence of aligned datasets for most artistic styles renders Pix2Pix impractical for generating synthetic images in this context. CycleGAN excels in unpaired image-to-image translation, allowing the conversion of images between domains without requiring aligned pairs. However, it focuses on domain translation rather than generating new, diverse samples within a specific domain. This limits its applicability when the objective is to expand a dataset with novel samples rather than transforming existing ones. Moreover, CycleGAN often struggles with generating highly diverse and stylistically consistent images within a single domain.

 

For all these reasons, StyleGAN was the model selected, as it allows the generation of highly diverse images while maintaining coherence within a given style. This level of control and quality is critical for generating synthetic data that effectively complements the dataset without introducing artifacts or inconsistencies that could negatively impact the performance of the classification model, making it the most suitable option for this study."

2. Provide a simple description of the Wasserstein loss function for those who are inexperienced with modern GAN techniques.

 

Answer: 

 

A detailed explanation of the Wasserstein loss function has been added to make the methodology accessible to readers unfamiliar with modern GAN techniques. This explanation clarifies the purpose and benefits of using the Wasserstein loss function, particularly in the context of GAN training stability and output quality. The following content was added in the section 3.1:

 

"The Wasserstein loss function was designed to address common challenges in training traditional GANs, such as instability and mode collapse. Unlike the cross-entropy-based loss, which classifies individual samples as real or generated, the Wasserstein loss is based on the Wasserstein distance. This metric measures how much "work" is required to transform the distribution of generated data into the distribution of real data. This approach quantifies the difference between real and generated data distributions by comparing their statistical properties rather than focusing on individual classifications. The key advantage of the Wasserstein loss lies in its ability to provide a smoother and more stable optimization process. Unlike traditional losses that can saturate when generated data deviates significantly from the real distribution, the Wasserstein loss varies continuously even in such cases. This stability mitigates training challenges and ensures that the generator receives meaningful gradients throughout the process. By stabilizing the training dynamics, the Wasserstein loss enables the GAN to converge to a more balanced state, where the generator produces images that closely resemble the target distribution. This leads to improved visual quality, with generated images exhibiting finer details, better texture consistency, and higher stylistic fidelity."

3. While the dataset preprocessing methods are discussed, they might go into further detail about the issues (for example, balancing stylistic diversity or resolving data biases).

 

Answer: 

 

The manuscript now provides a more detailed discussion of dataset preprocessing, specifically addressing imbalances and biases in the original dataset. This addition explains how stylistic diversity and balance were achieved in the dataset to improve the model's robustness and fairness. This new content is included in section 4.3.4.

 

"Analyses revealed imbalances in the original dataset, where underrepresented styles such as Ukiyo-e and Color Field Painting exhibited lower classification accuracy than more prevalent styles like Impressionism and Renaissance. A synthetic data generation strategy was implemented to address this, creating a uniform number of 300 synthetic images per style. This approach equalized the representation of all artistic movements in the training dataset, mitigating biases and enhancing the dataset’s stylistic diversity. Special attention was given to preserving the defining characteristics of each style, such as texture, composition, and color palette, to ensure stylistic coherence. Additionally, qualitative inspections of the generated images ensured that the synthetic data did not introduce repetitive patterns or artifacts, supporting a balanced and diverse training set."

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed most of the concerns. However, there are still some unclear points. The GAN's convergence cannot be assessed by a qualitative analysis. On the contrary, there are several measures for this task. This point is critical for the thorough study of the proposed method

Author Response

Manuscript ID: electronics-3341862

Type of manuscript: Article

Title: Improving Art Style Classification through Data Augmentation Using Diffusion Models

Authors: Miguel Ángel Martín-Moyano, Iván García-Aguilar, Ezequiel López-Rubio, Rafael M. Luque-Baena *

 

ANSWERS TO REVIEWERS

We thankfully acknowledge the reviewers for their constructive comments, which have indeed helped us to improve the paper. The reviewer’s concerns have been addressed and the paper has been deeply reviewed.

Yours faithfully,

The authors

 

1. The authors have addressed most of the concerns. However, there are still some unclear points. The GAN's convergence cannot be assessed by a qualitative analysis. On the contrary, there are several measures for this task. This point is critical for the thorough study of the proposed method

Answer:

We sincerely appreciate your comments and the opportunity to address the points raised. An error has been found in the model described in our work. Initially, StyleGAN was considered as the main model for our analysis. However, as the research progressed, we used Fooocus, a diffusion-based pre-trained model designed for imaging tasks. This decision was based on the proven quality of the model's results and its suitability for the objectives of our study. The use of Fooocus has been made clear in the manuscript, with a direct reference to its public repository at https://github.com/lllyasviel/Fooocus.

The Fooocus model was used pre-trained. Retraining a diffusion-based model was considered impractical in the limited six-day timeframe provided for the reviews and given the resources available to us. Diffusion models, in particular, require a lot of computation and time for effective retraining, which is beyond the scope of this review period. Nevertheless, we recognize the importance of retraining to fully explore the model's potential in future extensions of this work. 

Regarding the assessment of model convergence, we agree that quantitative measures, such as loss curves, are critical in model training. However, since this study focuses on a pre-trained model, such metrics were not applicable. Instead, our evaluation focused on qualitative inspections of the generated images to ensure the consistency of key stylistic elements (e.g., texture, color, and composition) and their alignment with the intended artistic styles. These inspections validated the suitability of the synthetic data for the augmentation and classification tasks.

In addition, one of the main contributions of this work is the design of detailed and carefully tailored prompts that allow the Fooocus model to generate relevant and high-quality augmented training images without the need for retraining. This approach reduces the computational burden without losing sight of the main goal of assessing the impact of synthetic data on the classification of pictorial styles.

- We have also improved multiple sections of the document, focusing on explaining the proposed methodology. The description of the methodology now provides a more precise understanding of the use of the Fooocus model, a pretrained diffusion-based framework, detailing the motivation for its selection and the design of tailored prompts to generate high-quality synthetic data without retraining.

Reviewer 3 Report

Comments and Suggestions for Authors

All the changes have been incorporated. 

Author Response

Manuscript ID: electronics-3341862

Type of manuscript: Article

Title: Improving Art Style Classification through Data Augmentation Using Generative Adversarial Networks

Authors: Miguel Ángel Martín-Moyano, Iván García-Aguilar, Ezequiel López-Rubio, Rafael M. Luque-Baena *

 

ANSWERS TO REVIEWERS

We thankfully acknowledge the reviewers for their constructive comments, which have indeed helped us to improve the paper. The reviewer’s concerns have been addressed and the paper has been deeply reviewed.

Yours faithfully,

The authors

 

1. All the changes have been incorporated. 

Answer:

Thank you for your thorough review and for confirming that all the changes have been incorporated. I appreciate your time and feedback throughout the process. 

Round 3

Reviewer 1 Report

Comments and Suggestions for Authors

All the reviewer's concerns have been addressed.