CSGO: Content-Style Composition in Text-to-Image Generation

Response to Reviewer a2Ei

Thank you for the reviewers' recognition.

---

Q1-1. Why not introduce a style metric during data collection?

A1-1: To address these concerns, we provide the following clarifications: (1) B-LoRA enables efficient and stable style transfer, consistently generating images that closely align with the style characteristics of the original references. This claim is substantiated by the following experimental validation: We constructed 300 unique content-style LoRA combinations by pairing 300 content LoRAs with 300 style LoRAs. For each combination, we generated 15 stylized images. Style fidelity was rigorously quantified using three key metrics: (a) The mean style similarity between generated images and original style references, (b) The average maximum style similarity score per content-style group, and (c) The average minimum style similarity score per content-style group. To ensure robustness in style assessment, we employed two distinct style evaluation models (CSD and OneIG-StyleEncoder). Both models output similarity scores normalized to the range [-1, 1], where higher values indicate greater style congruence. As shown in the table below, the results confirm that B-LoRA effectively captures stylistic representations and achieves stable style transfer.

(2) The observed instability in B-LoRA stems primarily from content preservation inconsistencies rather than style transfer limitations. This distinction was verified through Content Alignment Score (CAS) analysis between generated images and original content references. We report three key metrics: (a) The mean CAS across all generated images, (b) The average maximum CAS per content-style group, and (c) The average minimum CAS per content-style group. The CAS metric (range: [0, ∞)) employs an inverse scale where lower values indicate superior content preservation. As shown in the table below, CAS values exhibit significant variability, confirming B-LoRA's susceptibility to content degradation.

(3) To further validate the robustness of our dataset, we randomly sampled 1000 triplets and computed the average CAS, CSD, and OneIG-StyleEncoder scores of the selected images. The experimental results confirm that the samples obtained through our selection strategy are both reliable and stable.

（4） The screening results using only CAS are more aligned with the manual selection results. To evaluate the alignment rare (AR) between different screening strategies and human selections, we conducted some verification experiments. We introduced a*CAS + b*(1-CSD) to represent the final score used for screening. We filtered the results using four strategies respectively and calculated the alignment with the manually screened results (top 1). It can be observed that due to the instability in the content LoRA of B-LoRA, the results screened using only CAS are actually the closest to the manual screening results. Therefore, we choose to screen the dataset solely based on the CAS score.

---

Q1-2. Explanation of the "Dandelion" in Figure 4.

A1-2: Comparative analysis in Figure 4 demonstrates that our method achieves superior stylistic alignment while maintaining optimal content preservation. We posit that stylistic representation inherently encompasses perceptual subjectivity, where brushstroke techniques, color palettes, and compositional arrangements constitute fundamental stylistic components. Illustratively, in a feline-form painting composed of "grass" elements, these botanical features function dually as content carriers and stylization primitives. This ontological duality suggests that strategically incorporating content-distinct elements can enhance style expression in transfer tasks.

---

Q2. Concerns about the hyperparameters.

A2: Beyond the established Classifier-Free Guidance (CFG) technique commonly employed in diffusion models, the proposed CSGO method introduces two critical hyperparameters: the content scale δ_c and the style scale λs. Analogous to conditioning mechanisms in models such as Adapter or ControlNet, these hyperparameters are designed to enable finer-grained control over generative outcomes during inference. As illustrated in Figure 11 and Figure 12, we demonstrate the impact of all hyperparameters on style transfer results. The presented examples show that these hyperparameters provide stable and controllable influence: for instance, a CFG value of 10 yields optimal performance in both example sets, while a content scale δc of 0.5 also achieves the best results in both cases. Given the inherently subjective nature of style perception, these hyperparameters facilitate adjustable generation of style transfer outputs tailored to specific requirements, thereby accommodating the need for diverse stylistic interpretations. Crucially, these hyperparameters are not unduly complex or difficult to tune; their design aligns with an intuitive understanding of the style transfer task, ensuring both accessibility and robust control. The ease of use has been systematically validated in our experiments.

Table 1. Ablation studies of \δ_c.

Table 2. Ablation studies of λs.

---

Q3. Results of the new ablation experiments.

A3: We thank the reviewers for their insightful suggestions. In response, we refined our ablation study to systematically evaluate three critical components: (1) the content control module (cross-attention layers in downsampling blocks), (2) the style control module (cross-attention layers in upsampling blocks), and (3) the dual-control fusion strategy (the style features are inject to Controlnet ).(4) the dual-control fusion strategy without style feature. Quantitative results validating each module's contribution are presented as follow. Obviously, the hybrid content control module is missing, resulting in weak content feature retention capability. The style features are mainly related to the upsampling style control module and the style injection of the Controlnet branch. In order to balance the migration capabilities of style and content, we finally integrated the above-mentioned modules to achieve the optimal results.

---

Q4. Please provide more details about the style token (Perceiver Resampler), as well as the formulation of (line 177) and F(S)' in style control.

A4: (1) The style token refers to the number of tokens obtained after mapping the original style features using a resampler structure. A larger number of tokens means that more style features are retained. （2）Line 177 mainly describes that we not only inject style features into the upsampling layer, but also inject them into the Controlnet control module through an independent cross-attention module. （3） Yes, we use the same F(S)', but with two independent A.

---

Q5. A quantitative comparison may help to demonstrate the style-content trade-off.

A5: We thank the reviewers for their constructive feedback. In addition to the previously reported quantitative ablation studies on content- and style-related hyperparameters , we have supplemented the following: (1) Quantitative results analyzing the impact of style token t,

(2) Performance metrics for text-driven stylized synthesis tasks

Collectively, these comprehensive quantitative analyses—supported by our visualization results—demonstrate CSGO's robustness in style transfer applications. We believe this expanded evidence addresses the reviewers' methodological concerns regarding the approach's validity.

---

Q6. inference time and training efficiency.

A6: Regarding the training efficiency of B-LoRA: On a single H800, it only takes 6 minutes to train a B-LoRA. Regarding the inference efficiency of CSGO: On a single H800, without any optimization measures, it takes 4.5 seconds at 512 resolution and 12 seconds at 1024 resolution to generate results. In fact, the inference time of CSGO is significantly lower than that of other style transfer models. Most other stylization methods rely on DDIM inversion, which almost requires doubling the time for DDIM inversion.

---

Q7. The authors have discussed the limitations of other failure cases and social impact.

A7: Thank you for the reviewers' suggestions. In CSGO, due to the introduction of Controlnet, the content signal is very strong, and unreasonable inference hyperparameters may lead to excessive content, that is, style loss; setting the content control hyperparameter to 0.5 will help. Our research is the first to verify that large-scale synthetic data can improve style transfer, providing a new baseline, which will be helpful to the community.

Dear Reviewer a2Ei,

We sincerely thank you again for reviewing our response. We are keen to confirm whether the answers and experiments provided have sufficiently addressed the relevant questions and concerns. Please do not hesitate to raise any further inquiries you may have. Thank you sincerely for your dedicated efforts and ongoing support.

With kind regards, The CSGO Author Team

Response to Reviewer TeEm

We sincerely appreciate the reviewers' recognition of the strengths of our method. Next, we will respond to the reviewers' questions and concerns one by one.

---

Q1. Some typos in the paper. For example, in line 97, loRA should be modified to LoRA. imges should be modifed to images.

A1: Thank you for the careful review. We will correct any typos.

---

Q2. The current form of Figure 3 is a little bit messy. Maybe the authors should consider re-draw the pipleline to make it more readable.

A2: We thank reviewers for their valuable suggestion. We will implement the proposed architectural decomposition separating control mechanisms into: (1) content control in downsampling blocks, (2) style control in upsampling blocks, and (3) dual-control fusion. Due to current response constraints, this refinement will be implemented in the camera-ready manuscript.

---

Q3. According to Figure 4, I can not find the strength of the proposed method. By contrast, the compared method is better regarding the generation quality, such as instantstyle.

A3: Thank you for your attention. In Figure 4, the advantages of our method over InstantStyle are as follows: In the first row, InstantStyle has the issue of "the trees next to the house being lost"; in the second row, it suffers from the loss of key facial attributes and insufficient style restoration; in the third row, InstantStyle loses background attributes; in the fourth row, its generated results lack the "stone texture" possessed by the style image. The reason for these differences is that InstantStyle relies on DDIM inversion based on content images to preserve content features, while DDIM inversion not only tends to lose content information of the original image but also takes almost twice as much time. Moreover, through large-scale training, the effectiveness of our proposed method in style transfer has been further improved.

---

Q4. Authors mainly explore the data-driven way to achieve the stylized image generation. I am thinking about, if there is a large amount of data, the model architecture could be more simplified. For example, we can simply use the architecture of MLLM, such as Kosmos-G and Bagel to achieve this task.

A4: We thank the reviewers for their insightful observations. We concur that integrated generation-understanding architectures (e.g., Bagel, GPT-4o) represent an emerging paradigm where downstream task incorporation can streamline model design. While we acknowledge that sufficient data could theoretically enable Bagel to perform style transfer through full fine-tuning, such adaptation to multi-image inputs would necessitate substantial computational resources—a promising but resource-intensive research direction. Our plugin-based approach offers distinct advantages: it preserves base model capabilities while enabling parameter-efficient adaptation. We are actively exploring unified architectures for stylization and multimodal understanding-to-generation pipelines to bridge this methodological gap.

In fact, our CSGO is still a basic and simple pipeline. We mainly verified that the automated large-scale construction of stylized datasets can significantly improve the effect of style transfer. We will actively explore a more elegant framework.

---

Q5. Could the authors explain more about the potential to expand the dataset, such as the possibility to scale the dataset to infinite amount of data. That could be more useful for the community.

A5: The scalability of IMAGStyle is architecturally facilitated through an automated pipeline. By training m content-specific and n style-specific B-LoRAs, we theoretically obtain m × n triplets (content image, style image, stylized output). Each additional B-LoRA linearly extends the combinatorial space by m or n triplets respectively. Crucially, B-LoRA training exhibits high efficiency—requiring merely ≈6 minutes per module on NVIDIA H800 hardware. Thus, given sufficient content/style images and GPU resources, the system can rapidly scale to massive triplet volumes. Furthermore, CSGO serves as a complementary expansion mechanism: rather than training new B-LoRAs, it enables immediate generation of novel triplets through direct application, substantially accelerating dataset growth with minimal marginal cost.

Dear Reviewer TeEm,

We have proactively addressed the concerns raised by the reviewers and engaged in discussions regarding the dataset expansion and pipeline improvements of CSGO. We hope our responses will meet with your approval.
We are keen to confirm whether the answers and experiments provided have sufficiently addressed the relevant questions and concerns. Please do not hesitate to raise any further inquiries you may have. Thank you sincerely for your dedicated efforts and ongoing support.

With kind regards, The CSGO Author Team

Response to Reviewer nay9

We appreciate the reviewer's recognition and affirmation of various aspects of this study; your positive comments have provided significant motivation for us to further improve the research. Regarding the shortcomings you pointed out, we will attach great importance to them and handle them carefully.

---

Q1. A total of 11,000 content images and 10,000 style images would theoretically yield 110 million potential combinations, yet the experimental output amounts to only 210,000 images. What explains this notable discrepancy between the theoretical combinatorial possibility and the actual production yield?

A1: Firstly, the theoretical combinatorial calculation is a brute - force estimate based on the simple multiplication of the number of content and style images. In practice, not all combinations are viable or meaningful. Our experimental design incorporates a screening mechanism to eliminate such unpromising combinations. This screening process significantly reduces the number of combinations that are actually processed and output. Secondly, resource constraints play a major role. Generating and processing each image combination demands substantial computational resources, including GPU time, memory, and storage. Given the limited hardware resources available for our experiments, we are unable to exhaustively compute all possible combinations within a limited time. To optimize resource utilization, we prioritize the combinations that are more likely to produce high - quality and relevant results according to our predefined criteria.

---

Q2. Why injecting style features into ControlNet warrants a more detailed analysis.

A2: We will elaborate on the reasons for choosing to pre-inject style features into ControlNet from two perspectives: (1) Style features require multi-level injection. Observations from InstantStyle experiments show that although style features can be condensed into certain two layers during downsampling, increasing the number of injection layers can significantly improve style similarity. Furthermore, ControlNet is essentially analogous to a part of the basic backbone network, such as the downsampling layers of U-Net. Moreover, for inpainting-oriented ControlNet, the downsampling layers themselves carry content information. Relying solely on feature injection in the upsampling layers may make it difficult to achieve high-quality style transfer. Therefore, injecting style features into the ControlNet branch, consistent with the injection method in InstantStyle, is intuitively reasonable. (2) The results of our ablation experiments are shown in the Table 1. The experimental results clearly demonstrate that injecting style features into the ControlNet branch can improve style similarity without causing style leakage or content loss.

Table 1. Ablation studies style control.

---

Q3. Table 2 requires a more detailed explanation. Table 2 references appear to be missing.There is a typo in line 250.

A3. Table 2 presents human evaluation results comparing our method against StyleShot [12], InstantStyle [32], and StyleAligned. We aggregated evaluators' preference distributions across three categories: preference for our method ("better"), baseline ("worse"), or perceptual equivalence ("tie"). Key conclusions emerge: (1) Our approach demonstrates statistically significant superiority in style transfer quality, suggesting the efficacy of large-scale automated dataset construction for this domain; (2) Human preferences strongly correlate with quantitative metrics (CSD: Content Structure Distance, CAS: Content Awareness Score), validating their utility as perceptual consistency indicators.

Thank you for the careful review. We will correct any typos.

Dear Reviewer nay9,

We have proactively addressed the concerns raised by the reviewers. We hope our responses will meet with your approval.
We are keen to confirm whether the answers and experiments provided have sufficiently addressed the relevant questions and concerns. Please do not hesitate to raise any further inquiries you may have. Thank you sincerely for your dedicated efforts and ongoing support.

With kind regards, The CSGO Author Team

Response to Reviewer Exzg

We thank reviewer for the constructive comments. We provide our feedbacks as follows.

---

Q1. Please provide a quantitative study showing how often the content LoRA and style LoRA truly isolate their intended factors.

A1: Regarding the risk of B-LoRA. Our conclusion is first derived from the original literature [1], which verifies that B-LoRA has good decomposition properties. In addition, to address reviewer concerns regarding evaluation robustness, we implement the following protocol: (1) 300 content LoRAs and 300 style LoRAs are combinatorially paired, generating stylized images per pair (total: 4500 images) (2) Three style consistency metrics are quantified: (i) Mean style similarity between outputs and reference style images (ii) Inter-group average of maximum style similarity (iii) Inter-group average of minimum style similarity (3) Two established style assessment models provide metric validation: - CSD [1] - OneIG-StyleEncoder [2] (Both output normalized scores ∈ [-1,1], where higher values indicate stronger style fidelity)

Table 1．The average style similarity of 300 groups of images, as well as the inter-group averages of the maximum and minimum similarity within each group.

Table 2．The average CAS score of 300 groups of images, as well as the inter-group averages of the maximum and minimum similarity within each group.

As shown in Table 1, for unfiltered outputs, both assessors report mean style similarity ≈0.6, confirming that B-LoRA syntheses preserve high style fidelity regardless of content alignment quality. On the other hand, we also present visualization results. The 10 sets of results shown in Figures 5-9 all indicate that the samples within these sets are relatively close to the original style images.

Finally, we present in Table 2 the average CAS score among these 300 groups, the inter-group average of the maximum CAS score within each group, and the inter-group average of the minimum CAS score within each group. It can be observed that among the generated results, we are able to obtain images with excellent alignment, so the inter-group average of the minimum CAS score within each group is 1.13. We show the visualization results of the CAS quantitative alignment in Figure 2 of the supplementary materials. It is quite consistent with the subjective perception that if the CAS is less than 1.2 points, the content can be considered relatively aligned. Therefore, from the experimental results of both aspects and the argumentation in the original article, it can be verified that B-LoRA can effectively decouple and combine content and style.

[1] Somepalli, Gowthami, et al. "Measuring style similarity in diffusion models." arXiv preprint arXiv:2404.01292 (2024). [2] Chang, Jingjing, et al. "OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation." arXiv preprint arXiv:2506.07977 (2025).

---

Q2. Why not incorporate a perceptual style metric like CSD alongside CAS when choosing the target image?

A2: The reasons for not using the CSD score for screening are threefold: (1) B-LoRA can easily generate results that are close to the original style images. Empirical evidence from Table 1 confirms that there is high style fidelity between the synthesized images and the reference images (with an average similarity > 0.6). (2) Based on the intrinsic properties of B-LoRA decomposition: style features exhibit significantly higher decomposition stability, while content features show substantially greater variance. Therefore, it is more important to prioritize content accuracy when considering screening results. Visual verification results can be found in Figures 4 to 9 of the supplementary materials. (3)The screening results using CAS have the highest alignment with manual screening. To evaluate the alignment between different screening strategies and human selections, we conducted some verification experiments. We introduced a*CAS + b*(1-CSD) to represent the final score used for screening. We filtered the results using four strategies respectively and calculated the alignment with the manually screened results (top 1), as shown in Table 3. It can be observed that due to the instability in the content LoRA of B-LoRA, the results screened using only CAS are actually the closest to the manual screening results. Therefore, we choose to screen the dataset solely based on the CAS score.

Table 3. The alignment between the screening results of different strategies and manual screening.

---

Q3. AdaIN should divide by standard deviation, not variance.

A3： We thank the reviewer for highlighting the precise terminology requirement; the manuscript has been corrected to explicitly reference "standard deviation" in the computation procedure.

Dear Reviewer Exzg,

We would like to express our gratitude to the reviewers for their valuable feedback. We have carefully responded to the issues of concern raised by the reviewers:

1. Quantitative study of B-LoRA,
2. why not incorporate a perceptual style metric like CSD alongside CAS when choosing the target image,
3. and some typos

We are keen to confirm whether the answers and experiments provided have sufficiently addressed the relevant questions and concerns. Please do not hesitate to raise any further inquiries you may have. Thank you sincerely for your dedicated efforts and ongoing support.

With kind regards, The CSGO Author Team

Response to Reviewer QqDz.

We thank reviewer for the constructive comments. We provide our feedbacks as follows.

---

Q1. CAS uses feature normalization to isolate content by removing style. Could you provide further justification or ablation studies on the effectiveness of this approach compared to other potential methods for disentangling content and style?

A1: We verified through preliminary experiments that the proposed CAS has a very high similarity to manual screening. Secondly, we set up a strict screening mechanism and verified the reliability of the screening results. It is hoped that these experiments can alleviate the reviewers' concerns about CAS. First, to assess whether CAS aligns with human screening results, we randomly curated 300 content-style pairs and generated stylized results using the B-LoRA combination method introduced in our paper. For each of these 300 groups, each group contain images as candiate image, we applied three distinct screening strategies to select the top-1 and top-2 results: (1) CAS scores computed using the CLIP visual encoder, (2) CAS scores computed using the DINOV2 model, and (3) manual human selection. The alignment rates between these computational methods (CLIP-based CAS and DINOV2-based CAS) and human choices are reported in Tables 1 (The visualization results (10 groups) are shown in Figures 5 to 9 of the supplementary material.).

Table 1. Alignment results between CLIP-based CAS, DINOV2-based CAS and users.

Experimental results demonstrate that DINOV2-computed CAS scores achieve an high alignment accuracy. Representative visual examples are provided in Figures 9–14. We analyzed the underlying factors contributing to this high alignment performance. As illustrated in Figure 3 of the supplementary material, we present the generated samples with their corresponding CAS-DINOV2 scores relative to the original content image. These scores exhibit strong consistency (aligning with subjective human assessment).

Second, the technical details of the dataset screened based on CAS scores: Selecting the sample with the lowest CAS score per group constitutes merely one candidate solution. In practice, we establish a global threshold by aggregating the 25th percentile of minimum CAS scores across all groups. Consequently, for each group, only those containing at least one sample below this threshold are retained; all others are discarded. Furthermore, we implement globally normalized scores as the evaluation metric to maintain cross-dataset consistency. This systematic approach significantly enhances the reliability of our proposed automated dataset construction framework.

---

Q2. Your ablation study shows that injecting style features into the ControlNet branch is beneficial. Did you observe any trade-offs? Specifically, does this pre-injection of style into a content-focused module ever lead to undesirable content distortions or artifacts?

A2: In our experiments, injecting style features into the Controlnet branch can further improve the quality of style transfer without generating artifacts. Since the proposed CSGO framework is trained end-to-end, unlike some training-free methods, it can correct some artifacts and abnormal states during training. The advantages of style feature injection in ControlNet are twofold: (1) Ablation studies confirm the efficacy of style feature injection in ControlNet, contingent upon consistent application during both training and inference phases. (2)Our pre-injection design preserves content integrity without inducing distortions or artifacts. As empirically validated in Figure 11 and Figure 12, content and style hyperparameters exclusively mediate the preservation fidelity of respective attributes in output images.

---

Q3. It has been mentioned that the model struggles with face stylization. Could you elaborate on the specific failure modes? Is this primarily an issue with the underlying ControlNet, a lack of sufficient facial diversity in the IMAGStyle dataset, or the feature separation itself?

A3: The suboptimal face stylization performance stems from two primary factors: First, the core limitation resides in training data paucity. As noted in Section 3, content prompts for our large-scale dataset derive from open-source salient object datasets (MSRA10K and MSRA-B). These collections exhibit strong bias toward non-human subjects, with severely insufficient portrait samples—directly causing deficient facial stylization data in our training corpus. Second, inherent instabilities in the B-LoRA module compound this issue. During content-style decoupled LoRA training, B-LoRA fails to consistently preserve critical portrait content—precisely retaining facial feature details constitutes an intrinsically challenging task [1]. Finally, we clarify that while our method exhibits elevated failure rates in portrait style transfer, it maintains generalization capabilities within this domain—as explicitly demonstrated by the successful cases in Row 2 of Figure 4.

[1]Frenkel, Yarden, et al. "Implicit style-content separation using b-lora." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

---

Q4. The paper acknowledges that it does not report error bars or conduct statistical significance tests for its experimental results. This makes it more difficult to assess the consistency and reliability of the reported performance gains over other methods.

A4: To address reviewer concerns regarding result stability, we compute the final metric as the mean of five independent evaluations with standard error bounds (Table 4). These results confirm that the CSGO model achieves both statistically significant improvements over baselines and consistent performance across trials. Table 4.More comprehensive quantitative comparison results between the proposed method and the other methods.

Dear Reviewer QqDz,

We would like to express our gratitude to the reviewers for their valuable feedback. We have carefully responded to the issues of concern raised by the reviewers:

1. The effectiveness of CAS,
2. explanations of CSGO's limitations,
3. explanations of style feature injection into ControlNet,
4. and error bars

We are keen to confirm whether the answers and experiments provided have sufficiently addressed the relevant questions and concerns. Please do not hesitate to raise any further inquiries you may have. Thank you sincerely for your dedicated efforts and ongoing support.

With kind regards, The CSGO Author Team