acceptodds
Under review as a conference paper at ICLR 2027

CoStitch: Disentangled Color Style and Stitch Texture Generation for Controllable Embroidery Synthesis

Abstract

Embroidery image synthesis aims to generate realistic embroidery images with specified color styles and stitch textures while preserving target pattern structure. However, existing methods typically encode color style and stitch texture within a unified appearance representation, hindering their disentanglement and independent reuse on novel patterns. To address this issue, we propose CoStitch, a framework for disentangled color style and stitch texture generation. CoStitch exploits three progressively enriched image states to provide differentiated supervision for the two attributes. Specifically, CoStitch constructs an aligned triplet comprising a sketch, a color design image, and an embroidery image, where the color design image serves as an intermediate state that distinguishes color-related variation from stitch-texture variation. CoStitch further performs joint image-text response analysis to identify color-sensitive and texture-sensitive blocks, defining attribute-specific learning subsets for Color LoRA and Stitch LoRA to reduce cross-attribute interference. The three image states are then reconstructed successively with corresponding LoRA updates, progressively learning color style and fine-grained stitch texture. Finally, texture enhancement stops sketch reconstruction and further refines stitch direction, length, density, and arrangement while preserving the learned color style. Experiments demonstrate that CoStitch preserves target pattern structure while enabling color style and stitch texture to be reused independently and flexibly combined across different references.

Primary area:
generative models
Keywords:
Embroidery Image Generation, Collaborative Generation, Color Transfer, Stitch Texture Transfer, Low-Rank Adaptation, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_acDKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a framework called CoStitch for disentangled color style and stitch texture generation in embroidery image synthesis. CoStitch consists of three key components: an attribute-sensitive block localization method for selecting Color LoRA and Stitch LoRA parameters, a progressive reconstruction strategy using aligned sketch, color design, and embroidery images, and a texture enhancement strategy for refining fine-grained stitch characteristics while preserving color appearance.

Critical strengths
  1. This paper addresses the disentanglement of color style and stitch texture in embroidery image synthesis, which is an interesting and practically meaningful problem.

  2. The proposed method enables independent control and flexible combination of color styles and stitch textures from different references.

  3. The paper is well-written and well-organized.

  4. Extensive experiments are conducted to evaluate the proposed method.

Critical weaknesses
  1. The experimental settings are not sufficiently detailed. The paper does not clearly report the number of reference triplets and target patterns used for evaluation. Moreover, Section 3.1 mentions using HED and Wilcom to construct aligned sketch, color design, and embroidery images, but the detailed data preparation process is unclear. More information about the dataset and evaluation protocol should be provided.

  2. Some more recent studies from 2026 are neither discussed nor compared in this paper, such as StyleGallery [1], CRAFT-LoRA [2], and StyleExpert [3].
    [1] StyleGallery: Training-free and Semantic-aware Personalized Style Transfer from Arbitrary Image References. CVPR 2026.
    [2] CRAFT-LoRA: Content-Style Personalization via Rank-Constrained Adaptation and Training-Free Fusion. CVPR 2026.
    [3] Mixture of Style Experts for Diverse Image Stylization. CVPR 2026.

  3. In Section 4.2.2, the authors mention "the reproduced first-stage EmoLoRA result." Does this mean that only the first stage of EmoLoRA is used for comparison? If so, why is the complete EmoLoRA method not considered?

  4. Although Figure 6 demonstrates cross-reference generation, the results are only evaluated qualitatively. More quantitative experiments should be conducted to evaluate whether changing the color reference affects stitch texture and vice versa. For example, the authors could fix one attribute while varying the other and measure the resulting changes using color and texture similarity metrics.

  5. Some experimental results deserve further discussion. As shown in Table 2, the complete method achieves a lower Silhouette Dice score (0.9593) than all the ablated variants, despite structure preservation being one of the main objectives. Moreover, although the quantitative results are promising, no user study is conducted to evaluate the perceptual quality and realism of the generated embroidery images.

Questions for authors

Please see Weaknesses.

Others:

What is the time efficiency of different methods? Some evaluations regarding this should be conducted.

3 comments

Weaknesses1. Experimental Scale and Reference Triplet Construction

Answer:

We thank the reviewer for the valuable suggestion.

Evaluation Protocol. Following the experimental settings of existing methods [1–3], we use 30 reference triplets and 40 unseen target sketches, resulting in 1,200 evaluation combinations. The original experimental data scales of the compared methods and the evaluation setting of CoStitch are summarized below:

MethodReference DataTest Data
B-LoRA [1]25 style reference images23 content subjects
StyleID [2]40 style reference images20 content images
EmoLoRA [3]30 embroidery images50 design images
Ours30 reference triplets40 target sketches

Each reference triplet is used for independent personalization training. All methods are evaluated on the same 1,200 reference-target combinations, and the six evaluation metrics are averaged over all 1,200 test results.

Data Preparation. Following practical industrial embroidery sampling workflows, we use Wilcom, which is employed in industrial embroidery digitization, to generate color design images IdI_d and their corresponding embroidery images IeI_e. We then apply an edge detection method [4] to extract sketches IsI_s from the embroidery images. Through data cleaning and filtering, we ensure that the triplets are structurally aligned and that the color design and embroidery images maintain consistent dominant colors.

References

[1] Y. Frenkel et al. Implicit Style-Content Separation Using B-LoRA. In Proceedings of the European Conference on Computer Vision (ECCV), 2024.

[2] J. Chung et al. Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.

[3] J. Ma et al. One-shot Embroidery Customization via Contrastive LoRA Modulation. ACM Transactions on Graphics, 44(6), Article 272, 2025.

[4] S. Xie and Z. Tu. Holistically-Nested Edge Detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015.

Weaknesses2. Comparison with Recent Related Work

Answer:

We thank the reviewer for the valuable suggestion.

Quantitative Comparison. To extend our comparison with recent studies [5–7], we have added a quantitative evaluation of StyleGallery [5] under the same evaluation protocol. The results are presented below:

MethodDice ↑HFRD ↓GLCM-TD ↓SIFID ↓ΔE₀₀ ↓EMD ↓
StyleGallery [5]0.96360.01380.213738.833311.106030.6068
Ours0.95930.00480.050814.107210.583630.0398

CoStitch outperforms StyleGallery on five of the six metrics, achieving lower HFRD, GLCM-TD, and SIFID for stitch-texture similarity, as well as lower Palette ΔE₀₀ and EMD for color consistency. Although StyleGallery achieves a slightly higher Silhouette Dice, CoStitch demonstrates stronger overall performance in stitch-texture and color matching.

Technical Comparison. We compare these methods in terms of their learning objectives and core technical mechanisms. The corresponding discussions have also been added to the Related Work section.

MethodLearning ObjectiveCore Technical Mechanism
StyleGallery [5]Content semantics and reference styleSemantic region matching and style-loss-guided diffusion sampling
CRAFT‑LoRA [6]Content and styleRank-constrained LoRA adaptation and training-free adapter fusion
StyleExpert [7]Diverse image stylesStyle encoding and dynamic mixture-of-experts routing
OursColor style and stitch textureReconstruction supervision using sketch, color design, and embroidery images, attribute-sensitive block localization, and separate optimization of color- and stitch-sensitive parameters

References

[5] B. He, Y. Ye, C. Liu, W. Wu, F. Liu, and Z. Cai. StyleGallery: Training-free and Semantic-aware Personalized Style Transfer from Arbitrary Image References. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 29092–29102, 2026.

[6] Y. Li, Y. Cai, and C. Zhang. CRAFT-LoRA: Content-Style Personalization via Rank-Constrained Adaptation and Training-Free Fusion. arXiv preprint arXiv:2602.18936, 2026.

[7] S. Zhu, Z. Ouyang, Y. Kang, Q. Wang, M. Zhou, B. Li, M.-M. Cheng, and Q. Hou. Mixture of Style Experts for Diverse Image Stylization. arXiv preprint arXiv:2603.16649, 2026.

Weaknesses3. Scope of the EmoLoRA Comparison

Answer:

We thank the reviewer for pointing this out.

The original experiments indeed included only the reproduced first-stage results of EmoLoRA. Since the complete training code of EmoLoRA has not been publicly released and its official pretrained weights correspond to only one stitch style, we reproduced its first stage based on the methodology described in the paper and evaluated it on the same 1,200 reference-target combinations to examine its generation performance across different stitch styles.

Additional Quantitative Evaluation. We further evaluated the official EmoLoRA model using its publicly released pretrained weights on the same 40 unseen target sketches, with the six metrics reported in Table 1. The results are presented below:

MethodDice ↑HFRD ↓GLCM-TD ↓SIFID ↓EMD ↓ΔE₀₀ ↓
Official EmoLoRA0.94720.02140.194145.421459.153211.7042
CoStitch (Ours)0.96620.00340.073514.955239.000515.3110

CoStitch achieves better Dice, HFRD, GLCM-TD, SIFID, and EMD scores, particularly on stitch-texture-related metrics, while the official EmoLoRA model achieves a lower Palette ΔE₀₀.

Weaknesses4. Quantitative Evaluation of Cross-Reference Control

Answer:

We thank the reviewer for the valuable suggestion.

Experimental Setup. We have newly added a quantitative cross-reference experiment using 40 unseen target sketches. Config A fixes the color reference while varying eight stitch-texture references, whereas Config B fixes the stitch-texture reference while varying eight color references. For each sketch, we generate eight embroidery images and report the mean and standard deviation of pairwise metric differences across reference combinations.

MetricConfig AConfig B
Palette EMD ↓13.2916 ± 1.349763.4266 ± 6.0085
Palette ΔE₀₀ ↓4.8900 ± 0.605521.2639 ± 2.3952
HFRD ↓0.006149 ± 0.0014990.001183 ± 0.000360
GLCM-TD ↓0.085504 ± 0.0195430.039867 ± 0.012628
SIFID ↓10.6759 ± 1.94779.2045 ± 2.7367

Results Analysis. Under Config A, the relatively low Palette EMD and Palette ΔE₀₀ indicate that varying stitch-texture references has a limited effect on color distribution. Under Config B, the relatively low HFRD and GLCM-TD indicate that varying color references has a limited effect on stitch-texture characteristics. These results support CoStitch's ability to control the two attributes independently.

Weaknesses5. Structural Metric Degradation and User Study

Answer:

We thank the reviewer for the valuable suggestions.

Structural Consistency Evaluation. To further evaluate boundary consistency between generated and target images, we have newly added the Boundary F1 (BF1) metric and evaluated 1,200 generated embroidery images using consistent mask extraction and distance tolerances of 1, 2, and 4 pixels. The results are presented below:

Metric1 px2 px4 px
BF1 ↑0.32530.56880.8278

BF1 increases from 0.3253 to 0.8278 as the tolerance increases from 1 to 4 pixels, indicating that the generated images exhibit local boundary deviations but achieve relatively high boundary consistency when small spatial deviations are allowed. This observation is consistent with the potential influence of fine-grained stitches on local contours.

User Study. We have newly conducted a perceptual evaluation involving 20 participants without professional embroidery experience. The evaluation pool contains 50 real embroidery images and 50 CoStitch-generated images. Each participant randomly selected 15 real and 15 generated images and rated their overall visual quality and perceived embroidery realism on a five-point scale. The results are presented below:

Image TypeVisual Quality ↑Embroidery Realism ↑
Real Embroidery4.354.17
CoStitch (Ours)4.324.15

CoStitch achieves average scores of 4.32 for visual quality and 4.15 for embroidery realism, approaching the corresponding scores of real embroidery images (4.35 and 4.17). These findings provide additional perceptual evidence supporting the visual quality and realism of CoStitch-generated embroidery images.

Q. Time Efficiency Comparison

Answer:

We thank the reviewer for the valuable suggestion.

Training Efficiency Evaluation. We have newly added a comparison of training times for personalization-based methods on an NVIDIA L20 GPU. The results are presented below:

MethodTraining Time
B-LoRA35 min
DB-LoRA37 min
PairCustomization26 min
Ours45 min

CoStitch requires slightly more training time than the compared methods due to its successive reconstruction and stitch-texture enhancement procedures.

Thanks for the authors' rebuttal. I will maintain my original score.

Reviewer_FZepOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper performs controllable embroidery style transfer by disentangling two attributes, color and texture. It is a two-stage model: the first stage performs color design, and the second stage transfers the stitch style. Qualitative and quantitative comparisons are used to show the model’s disentanglement ability and generation quality.

Critical strengths

An interesting topic of controllable embroidery style transfer with color and texture disentanglement. Color LoRA and Stitch LoRA are used to to separate color style and stitch texture learning. According to the experiment results, CoStitch maintains the target pattern’s structure and lets users apply the color of one reference and the stitches of another.

Critical weaknesses
  1. Overlap between the color and stitch LoRAs. The color and stitch LoRAs both act on the attention blocks at layers 7 and 11. During training, some of the color LoRA parameters are optimized by L_d together with the stitch LoRA. The authors later freeze some parameters, but I remain concerned that this design cannot fully remove the interference that arises when both LoRAs modify the same base feature transformation. An ablation comparing a disjoint block assignment with the current overlapping design is needed to support the disentanglement claim. In addition, a closely related dual-LoRA disentanglement method [1] should be included in the comparison and discussed. [1] Abdal, Rameen, et al. “Dynamic Concepts Personalization from Single Videos.” SIGGRAPH 2025.

  2. Blurring and inconsistency in the results. The final embroidery results are blurry in places and show many inconsistencies with the input:

  • In the top-left example of Fig. 1, the colors of the input flowers are not fully or reasonably transferred to the output.
  • In the bottom-right example of Fig. 1, the buildings and the bridge turn into something that looks like a forest.
  • Figs. 5, 7, and 9 all show deformed mouths and eyes. Mountains, trees, and clouds change in shape, and their number is not preserved.

The inconsistency grows as the input images become more complex, which raises concerns about the model’s stability. Also, the text prompts used in training and inference are not reported, so it is impossible to tell whether these inconsistencies come from the model itself or from variation in the text.

Questions for authors
  1. How are the text prompts designed? Are they written manually or generated by AI?

  2. The core of the paper is embroidery synthesis, so the results are textures with strongly regular patterns. Could you show results on irregular patterns such as leather or leaves, so that we can judge whether the method transfers to other textures?

  3. Can users specify the colors? My current impression is that the method mainly transfers the overall color scheme.

  4. Going from the color output to the final result often introduces blur or loss of detail because of the embroidery style. Is there a hyperparameter that controls this, such as the LoRA injection strength?

3 comments

Weakness 1. Overlapping Sensitive Blocks and Related Dual-LoRA Methods

Answer:

We thank the reviewer for the valuable suggestions.

Impact of Overlapping Sensitive Blocks. We clarify the overlapping block configuration and its optimization mechanism in the revised manuscript:

ResponseManuscript Revision
CoStitch separately selects color-sensitive and stitch-texture-sensitive blocks. Some blocks respond to both attributes, resulting in overlaps at the 7th and 11th attention blocks. The two LoRAs maintain independent parameters, while their joint optimization under LdL_d may introduce cross-attribute interference.Appendix A.2 (Attribute-Sensitive Block Configuration): The selected sensitive blocks remain fixed during training. Under LdL_d, Color LoRA parameters in color-sensitive blocks and all Stitch LoRA parameters are jointly optimized. During embroidery reconstruction, Color LoRA is frozen, while Stitch LoRA is updated within the selected texture-sensitive blocks.

To evaluate the impact of overlapping sensitive blocks, we conducted an additional ablation experiment using non-overlapping block assignments. The quantitative results are presented below:

MethodDice ↑HFRD ↓GLCM-TD ↓SIFID ↓ΔE₀₀ ↓EMD ↓
Non-overlapping Blocks0.98380.00360.214731.772819.105051.9917
Ours0.95930.00480.050814.107210.583630.0398

CoStitch achieves better GLCM-TD, SIFID, Palette ΔE₀₀, and Palette EMD, while the non-overlapping configuration performs better in Dice and HFRD. These results demonstrate that the overlapping configuration achieves stronger color and stitch-texture matching, while maintaining competitive structural preservation. We have incorporated this ablation into Section 4.3.

Comparison with Related Dual-LoRA Methods. We clarify the relationship between CoStitch and Set-and-Sequence [1] in the revised manuscript:

ResponseManuscript Revision
Both methods adopt sequential learning. However, Set-and-Sequence [1] models dynamics through residual coefficients over a shared low-rank basis, whereas CoStitch learns color style and stitch texture through successive reconstruction and attribute-sensitive parameter optimization.Section 2.3 (Related Work): Set-and-Sequence [1] disentangles static appearance and dynamic information by first learning a shared low-rank appearance basis from unordered video frames, then freezing it and modeling dynamics through motion residual coefficients.

Reference

[1] R. Abdal et al. Dynamic Concepts Personalization from Single Videos. SIGGRAPH 2025.

Weakness 2. Blurring and Structural Inconsistencies

Answer:

We thank the reviewer for the careful observations.

Boundary Consistency Evaluation. CoStitch aims to independently learn and flexibly transfer reference color styles and stitch textures, with color transfer focusing on overall color relationships. To further investigate the local blurring and structural changes highlighted by the reviewer, we introduced the Boundary F1 (BF1) metric to evaluate boundary consistency between generated and target images under different pixel distance thresholds.

Let BgB_g and BtB_t denote the boundaries extracted from the generated and target masks, respectively. Boundary precision and recall are defined as:

Pb(τ)=∣{p∈Bg:d(p,Bt)≤τ}∣∣Bg∣,P_b(\tau)=\frac{|\{p\in B_g:d(p,B_t)\leq\tau\}|}{|B_g|},

Rb(τ)=∣{q∈Bt:d(q,Bg)≤τ}∣∣Bt∣.R_b(\tau)=\frac{|\{q\in B_t:d(q,B_g)\leq\tau\}|}{|B_t|}.

BF1 is calculated as their harmonic mean:

BF1(τ)=2Pb(τ)Rb(τ)Pb(τ)+Rb(τ)+ϵ.\mathrm{BF1}(\tau)=\frac{2P_b(\tau)R_b(\tau)}{P_b(\tau)+R_b(\tau)+\epsilon}.

Here, d(p,B)d(p,B) denotes the minimum Euclidean distance from pixel pp to boundary set BB, and τ\tau is the distance tolerance in pixels. Using a consistent mask extraction procedure, we evaluated 1,200 images generated by CoStitch. The results are presented below:

Metric1 px2 px4 px
BF1 ↑0.32530.56880.8278

As the boundary matching tolerance increases from 1 px to 4 px, BF1 rises from 0.3253 to 0.8278, indicating that some local boundary deviations between generated and target images fall within the larger matching tolerance. These results provide a quantitative assessment of boundary consistency across different spatial tolerances.

Text Prompt Design. We clarify the text prompt construction during training and inference in the revised manuscript:

ResponseManuscript Revision
CoStitch uses manually designed prompt templates rather than AI-generated prompts. Training and inference follow the same prompt structure and attribute trigger tokens, with only the subject description adapted to the target pattern.Appendix A.1 (Model and Training Settings): We adopt three manually designed prompt templates: "a [object]", "a [object] in [col] style", and "a [object] in [col] [emb] style", where [object] denotes the subject category. During inference, the same templates are retained with the target subject substituted.

Q1. Text Prompt Design

Answer: We thank the reviewer for the question. We have clarified the prompt design in the revised manuscript:

ResponseManuscript Revision
CoStitch uses manually designed prompt templates rather than AI-generated prompts. For example, "a clown," "a clown in [col] style," and "a clown in [col] [emb] style."Appendix A.1 (Model and Training Settings): We adopt three manually designed prompt templates: "a [object]," "a [object] in [col] style," and "a [object] in [col] [emb] style," where [object] denotes the target subject.

Q2. Generalization to Irregular Textures

Answer: We thank the reviewer for the valuable suggestion. We conducted additional experiments on leaf, fine-grained leather, and wrinkled leather textures, using one reference sample per material and the same 40 target sketches. The quantitative results are presented below:

MaterialDice ↑GLCM-TD ↓SIFID ↓
Leaf0.98400.144711.7085
Fine-grained Leather0.98510.252415.8134
Wrinkled Leather0.98450.158415.1411

All reported values are averages. We have added the corresponding discussion to the revised manuscript:

ResponseManuscript Revision
The results suggest that CoStitch has the potential to generalize beyond embroidery through one-shot adaptation. However, performance varies across materials, and preserving complex details and large-scale wrinkles remains challenging.Section 4.4 (Extended Experiments): Additional experiments on leaf, fine-grained leather, and wrinkled leather suggest potential generalization to irregular textures through one-shot adaptation, although preserving complex details and large-scale wrinkles remains challenging.

Q3. Specific Color Control

Answer: We thank the reviewer for the question. We have clarified the scope and limitations of color control:

ResponseManuscript Revision
CoStitch supports reference-driven color transfer, allowing users to guide the overall color scheme through different color design references. However, precise specification of individual colors or region-level color assignments is not currently supported.Section 5 (Conclusion): CoStitch enables reference-driven color transfer through different color design references, but primarily controls the overall color scheme and does not currently support precise specification of individual colors or region-level color assignments.

Q4. Controlling Blurring and Detail Loss

Answer: We thank the reviewer for the question. We have clarified the adjustable stitch texture strength during inference:

ResponseManuscript Revision
The stitch texture strength parameter λtex\lambda_{\mathrm{tex}} controls the injection strength of Stitch LoRA during inference. Reducing λtex\lambda_{\mathrm{tex}} weakens the transferred stitch texture and may alleviate local blurring or detail loss.Appendix A.3 (Inference Details): The stitch texture strength parameter λtex\lambda_{\mathrm{tex}} controls the injection strength of Stitch LoRA during inference. Lower values may reduce local blurring and detail loss, but also weaken the transferred stitch texture.

Comments on the paper

9 comments

Weaknesses 1. Disentanglement and Dependency on External Triplets

Answer:

We thank the reviewer for the valuable comments.

We have clarified the disentanglement mechanism and the construction of aligned triplets as follows:

ResponseManuscript Revision
CoStitch aims to disentangle the learning of color style and stitch texture from a single embroidery image. Thus, starting from an embroidery image IeI_e, we construct a triplet consisting of a sketch IsI_s, a color design IdI_d, and an embroidery image IeI_e to provide differentiated supervision. Meanwhile, by separately learning color-sensitive and texture-sensitive parameters, the model enables the two attributes to be reused independently.Section 3.1: We have revised the description to clarify the disentanglement objective and its implementation: To achieve the disentanglement of color style and stitch texture, CoStitch employs differentiated supervision from aligned sketch, color design, and embroidery images, together with separate learning of color-sensitive and texture-sensitive parameters, allowing the two attributes to be learned and reused independently.
CoStitch follows practical industrial embroidery design and sampling workflows. We employ Wilcom, which is used in industrial embroidery digitization, to generate colored embroidery images and corresponding design images. We then apply an edge detection algorithm [1] to extract sketches from the embroidery images. The triplets are cleaned and filtered to ensure structural consistency among IsI_s, IdI_d, and IeI_e, as well as consistency in dominant colors between IdI_d and IeI_e.Dataset Construction: We have added the following description: Following practical industrial embroidery design and sampling workflows, we employ Wilcom, which is used in industrial embroidery digitization, to generate colored embroidery images and their corresponding design images. We then apply an edge detection algorithm [1] to extract sketch images from the embroidery images. To ensure triplet quality, we clean and filter the data to maintain structural consistency among the three images, as well as consistency in dominant colors between the color design and embroidery images.

References

[1] S. Xie and Z. Tu. Holistically-Nested Edge Detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015.

Weaknesses 2. Stitch Geometry Is Neither Modeled nor Measured

Answer:

We thank the reviewer for the valuable comments.

Scope of Stitch Geometry Modeling. CoStitch aims to separately learn color style and stitch texture from a single embroidery reference and transfer them to different target patterns, rather than explicitly modeling the geometry of stitch orientation, length, and density.

Quantitative Evaluation of Stitch Characteristics. To further evaluate the local stitch characteristics of generated embroidery images, we introduce two objective metrics: Stitch Orientation Coherence Distance (SOCD) and Stitch Density Scale Distance (SDSD).

SOCD measures differences in local stitch orientation coherence between generated and reference embroidery images. Given the eigenvalues λ1(p)≥λ2(p)\lambda_1(p)\geq\lambda_2(p) of the structure tensor at pixel pp, local orientation coherence is calculated as:

c(p)=λ1(p)−λ2(p)λ1(p)+λ2(p)+ϵ.c(p)=\frac{\lambda_1(p)-\lambda_2(p)}{\lambda_1(p)+\lambda_2(p)+\epsilon}.

The distributions of local coherence values are then compared using the 1D Wasserstein distance:

SOCD⁡(G,R)=W1(PGc,PRc).\operatorname{SOCD}(G,R)=W_1(P_G^c,P_R^c).

SDSD measures differences in local texture scale characteristics, such as stitch thickness and spacing, through spectral analysis. For each valid foreground window kk, we calculate the high-frequency power spectrum:

Pk(u,v)=∣F{w⊙(Ik−Iˉk)}(u,v)∣2P_k(u,v)=\left|\mathcal F\{w\odot(I_k-\bar I_k)\}(u,v)\right|^2

The radial spectral centroid is then computed as:

fk=∑(u,v)∈Br(u,v)Pk(u,v)∑(u,v)∈BPk(u,v).f_k=\frac{\sum_{(u,v)\in\mathcal{B}}r(u,v)P_k(u,v)}{\sum_{(u,v)\in\mathcal{B}}P_k(u,v)}.

Here, r(u,v)r(u,v) denotes the radial frequency and B\mathcal{B} denotes the selected frequency band. SDSD is defined as:

SDSD⁡(G,R)=W1(PGf,PRf).\operatorname{SDSD}(G,R)=W_1(P_G^f,P_R^f).

Comparative Results. Here, GG and RR denote the generated and reference embroidery images, respectively. PGcP_G^c and PRcP_R^c represent their respective distributions of local orientation coherence, while PGfP_G^f and PRfP_R^f represent their distributions of radial spectral centroids. Lower values indicate better agreement with the reference for both metrics. The comparative results are presented below:

MethodSOCD ↓SDSD ↓
Pair Customization0.083860.03419
DB-LoRA0.104130.03151
RB-Modulation0.153050.03454
StyleGallery0.181550.04126
B-LoRA0.191320.03361
StyleID0.218040.06163
InstantStyle0.236530.04264
Ours0.083210.03134

We have added the following analysis to the experimental section:

As shown in Table 1, CoStitch achieves the lowest SOCD and SDSD, indicating closer agreement with reference embroidery images in local stitch orientation coherence and texture scale characteristics.

Weaknesses 3. Training and Inference Discrepancy

Answer:

We thank the reviewer for the valuable suggestions.

Use of ControlNet During Inference. Previous studies [2–4] have combined trained models with pretrained ControlNet during inference to provide structural guidance. Our proposed CoStitch follows a similar strategy. The training and inference strategies are summarized below:

MethodTraining and Inference Strategy
OPS-GA [2]Uses Canny ControlNet for structural guidance during portrait stylization.
Style Composition [3]Uses depth-conditioned ControlNet for structural coherence during multi-style generation.
EmoLoRA [4]Combines Canny and Tile ControlNet to preserve design contours and content during embroidery generation.
OursCoStitch combines with pretrained ControlNet during inference to preserve the structure of target sketches and generate embroidery images.The term “img2img” was inaccurately used to describe the overall inference pipeline in the original manuscript and has been removed to avoid ambiguity.

Color Design Distribution Discrepancy. To investigate the discrepancy between the generated color design used during inference and the Wilcom color design used to train Stitch LoRA, we conduct a quantitative analysis of their color distributions. Palette EMD and Palette ΔE₀₀ measure the color differences between generated color designs and their corresponding Wilcom designs:

MethodPalette EMD ↓Palette ΔE₀₀ ↓
Pair Customization49.080115.0615
DB-LoRA49.727214.6750
RB-Modulation43.019312.7530
StyleGallery41.802312.9793
B-LoRA59.698521.0733
StyleID54.972816.1110
InstantStyle41.574111.2206
Ours32.987610.9885

We have added the following analysis to the experimental section:

CoStitch achieves the lowest Palette EMD and Palette ΔE₀₀, indicating that its generated color designs are closer to the corresponding Wilcom designs in terms of color distribution. This suggests a smaller color difference between the color designs used during training and those generated during inference.

References

[2] X. Wang et al. One-shot Portrait Stylization via Geometric Alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 4964–4973, 2026.

[3] J. Lee et al. Style Composition Within Distinct LoRA Modules for Traditional Art. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pp. 4497–4504, 2025.

[4]J. Ma et al. One-shot Embroidery Customization via Contrastive LoRA Modulation. ACM Transactions on Graphics, 44(6), Article 272, 2025.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.