Content-style disentangled representation for controllable artistic image stylization and generation

In Appendix Figures 14 and 15, we present comparative experiments with DreamStyler and StyleDrop.

• Comparison with StyleDrop

For the comparison with StyleDrop, we used the original results provided in its paper. The experimental results demonstrate that StyleDrop and our method have distinct focuses. StyleDrop does a good job in transferring the color and texture of the reference image to the generated results by fine-tuning with a single reference image and a content prompt. In contrast, our method employs a disentangled design to extract the style characteristics from the reference image, capturing the artist's overall style. This allows our method to generate diverse results that remain consistent with the artist's style. To more intuitively highlight our method's ability, we included original works by the reference artist (Van Gogh) with themes similar to the content prompts. The results further validate that our method achieves higher consistency with the artist's distinctive style compared to StyleDrop. Under a unified style, our approach reflects different content in varied yet stylistically consistent ways.

• Comparison with DreamStyler

For comparison with DreamStyler, we observed major content leakage from the reference images, where the generated results almost replicate the structure of the reference images but fail to adequately reflect the content prompts. In the fourth row, the generated result resembles a photograph, failing to follow the style of the reference image. This limitation arises because DreamStyler focuses on enhancing content information in the generated results through text inversion during fine-tuning. As a result, when there is a significant mismatch between the content prompt and the reference style image, the disentanglement performance drops. In contrast, our method focuses on the explicit disentanglement of style and content features. This approach ensures that the generated results not only faithfully adhere to the input content prompt but also closely align with the artist's overall style.

In conclusion, while both DreamStyler and StyleDrop produce intriguing results, their core focus and methodologies differ significantly from ours. DreamStyler and StyleDrop primarily emphasize replicating the color and texture of the reference image by leveraging text inversion techniques during the fine-tuning stage. In contrast, our research focuses on achieving a more effective disentanglement of content and style from reference images, aiming to emulate the artist’s painting style—characterized by diverse yet consistent artistic expressions across different subjects.

While content-style disentanglement has been explored in prior works, we significantly advanced this field by achieving explicit disentanglement of content and style at the feature level, enabling us to capture the overall style of the artist. This progress is attributed to our innovative contributions in both dataset construction and model design. Specifically:

(1) Contributions of WikiStyle+:

WikiStyle+ is more than a curated dataset—it is designed to address the challenge of disentangling content and style, a foundational problem in artistic stylization. Compared to existing datasets, WikiStyle+ introduces:

• A significantly larger scale, encompassing over 146k images and 209 styles, which broadens its applicability for various downstream tasks.
• The construction of artistic image–style description–content description triplets, making it the first dataset of its kind. This structure not only supports disentanglement but also enables multimodal tasks such as style transfer, retrieval, and generation.

While incremental contributions to datasets are often perceived as less innovative, we believe WikiStyle+ fills a critical gap by emphasizing disentanglement, and our experimental results highlight its robustness even for lesser-known styles with general descriptions.

(2) Contributions of CSDN:

Although previous approaches to content-style disentanglement have utilized Q-Formers, their disentanglement methods were not explicit. Instead, they typically relied on synthesized image pairs (DEA-Diff) or assumed linear additivity between content and style features in the latent space (InstantStyle), which often resulted in noticeable content leakage in the generated outputs. To the best of our knowledge, no framework has yet been proposed to achieve explicit content-style disentanglement at the feature level. Our work is the first to introduce a framework specifically designed for explicit disentanglement by multimodal alignment , highlighting its robustness in comprehensive experiments.

• Furthermore, our approach goes beyond simply integrating Q-Formers:

  • We carefully integrate Q-Formers into a novel pipeline to achieve explicit disentanglement at the feature level. This is non-trivial, as it requires aligning features across modalities while preserving both content and style fidelity.
  • The disentanglement process leverages both WikiStyle+ and Q-Formers to achieve significant improvements in robustness and flexibility. The robustness of our disentanglement approach is validated by the consistent generation quality across varying levels of detail in input descriptions, which discussed in Appendix 3.5, Figure 12. Moreover, our method is capable of disentangling content not only from photographs but also from artistic images, effectively separating content and style, as demonstrated in Figures 7 and 17.

We agree that further exploration of specific style elements, such as brushstrokes and color palettes, is indeed an exciting direction. We thank the reviewer for highlighting these areas of improvement, as they align with our long-term goals for advancing the field of artistic stylization. While our current work focuses on building a robust foundation for content-style disentanglement, we see it as a stepping stone toward deeper and more granular exploration of artistic styles.

We appreciate the reviewer’s insightful comments regarding the expressiveness of style descriptions in WikiStyle+, particularly for lesser-known artistic styles. Addressing your concerns, we would like to clarify a few points:

(1) Style descriptions in WikiStyle+

The style descriptions in WikiStyle+ are indeed based on the existing tags from WikiArt, which were chosen for their scientific and objective definitions, ensuring both consistency and usability across the dataset. While these tags provide a general foundation for style representation, we acknowledge that they may lack nuanced details for lesser-known styles. we see this as a potential avenue for future exploration.

While the textual descriptions of styles are relatively simple, this does not negatively impact the experimental results. Through the first-stage multimodal alignment pre-training, the model is able to learn features in the latent space that effectively capture the differences between artistic styles, even if the textual descriptions do not explicitly mention specific brushstrokes, compositions, or other details. Additionally, WikiStyle+ covers a wide range of styles and artists, enable the model to form rich representations of artistic styles in the visual space, compensating for the simplicity of the textual descriptions. Consequently, as demonstrated in the experimental results in Appendix 3.5, even when the textual descriptions range from detailed (e.g., an entire reference image) to simplest (e.g., a single word), the fidelity of the generated styles does not show significant degradation, further validating the robustness of our approach.

(2) Performance for lesser-known styles with general style descriptions

For lesser-known styles, we agree that more detailed descriptors (e.g., brushstroke techniques, color palettes) could improve expressiveness and style fidelity. Still, in our experiments, we observed that even with relatively general descriptions, our model could generate visually alike results for lesser prominent styles. Examples include Kanstantin (Figure7, column 4, Figure 12, (b)), Samuel Peploe (Figure 4, row 4; Figure 5, column 4; Figure 6, row 4), and Berthe Morisot (Figure 4, row 2; Figure 5, column 5; Figure 6, row 2). These results demonstrate that our model not only excels with iconic styles but also effectively learns and generates for underrepresented styles, highlighting the robustness of our approach. Moreover, as shown in the experiments in A3.5, Figure 12 (b), a comparison between the fourth and fifth columns reveals that using a general description yields results comparable to directly using the style image as the description.

Still, the experiments in A3.5 indeed demonstrate a clear trend: more detailed descriptions can enhance the style expressiveness of the results. To address this and further explore the potential of detailed style description, we plan to extend the style descriptions in WikiStyle+ as our future work by incorporating more specific attributes , such as brushstroke details, compositional techniques, and color schemes.

In Appendix Section 3.5, Figure 12, we present an analysis of the impact of varying levels of detail in content and style descriptions on disentanglement performance. The results demonstrate the robustness of our method across a wide range of detail levels.

• Changes in Content Descriptions:

  As illustrated in Figure 12(a), whether the content description is as simple as a single word or as detailed as an image, our method achieves good disentanglement of the content information. The generated images accurately capture the described content, whether from minimal or detailed descriptions, while maintaining consistent stylistic alignment with the artist. This demonstrates the model's ability to effectively capture content semantics across varying levels of descriptive detail.

• Changes in Style Descriptions:

  As shown in Figure 12(b), as the style descriptions progress from single words to detailed depictions like an artist’s paintings, the brushstrokes and styles in the generated images become increasingly nuanced and refined. However, the themes of the generated images remain unaffected, demonstrating the model’s strong disentanglement capability.

These results highlight the robustness of our explicit disentanglement network design, as it consistently maintains stable disentanglement performance across varying levels of detail in content and style descriptions. Even with minimal descriptions, the model reliably achieves disentanglement. Richer style descriptions enhance the stylistic expressiveness of the generated results, while more detailed content descriptions result in correspondingly richer content in the output. However, these descriptions only affect their respective aspects, and variations in the amount of descriptive information do not impact the disentanglement performance.

• Mathematical expressions and symbols

  In the revised PDF, we made the following modifications for mathematical expressions and symbols:

  • Throughout the text, we use the uppercase letter $I$ to represent the image modality. In Equation (2), we replaced I with the lowercase letter $n$ to denote the sample index. Accordingly, modifications were also made to Equation (3).
  • We standardized the notations across the text. Following common conventions in the literature, we use the lowercase letter $t$ to denote the time step, the uppercase letter $T$ to represent text (Equation 2), and $M$ and $m$ to denote text length and text sequence index, respectively (Equation 4). The corresponding sections of the text have been updated to incorporate these changes.
  • $Z$ and $Z_t$ are not image features; they represent the noise states at time step $t$ during the diffusion process. We updated the framework diagram (now Figure 3) to provide a clearer depiction of the diffusion process module. We use $Z_0$ and $Z_T$ in Figure 3 to denote noise states at the first and last time step, respectively.

• Typos

  We greatly appreciate your attention to detail. We have carefully reviewed the text and corrected all identified typos in the revised version. Additionally, we have conducted a thorough proofreading of the entire manuscript to ensure that no other typos remain.

In disentanglement loss function, Image-Text Contrastive Loss (ITC) primarily handles the tasks of alignment and disentanglement in the network, while the Image-Text Matching Loss (ITM) and Image-grounded Text Generation Loss (ITG) provide auxiliary support for modality alignment based on ITC. To evaluate the contribution of each loss function and their impact on final results, we conducted an ablation study in Appendix 3.3, where four different configurations of the model are compared:

1. Training with only Image-Text Contrastive Loss (ITC). ITC is the core loss function in the first stage, enables the model to disentangle style and content more effectively by ensuring that features corresponding to style and content are aligned with their respective textual descriptions. When the CSDN is trained exclusively with ITC, the overall content structure is preserved. When the model is trained solely with ITC, the content alignment score achieves the highest value among all configurations. This indicates that $\mathcal{L}_{\mathit{itc}}$ as a contrastive loss, is highly effective in aligning the content features with their corresponding textual descriptions. However, the generated results lack intricate brushstroke details, and fail to capture the characteristic yellowish tone of traditional oil paintings, as illustrated in Fig. 10(a). The lack of fine-grained alignment with styles is corroborated by the decreased SS metrics in Table 5.
2. Training without Image-grounded Text Generation Loss (ITG): ITG trains the model to generate coherent style and content descriptions for a given image by predicting the next word or sentence based on the image and context. ITG ensures that the style and content information extracted from an image is not only disentangled but also interpretable and coherent, which enhances the overall textual-visual understanding of the model. As shown in Fig.10 (b), the absence of ITG affects the coherence of the style and content descriptions extracted from the image. For example, in the "A bus" prompt, the generated bus elements are inconsistent. Also, stylistic elements such as brushstrokes are less pronounced, leading to the drop of SS and TA in quantitative results.
3. Training without Image-Text Matching Loss (ITM): ITM operates as a binary classification task, predicting whether an image-text pair is a positive or negative match. This enables the model to focus on fine-grained correspondence between images and text, such as specific objects or elements mentioned in the prompts. When ITM is removed, the generated images show a loss of detail in both style and content alignment. For instance, in Fig. 10 (c), the ``house'' element explicitly mentioned in the content prompt is missing in the generated image. This indicates that, without ITM, the model’s capacity to maintain fine-grained alignment is compromised, resulting in less coherent and contextually relevant outputs.
4. Full-loss setting: When all three loss functions are used together, the SS metrics achieve the highest scores. From Fig. 10 (d), the generated images align closely with both style and content prompts, showcasing strong disentanglement and alignment. The style is faithfully preserved, while the content, such as the "house" or "bus," is accurately represented in the generated outputs. This demonstrates the complementary nature of the three loss functions in ensuring both disentanglement and multimodal coherence. While ITC prioritizes content alignment, the inclusion of ITM and ITG shifts the focus toward achieving a better balance, where both content and style are accurately disentangled and aligned with their textual descriptions.

In conclusion:

• ITC ensures effective alignment of visual and textual features.
• ITG ensures that the style and content information extracted from an image is not only disentangled but also interpretable and coherent. Without ITG, the generated images may lose stylistic fidelity or fail to fully capture content prompts, as the model would lack a strong textual grounding during training.
• ITM enhances fine-grained alignment between textual descriptions and corresponding visual elements, enabling the model to accurately reflect details in the generated images. Without ITM, the relevance of generated images to the input prompts degrades, with style and content features becoming less accurately matched to their textual descriptions.

CSDN is the network trained in the first stage of our method, with the objective of aligning the style features of artistic images with style description texts and the content features with content description texts in the feature space, thereby achieving the disentanglement of style and content. The structure of CSDN includes a pre-image encoder and a Q-Former, where the Q-Former is the trainable module designed to extract multimodal representations related to style and content (disentanglement). The training of CSDN involves three pretraining tasks: Image-Text Contrastive Learning (ITC), Image-Text Matching (ITM), and Image-Text Generation (ITG). The Q-Former consists of main parameters, a binary classifier for ITM, and a text decoder for ITG. The main parameters are jointly updated by all three tasks, while the binary classifier and text decoder are updated specifically by the ITM and ITG losses, respectively.

MCL is the structure that injects the disentangled embeddings from CSDN as control conditions into the diffusion process at different timesteps using cross-attention mechanisms. At each timestep of the diffusion process, the style and content embeddings are introduced as conditions through the cross-attention layers in MCL to guide the generation process.

We have provided a more detailed explanation of cross-attention in MCL in section 4.2 in the revised PDF. Additionally, we have made the distinction between CSDN and MCL more explicit in the framework figure (Figure 3) to better highlight their differences.

• MCL is mentioned in Line 251 but is not further elaborated upon.

We revised the writing in Section 4.2 to include a more detailed explanation of MCL: “We use multi-step learnable cross-attention layers (MCL) to inject the style embeddings into the denoising process of the SD model. At each timestep of the diffusion process, the style features are introduced as conditions through the cross-attention layers in MCL to guide the generation process. These cross-attention layers embed the style features into the current diffusion features using the attention mechanism.”

For a more comprehensive context, please refer to Section 4.2 of the revised PDF.

• In Eq 3, the specific workings of the binary classification network are not described, including its specific inputs.

In Eq. (3), the ITM loss is formulated as binary classification task, predicting whether an image-text pair is a match (positive) or not (negative). This encourages the model to focus on fine-grained correspondence between images and text. The binary classifier is updated alongside the main parameters of the Q-Former during the ITM task.

The ITM loss computes the cosine similarity between the image embedding $I$ and text embedding $T$, and then uses a linear layer to map the cosine similarity to a matching probability. For content correspondence matching, the inputs are ($I_c$ , $T_c$); for style correspondence matching, the inputs are ($I_s$ , $T_s$). The ITM loss employs a binary classification loss to optimize both the Q-Former and the classifier. We included a more detailed introduction for ITM task in Sec.4.1 of the revised PDF (L260-269).

• The text generation model used in Eq 4 is not introduced in the text or figures.

We use a lightweight text decoder as text generation model in ITG loss. Its primary structure includes a transformation module that applies a dense projection, an activation function, and layer normalization to refine hidden states, and a decoder layer that maps the processed hidden states to vocabulary logits using a linear layer. We included a more detailed introduction for ITG task in Sec.4.1 of the revised PDF (L270-279).

In Appendix 3.1 Table 3, we compared the inference speed of our method with two structurally similar methods, T2I-Adapter and Dea-Diff in the table below. All three methods utilize the pre-trained SD model as the generative backbone; however, they differ in the conditional modules used to control the diffusion process. The complexity of our method is comparable to that of Dea-Diff, resulting in similar inference times.

In Appendix 3.2, Table 4, we included quantitative comparisons with SD from Section 5.2.2 (stylized text-to-image generation) and ArtBank from Section 5.2.3 (collection-based stylization tasks). The results show that our method achieves performance metrics comparable to Stable Diffusion and surpasses ArtBank. In our quantitative experiments across different settings, SD outperforms all existing methods.

Table: Quantitative comparison with stylized text-to-image generation and collection-based stylization methods.

References:

1. Rombach et al. (2022). Stable Diffusion.
2. Zhang et al. (2024). ArtBank.

As shown in Figure 5, SD exhibits a more pronounced stylization, which contributes to its superior performance on quantitative metrics. However, as illustrated in Figure 16, SD sometimes also demonstrates limitations in disentangle content and style. The generated images include elements not specified in the input images (e.g., row 4, the person in the middle, which is absent from both the style image and the content image).

For the comparison with DALL-E, as it is a not open-sourced commercial software intended for testing purposes only, we conducted visual comparisons exclusively (Figure 5). The results reveal that DALL-E often generates repetitive patterns and textures across different style prompts, suggesting that its stylization performance is suboptimal.

• Clarification on selected metric

The style similarity metric we selected is from DEA-Diff (CVPR 2024). The calculation of this metric involves first generating descriptive text for the reference image and then removing content-related elements from the description. The style similarity is then calculated by computing the CLIP similarity between the generated image and the style description text.

We chose not to directly measure the CLIP similarity between the generated image and the reference image for two main reasons:

1. This metric is not calculated directly using the CLIP features of the reference image but rather based on the CLIP features of the style description derived from it. This ensures that the similarity is measured specifically between the generated results and the style characteristics of the reference image, rather than its overall similarity. Consequently, this metric is better suited for evaluating a model's ability to disentangle content and style in multimodal (text-to-image) tasks, aligning with the core objective of our approach, as well as that of DEA-Diff, ArtBank.
2. Since CLIP is pre-trained primarily for image recognition tasks, directly using the CLIP similarity between the reference image and the generated image would predominantly measure content similarity rather than style similarity.
3. Moreover, IP-Adapter and other T2I algorithms also use similar metrics, calculating the CLIP similarity between the generated image and the style prompt. Since this is a multimodal task, we believe it is a reasonable approach to adopt metrics that incorporate multimodal inputs.

This metric is indeed not suitable for evaluating performance on color transfer. However, for models aimed at disentangled representations, it is appropriate for assessing style similarity from a disentanglement perspective. While some comparison methods exhibit minimal color deviation in visual results, our method achieves better style similarity according to this metric, demonstrating that our approach effectively disentangles the artist's style from the reference image.

• Why not use Gram Matrix loss for style similarity metric

The Gram matrix, which computes the correlation between channels of CNN features, primarily measures color and texture similarity. It has predominantly been used as a style loss function in earlier image-to-image style transfer algorithms. In such works, Gram features are rarely employed directly as a metric for evaluating style similarity. Examples include AdaIN, SANet, AdaAttn, Avatar-Net, StyleBank, and others. Although the Gram matrix is used in loss functions to calculate color and texture correlations, it is less suitable than CLIP-based metrics for tasks supporting multimodal inputs. This metric favors algorithms that define style primarily as color but is not fair for algorithms focusing on the artist’s overall style (such as our method and ArtBank, which also emphasize the artist's overall style).

In contrast, recent studies on text-to-image(T2I) stylization methods predominantly adopt multimodal feature similarity in CLIP space as a quantitative metric for assessing style similarity and content consistency. This metric is especially relevant for methods that support multimodal input. While we observed that DreamStyler utilizes Gram loss, the majority of T2I methods, such as ArtAdapter, ArtBank, DEA-Diff, InstantStyle, IP-Adapter, T2I-Adapter, and StyleTokenizer, do not use Gram features as an evaluation metric. Instead, they rely on CLIP-based metrics as their primary evaluation criteria.

For the reasons mentioned above, we have not adopted this metric in our study.

In Appendix Figures 14 and 15, we present comparative experiments with DreamStyler and StyleDrop.

• Comparison with StyleDrop

For the comparison with StyleDrop, we used the original results provided in its paper. The experimental results demonstrate that StyleDrop and our method have distinct focuses. StyleDrop does a good job in transferring the color and texture of the reference image to the generated results by fine-tuning with a single reference image and a content prompt. In contrast, our method employs a disentangled design to extract the style characteristics from the reference image, capturing the artist's overall style. This allows our method to generate diverse results that remain consistent with the artist's style. To more intuitively highlight our method's ability, we included original works by the reference artist (Van Gogh) with themes similar to the content prompts. The results further validate that our method achieves higher consistency with the artist's distinctive style compared to StyleDrop. Under a unified style, our approach reflects different content in varied yet stylistically consistent ways.

• Comparison with DreamStyler

For comparison with DreamStyler, we observed major content leakage from the reference images, where the generated results almost replicate the structure of the reference images but fail to adequately reflect the content prompts. In the fourth row, the generated result resembles a photograph, failing to follow the style of the reference image. This limitation arises because DreamStyler focuses on enhancing content information in the generated results through text inversion during fine-tuning. As a result, when there is a significant mismatch between the content prompt and the reference style image, the disentanglement performance drops. In contrast, our method focuses on the explicit disentanglement of style and content features. This approach ensures that the generated results not only faithfully adhere to the input content prompt but also closely align with the artist's overall style.

In conclusion, while both DreamStyler and StyleDrop produce intriguing results, their core focus and methodologies differ significantly from ours. DreamStyler and StyleDrop primarily emphasize replicating the color and texture of the reference image by leveraging text inversion techniques during the fine-tuning stage. In contrast, our research focuses on achieving a more effective disentanglement of content and style from reference images, aiming to emulate the artist’s painting style—characterized by diverse yet consistent artistic expressions across different subjects.

We understand that the perception of style similarity can vary among individuals, which may explain the reviewer’s concern that our generated results deviate from the reference style (e.g., in terms of color). While color similarity is one interpretation of style correlation, our primary goal is to emulate the artist’s overall style—specifically, the diverse yet consistent artistic expressions when portraying different themes. We would like to clarify that our statement of result quality is based on our aim to capturing the broader stylistic essence of an artist for stylization, rather than transferring colors. Under a unified style for an artist or genre, our approach produces results that are diverse yet remain stylistically consistent.

Achieving this objective relies on disentanglement, which is why we also place significant emphasis on comparing performance in terms of disentanglement when analyzing the results. While some methods succeed from the perspective of color preservation, they often fail to generate results that faithfully align with the content prompt. Take Figure 4 (formerly 5) for example, our method achieves the best balance between generating content consistent with the content prompt and adhering to the style of the artist represented by the reference image.

We would also like to highlight the observations made by two other reviewers, fLvE and gsAG, who noted that our method “captures the nuanced brushstroke techniques of artists rather than simply replicating signature swirling patterns” (Figure 1) or merely transferring colors (Figures 4 and 6), and “achieved more fine-grained control in generation, yielding excellent results” . These comments align with our definition of style, which goes beyond simple replication of color or texture, and further validate our claims about the effectiveness of our approach.

Although InstantStyle and DEA-Diff also aim to disentangle content and style, their generated results often reproduce the style of the content image or the content of the style image, indicating insufficient disentanglement. Specifically, the brushstrokes of DEA-Diff resemble those of comic art rather than oil paintings, while InstantStyle exhibits content leakage from the reference image. For example, in the first row, the generated image includes elements such as wheat fields and trees from the reference image, even though these are not mentioned in the content prompt. Similarly, IP-Adapter and T2I-Adapter maintain style consistency but fail to successfully generate content specified in the prompt.

To further support our explanation, we have included additional visual experiments in Appendix Figure 11. It is important to note that our method is capable of transferring colors and textures. However, much like emulating an artist’s painting style, this occurs only when the content prompt is thematically aligned with the reference image. In such cases, our method produces colors and textures that closely match those of the reference image.

• What are the outputs and training objectives of the Content and Style Disentangled Network (CSDN)? What is its structure?

  • The output of CSDN consists of disentangled style and content embeddings. The training objective of CSDN is to align the style features of artistic images with style descriptions and the content features of images with content descriptions in the feature space, thereby achieving disentanglement of style and content. The structure of CSDN includes a pre-image encoder and a Q-Former. The Q-Former, as a trainable module, is designed to explicitly disentangle features by extracting multimodal representations related to style and content, respectively.

  • CSDN is trained using three pretraining tasks: Image-Text Contrastive Learning (ITC), Image-Text Matching (ITM), and Image-Text Generation (ITG). The trainable parameters of the Q-Former include the main module parameters, the binary classifier parameters for ITM, and the text decoder parameters for ITG. Among these, the main module parameters are updated jointly by all three tasks, while the binary classifier and text decoder are updated specifically by the ITM and ITG losses, respectively.

  • To enhance clarity, we optimized the framework figure by including some intermediate variables to make the structure of CSDN more explicit.

• The image-grounded text generation loss involves training a model to generate descriptive text that corresponds to a given input image. To whom does the 'model' refer in this context?

  • Here, the ‘model’ refers to CSDN, specifically the main parameters of the Q-Former and the text decoder updated through the ITG task in first stage training. A lightweight text decoder is used to generation the text sequence. It consists of two main components: a transformation module that applies a dense projection, an activation function, and layer normalization to refine hidden states, and a decoder layer that maps the processed hidden states to vocabulary logits using a linear layer. We have provided a more detailed introduction of ITG in Section 4.1 of the revised PDF.

• What is the 'two-class linear classifier' mentioned in the image-text matching loss, and where does it come from?

  • In the image-text matching loss, the 'two-class linear classifier' refers to an auxiliary binary classifier that predicts whether an image-text pair is a match (positive) or not (negative). This binary classifier is updated alongside the main parameters of the Q-Former during the ITM task. ITM computes the cosine similarity between image embedding and text embedding, then using a linear layer to map the cosine similarity into matching probability. ITM uses binary classification loss to optimize the Q-former and classifier. As one of the three training tasks for CSDN, the ITM task focuses on learning fine-grained multimodal relationships between images, style text, and content text. This enables the model to focus on fine-grained correspondence between images and text. We have provided a more detailed introduction of ITM in Section 4.1 of the revised PDF.

In Eq.(1), Image-Text Contrastive Loss (ITC) primarily handles the tasks of alignment and disentanglement in the network, while the Image-Text Matching Loss (ITM) and Image-grounded Text Generation Loss (ITG) provide auxiliary support for modality alignment based on ITC. To evaluate the contribution of each loss function and their impact on final results, we conducted an ablation study in Appendix 3.3, where four different configurations of the model are compared:

1. Training with only Image-Text Contrastive Loss (ITC). ITC is the core loss function in the first stage, enables the model to disentangle style and content more effectively by ensuring that features corresponding to style and content are aligned with their respective textual descriptions. When the CSDN is trained exclusively with ITC, the overall content structure is preserved. When the model is trained solely with ITC, the content alignment score achieves the highest value among all configurations. This indicates that $\mathcal{L}_{\mathit{itc}}$ as a contrastive loss, is highly effective in aligning the content features with their corresponding textual descriptions. However, the generated results lack intricate brushstroke details, and fail to capture the characteristic yellowish tone of traditional oil paintings, as illustrated in Fig. 10(a). The lack of fine-grained alignment with styles is corroborated by the decreased SS metrics in Table 5.

2. Training without Image-grounded Text Generation Loss (ITG): ITG trains the model to generate coherent style and content descriptions for a given image by predicting the next word or sentence based on the image and context. ITG ensures that the style and content information extracted from an image is not only disentangled but also interpretable and coherent, which enhances the overall textual-visual understanding of the model. As shown in Fig.10 (b), the absence of ITG affects the coherence of the style and content descriptions extracted from the image. For example, in the "A bus" prompt, the generated bus elements are inconsistent. Also, stylistic elements such as brushstrokes are less pronounced, leading to the drop of SS and TA in quantitative results.

3. Training without Image-Text Matching Loss (ITM): ITM operates as a binary classification task, predicting whether an image-text pair is a positive or negative match. This enables the model to focus on fine-grained correspondence between images and text, such as specific objects or elements mentioned in the prompts. When ITM is removed, the generated images show a loss of detail in both style and content alignment. For instance, in Fig. 10 (c), the ``house'' element explicitly mentioned in the content prompt is missing in the generated image. This indicates that, without ITM, the model’s capacity to maintain fine-grained alignment is compromised, resulting in less coherent and contextually relevant outputs.

4. Full-loss setting: When all three loss functions are used together, the SS metrics achieve the highest scores. From Fig. 10 (d), the generated images align closely with both style and content prompts, showcasing strong disentanglement and alignment. The style is faithfully preserved, while the content, such as the "house" or "bus," is accurately represented in the generated outputs. This demonstrates the complementary nature of the three loss functions in ensuring both disentanglement and multimodal coherence. While ITC prioritizes content alignment, the inclusion of ITM and ITG shifts the focus toward achieving a better balance, where both content and style are accurately disentangled and aligned with their textual descriptions.

In conclusion:

• ITC ensures effective alignment of visual and textual features.
• ITG ensures that the style and content information extracted from an image is not only disentangled but also interpretable and coherent. Without ITG, the generated images may lose stylistic fidelity or fail to fully capture content prompts, as the model would lack a strong textual grounding during training.
• ITM enhances fine-grained alignment between textual descriptions and corresponding visual elements, enabling the model to accurately reflect details in the generated images. Without ITM, the relevance of generated images to the input prompts degrades, with style and content features becoming less accurately matched to their textual descriptions.

• We would like to clarify that the term "multi-step" refers to the multiple denoising steps during the training process of the diffusion model, rather than indicating that the style embeddings are injected into multiple blocks of the U-Net. In our implementation, the style embeddings are indeed only injected into the middle block of the U-Net through cross-attention at every denoising step across the diffusion process. To clarify this point, we have revised the original text in L311, changing "we inject the style embeddings only into the intermediate layer of the denoising process" to "we inject the style embeddings only into the middle block of U-Net" (L312 in the revised PDF).

• For the content text, we first extract text features using the original text encoder from SD, then concatenate them with the disentangled content embeddings $e_c$ extracted by CSDN before injecting them into the diffusion process. Previously, we simplified this step in the flowchart, but for better clarity, we have explicitly added this step to Figure 3 (formerly Figure 4) in the revised PDF.

• Dose the multi-step cross-attention layers accept the modulation from the time step?

  • The multi-step cross-attention layers (MCL) in our model do not directly accept modulation from the time step embeddings. Instead, time steps are incorporated into the U-Net's processing pipeline through the standard embedding mechanism of Stable Diffusion (SD). This embedding is injected into each layer of the U-Net to modulate its feature processing based on the current diffusion step. In our implementation, MCL specifically focuses on incorporating the disentangled style and content embeddings into the denoising process via cross-attention. While the time step information indirectly influences the MCL through the modulated features of the U-Net, it is not explicitly part of the attention mechanism in the MCL. We omitted the detailed depiction of the time step embedding process in the framework for simplicity, as it follows the standard practice of SD and is orthogonal to the design of our MCL.

Thanks for the efforts for resolving my concerns about ablation study, model details and dataset comparisons. My main concern is still not addressed:

1. Generally, the overall painting habits of an artist include both the strokes and colors. With only one single out-of-domain reference image as input, the stylized image should be aligned with both of these terms. In this task, the content leakage from the reference image is hard to define. It is inappropriate to regard all the semantic objects with unusual colors as content leakage in art work generation.

2. The authors claim that the proposed method can better retain the artist’s actual painting habits than existing methods. However, overall color deviation from the reference images is also observed (e.g., in Fig. 4, 6, 11, 14 and 16).

Considering the revised paper and detailed responses from the authors, I change my rating from 3 to 5.

We understand that the perception of style similarity can be subjective, which is likely the main reason behind the reviewer’s impression that our generated results lack style similarity. However, our work adheres to the definition of style in art history, viewing it as the artist’s overall expressive techniques rather than merely replicating specific iconic textures or colors from an individual painting.

Here, we would like to reference the comments made by two other reviewers (fLvE and gsAG). Our method “captures broader brushstroke techniques of the artist rather than simply replicating signature swirling patterns” and “achieved more fine-grained control in generation and yielding excellent results (fLvE)”.

In Figure 1, the results generated by DeaDiff exhibit content leakage from the reference image, as the background of “Starry Night” appears consistently across different themes. We view this not as a follow of style but rather a replication of fixed patterns resulting from content leakage. Similarly, in Figure 4(original 5), InstantStyle also demonstrates content leakage. For example, in the first row, the generated image includes elements such as wheat fields and trees from the reference image, despite these not being mentioned in the content prompt. In contrast, our model explicitly disentangles the content and style of the reference image at the feature level, ensuring that the generated image's content is solely guided by the input content prompt, proving that our approach minimizes content leakage from the reference image.

In Appendix Figure 11, we provide a detailed comparison with InstantStyle and DEA-Diff, including the artist's original painting in the last column to offer a more intuitive demonstration of how our results more closely follow the artist's overall style. From the experimental results, InstantStyle focuses on replicating the colors and content of the reference image. DEA-Diff, on the other hand, faithfully reflects the content prompt but produces brushstrokes that deviate significantly from the reference image's style, resembling comic illustrations rather than oil paintings. In contrast, our method emphasizes imitating the artist's overall style, delivering more stylistically accurate results. This is because DEA-Diff and InstantStyle define style differently from our approach. They focus on imitating the color and texture, whereas our method aims to replicate the artist’s overall painting habits.

Under a unified style, our approach reflects different content in varied yet stylistically consistent ways. From Figure 11, regardless of the input content, DEA-Diff and InstantStyle consistently produce the same color palette and textures, which do not align with the artist’s actual painting habits. On the contrary, our generated results incorporate similar colors and distinctive swirl patterns from Starry Night only when the input content prompt explicitly includes "starry night." Additionally, when the content prompt specifies a "city skyline," our results accurately depict urban scenes, while InstantStyle continues to replicate the rural content from the reference image. Meanwhile, DEA-Diff produces results that resemble anime rather than an oil painting. These observations demonstrate that our method effectively disentangles style and content, producing outputs that align more closely with the artist’s painting habit and overall style.

In Appendix, Figure 15 further demonstrates the superiority of our disentanglement design. DeaDiff exhibits several failure cases in disentangle content information from images. For example, under the "pencil" style, the color of the mountains in the content image is incorrectly treated as style information and transferred to the generated result. With "Samuel Peploe" style, DeaDiff struggles to produce a stylized output, instead following the photographic style from the content image. These issues arise because Dea-Diff lacks the explicit disentanglement design proposed in our method. Consequently, it often reproduces the style of the content image or the content of the style image in the generated results. In contrast, our method effectively avoided content or style leakage in the generated images while capturing the artist's style.

Both our proposed WikiStyle+ dataset and the dataset in [1] are based on the publicly available WikiArt website, resulting in some overlap in images. WikiStyle+ is more than an extension version of [1]—it is designed to address the challenge of disentangling content and style, a foundational problem in artistic stylization. The primary difference between WikiStyle+ and the dataset from [1] lies in their scale, diversity, and data structure:

1. Scale and Diversity:

• WikiStyle+ contains over 146k images spanning 209 styles, while the dataset from [1] includes 85k images covering 25 styles. This broadens WikiStyle+’s applicability for various downstream tasks.

2. Data Structure:

• WikiStyle+ introduces artistic image–style description–content description triplets, making it the first dataset of its kind. This structure not only supports disentanglement but also enables multimodal tasks such as style transfer, retrieval, and generation.

We commit to open-sourcing the dataset as well as source code upon acceptance, as we believe it holds significant value for advancing research in content-style disentanglement—a critical challenge in artistic stylization.

Thank you for taking the time to review and consider our responses!

• The style of an image should be reflected in the artist's color usage habits rather than the specific colors present in that particular image. These habits encompass commonly used tones (e.g., bright, soft, or dark), color schemes (e.g., complementary, analogous, or monochromatic), as well as attributes like saturation and contrast. If style were defined solely by the specific colors in an image, then by that logic, each of Van Gogh's works with a different color palette would represent a different style.

• The single reference image is not the only color source for the generated results; the model also utilized color usage patterns embedded in the pre-trained dataset. In pretraining, the model learns from large-scale artworks and clusters similar color usage patterns (e.g., artists who paint blue skies also tend to depict sunflowers in gold). In the inference stage, the model extracts color habits from the reference image and combines this with the knowledge acquired during pretraining. As a result, the model does not simply transfer colors from the reference image but infers suitable colors for different content, ensuring better stylistic fidelity. Otherwise, when the content prompt is "sunflowers" and the reference image is Van Gogh's Starry Night, would generating "blue sunflowers sparkling with stars" truly reflect Van Gogh’s style?

These clarifications may be insufficient in the manuscript, leading to potential misunderstandings. We have revised the introduction section based on the response above. We hope this response addresses the reviewer's concern.

We sincerely thank all the reviewers for taking the time to review our rebuttal and thoughtfully consider our responses!

We address each reviewer’s second-round concerns below. We hope our responses sufficiently resolve these issues and invite you to review them.

• Color deviation from single reference image (Reviewer USxT)

• Style similarity metric (Reviewer ikfg)

• Learn style from single reference image (Reviewer ikfg)

• Experimental setting in comparison with DreamStyler (Reviewer gsAG)