acceptodds
Under review as a conference paper at ICLR 2027

SimSew: Towards Simulation-Ready Sewing Pattern Generation

Abstract

In this paper, we focus on sewing pattern generation for digital garments. Existing methods are primarily trained with supervised objectives on ground-truth sewing patterns. However, such objectives do not directly constrain key factors related to simulation success, such as garment constructability and mesh quality. Tackling this issue, we present SimSew, a reinforcement-learning post-training framework for simulation-ready sewing pattern generation. SimSew directly incorporates downstream garment-construction feedback into post-training, encouraging the generator to produce patterns that are more suitable for physical simulation. To provide informative learning signals, we introduce failure-aware diagnostic rewards for unsuccessful generations and mesh-quality rewards for successfully constructed garments. The former differentiates diverse construction failures even when all sampled candidates are unsuccessful, while the latter further promotes high-quality meshes for downstream simulation. Extensive experiments demonstrate that SimSew improves downstream simulation readiness while preserving sewing-pattern quality. On GCD-MM, SimSew increases the simulation success rate from to ; reward-aligned evaluation further improves One-to-one Stitch from to and reduces Corner MAE from to .

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Sewing Pattern Generation, Reinforcement Learning, Garment Simulation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_wNCrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SimSew, an RL-based post-training framework that enhances SewFusion with simulation-aware rewards. The paper introduces three categories of rewards and jointly optimizes the network using their weighted combination.

Critical strengths
  1. The paper is well-written.
  2. The observation that reference fidelity does not necessarily reflect simulation readiness is interesting and potentially important.
Critical weaknesses
  1. The paper introduces an RL-based post-training module built upon the SewFusion architecture. However, SimSew does not consistently outperform SewFusion across all metrics, particularly under the editing setting. It is recommended to investigate setting-specific reward weighting or adaptation strategies.

  2. The observation that reference fidelity does not necessarily reflect simulation readiness is interesting and potentially important. It is recommended to supplement additional statistical evidence, including correlation analyses between reference-based metrics and simulation success, together with failure rate statistics for different error types.

  3. The paper defines three top-level rewards decomposed into seven sub-rewards. Nevertheless, weighted-sum aggregation of multiple rewards cannot guarantee simultaneous improvements over all objectives. Certain reward terms may be hard to optimize or mutually conflicting, which may cause unstable optimization or unwanted trade-offs. It is recommended to explicitly address such reward conflicts.

  4. Moreover, the reward formulation relies on several hand-crafted hyperparameters that can alter interactions between reward components. It is recommended to conduct reward-weight sensitivity analysis and examine potential gradient conflicts or competing objectives in the reinforcement learning pipeline.

Questions for authors

See weekness.

4 comments
AuthorsOpenReview ↗

Pointwise Response to Reviewer wNCr (Part 1/2)

We are grateful to the reviewer for the insightful feedback, which has helped us improve our work. We address each concern individually below. Unless otherwise specified, the additional experiments on GCD-MM are conducted under the image setting, using 500 training steps and 50% of the dataset for both training and evaluation.

Response to 'Critical Weaknesses':

The paper introduces an RL-based post-training module built upon the SewFusion architecture. However, SimSew does not consistently outperform SewFusion across all metrics, particularly under the editing setting. It is recommended to investigate setting-specific reward weighting or adaptation strategies.

A1:

We conducted additional reward weight sensitivity experiments under both the image-conditioned and editing settings on GCD-MM. Specifically, we varied the weights of the Pattern-Format, Stitch-Aware, and Geometry-Aware rewards while keeping the random seed and other training configurations unchanged. The quantitative results are presented below.

Image-settingλfmt\lambda_{\mathrm{fmt}}λstitch\lambda_{\mathrm{stitch}}λgeom\lambda_{\mathrm{geom}}Panel Acc. (%) ↑Edge Acc. (%) ↑Stitch Acc. (%) ↑Vertex L2 ↓SSR (%) ↑
Format-low0.100.450.4593.1687.1583.723.730876.03
Format-high0.600.200.2093.2187.1783.703.747275.97
Stitch-low0.450.100.4593.0087.1983.503.795776.09
Stitch-high0.200.600.2093.1786.9183.713.731975.95
Geometry-low0.450.450.1093.2087.1383.793.752976.05
Geometry-high0.200.200.6093.5987.0283.753.763776.01
Default1/31/31/393.5287.2283.873.797676.13
Editing-settingλfmt\lambda_{\mathrm{fmt}}λstitch\lambda_{\mathrm{stitch}}λgeom\lambda_{\mathrm{geom}}Panel Acc. (%) ↑Edge Acc. (%) ↑Stitch Acc. (%) ↑Vertex L2 ↓SSR (%) ↑
Format-low0.100.450.4588.1696.2597.326.803873.09
Format-high0.600.200.2088.0996.2097.336.816573.18
Stitch-low0.450.100.4588.1896.4597.226.903173.17
Stitch-high0.200.600.2088.1596.5897.347.118373.13
Geometry-low0.450.450.1088.1796.6897.356.991573.15
Geometry-high0.200.200.6088.1196.5697.297.194073.10
Default1/31/31/388.0696.2997.376.972973.24

The results demonstrate that SimSew maintains stable SSR across different reward configurations, with equal weights achieving the highest SSR in both settings. These findings demonstrate the robustness of our reward design to weight variations. Therefore, we retain equal weights across both settings rather than introducing additional setting-specific strategies.

We have included these results in the revised manuscript. (Sec. 5.3, line 519)

The observation that reference fidelity does not necessarily reflect simulation readiness is interesting and potentially important. It is recommended to supplement additional statistical evidence, including correlation analyses between reference-based metrics and simulation success, together with failure rate statistics for different error types.

A2:

  • Correlation Analysis. We analyze sample-level Pearson and Spearman correlations between reference-based metrics and simulation success.
Pearson r / Spearman ρPanel Acc.Edge Acc.Stitch Acc.Vertex L2
SSR0.39 / 0.410.46 / 0.510.47 / 0.55-0.23 / -0.31

The results show weak-to-moderate correlations, indicating that reference-based metrics are associated with simulation success but do not fully capture simulation readiness.

  • Failure Rate Statistics by Error Type. We further analyzed two major simulation failure types: StitchingError and DegenerateTrianglesError.
MethodStitchingError (Failed/All)DegenerateTrianglesError (Failed/All)
SewFusion421/2907 (14.48%)605/2907 (20.81%)
SimSew (Ours)346/2907 (11.90%)472/2907 (16.24%)

These results demonstrate that our simulation-aware RL framework effectively reduces both stitching-related and geometry-related failures, further supporting its effectiveness in improving downstream simulation readiness.

AuthorsOpenReview ↗

Pointwise Response to Reviewer wNCr (Part 2/2)

The paper defines three top-level rewards decomposed into seven sub-rewards. Nevertheless, weighted-sum aggregation of multiple rewards cannot guarantee simultaneous improvements over all objectives. Certain reward terms may be hard to optimize or mutually conflicting, which may cause unstable optimization or unwanted trade-offs. It is recommended to explicitly address such reward conflicts.

A3:

To investigate potential conflicts, we analyze gradient conflicts over 100 RL training steps. At each step, we compute reward-specific policy gradients on the same rollouts, and measure their pairwise cosine similarities. We report the average cosine similarity across the 100 steps.

Reward PairCosine Similarity ↑
Format vs. Stitch0.1063
Format vs. Geometry0.0827
Stitch vs. Geometry0.1159

The results suggest that the three reward objectives are generally compatible. The FlowOPD framework offers a potential direction for further mitigating these conflicts. However, exploring this approach is beyond the scope of the current manuscript. We thank the reviewer for this valuable suggestion and will investigate this direction in future work.

Moreover, the reward formulation relies on several hand-crafted hyperparameters that can alter interactions between reward components. It is recommended to conduct reward-weight sensitivity analysis and examine potential gradient conflicts or competing objectives in the reinforcement learning pipeline.

A4:

Thank you for this valuable and constructive suggestion. We have conducted additional reward weight sensitivity experiments and gradient conflict analyses to address these concerns. Please refer to A1 and A3 for detailed results and discussions.

Thank you for the additional analyses. This additional experiment cannot verify the soundness of compatible objectives or weighting strategy (average cosine similarity of gradients). Given that the pipeline is standard GRPO with some hand-crafted rewards, I remain concerned about novelty and keep my original score.

We thank the reviewer for the thoughtful comment. To the best of our knowledge, SimSew is the first work to apply reinforcement learning to improve the simulation success rate of generated sewing patterns. We argue that the main contribution of our work lies in designing effective rewards to address this challenge. Regarding potential conflicts among multiple rewards, existing works such as FlowOPD offer promising solutions. We plan to explore this direction in future work, as it is out of scope of this work.

Reviewer_9SBwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper fine-tunes SewFusion, a multimodal sewing pattern generator, with reinforcement learning to reduce failures. It uses GRPO with a LoRA adapter. For each input, the model samples eight patterns, and GRPO raises the probability of those that score above the group average. The reward has three parts, and most of them compare the generated pattern with the reference pattern. A pattern-format term checks the numbers of panels and edges. A stitch-aware term checks the number of stitches, whether each stitched edge has exactly one correct partner, and whether the length ratio of stitched edges matches the reference. A geometry-aware term checks panel size and corner angles. It also rewards the quality of the worst-shaped mesh triangles, but only when simulation succeeds. Compared with SewFusion, the simulation success rate (SSR) on GCD-MM rises in all four input settings, for example from 71.89% to 76.92% for image input and from 58.52% to 65.28% for text input. On SewFactory it rises from 88.72% to 89.86%.

Critical strengths
  1. Simulation readiness is a real problem for sewing pattern generators, and the paper states it clearly.
  2. The reward is dense. It still gives a learning signal when simulation fails.
  3. The SSR gains on GCD-MM appear for all four input settings (image, text, image-text and editing).
Critical weaknesses
  1. The technical contribution is limited and partly undescribed. The method is standard GRPO on one SFT model with a hand-designed reward. The continuous geometry is generated by flow matching, but the paper does not explain how the policy ratio is computed for it.
  2. The claim to be "the first RL framework that improves downstream simulation success through dedicated reward design" should also be limited to the domain of sewing patterns. DSO (Li et al., 2025) already fine-tunes 3D generators with physics simulation feedback.
  3. Simple baselines are missing. The ablation in Table 2 suggests that most of the gain comes from two terms. Without the stitch-count term, SSR drops from 76.13 to 73.63, and without the mesh term, it drops to 73.54. Removing the whole geometry reward gives 73.49, so the size and corner terms add little beyond the mesh term. The mesh term is zero unless simulation succeeds, so it acts largely as a success reward. Please compare with (a) a success-only reward, (b) best-of-N sampling from SewFusion with the simulator as a filter, at the same number of simulator calls, and (c) fine-tuning on successful samples only.
  4. The paper reads as unfinished. Appendix B is empty and has the same title as Appendix C. The reward ablation in Table 2 is never discussed, and a sentence in Section 5.3 is cut off (“reported in.”). Appendix D.1 refers to “Fig. ??”. “Meshgen” in Section 4.2 is not defined. The introduction cites BCNet, GarmentNets, GarmentTracking, CLOTH4D, and DeClotH as methods trained to match annotated sewing patterns, but none of these works predict sewing patterns, and CLOTH4D is a dataset.
Questions for authors
  1. How is the likelihood ratio computed for the flow-matching geometry, and which sampling steps are stochastic?
  2. How is SSR defined, which simulator and settings does it use, and how long does one reward evaluation take?
  3. What are λ, ω and γ, and how were they chosen?
  4. Which dataset and input setting do Table 2, Table 3 and Table 4 use? Why does the full model in Table 2 differ from Table 1?
5 comments
AuthorsOpenReview ↗

Pointwise Response to Reviewer 9SBw (Part 1/3)

Response to 'Critical Weaknesses':

We appreciate the reviewer's valuable comments and provide point-by-point responses to the concerns raised below. Unless otherwise specified, the additional experiments on GCD-MM are conducted under the image setting, using 500 training steps and 50% of the dataset for both training and evaluation.

The technical contribution is limited and partly undescribed.

A1:

  • Technical Contribution. Our main contribution is to improve the simulation success rate of generated sewing patterns through simulation-aware reward design. To this end, we introduce three categories of simulation-aware rewards and employ GRPO-based RL to incorporate simulation feedback into policy optimization. Initially, we stated this in line 54, line 81 and line 123.
  • Effectiveness of Simulation-aware Reward Design. As shown in Table 3, removing the Pattern-Format, Stitch-Aware, and Geometry-Aware rewards reduces SSR from 76.13% to 75.91%, 73.60%, and 73.49%, respectively, demonstrating the individual contributions of all three reward components.
  • Generalizability Across Different Generators. We further applied our simulation-aware rewards design to AIpparel and SewingLDM, with quantitative results presented below.
MethodPanel Acc. (%) ↑Edge Acc. (%) ↑Stitch Acc. (%) ↑Vertex L2 ↓SSR (%) ↑
SewingLDM76.2372.6571.945.940856.33
SewingLDM + SimSew Rewards77.9173.8473.975.579260.71
AIpparel86.3480.9573.495.460566.07
AIpparel + SimSew Rewards87.0582.7275.134.965770.12
SimSew (Ours)93.5287.2283.873.797676.13

The results demonstrate that our approach consistently improves simulation success rates across different generative models, confirming its generalizability beyond SewFusion.

The method is standard GRPO on one SFT model with a hand-designed reward.

A2:

Our primary focus is on designing effective sewing pattern-specific rewards to improve the simulation success rate of generated sewing patterns. As stated in the manuscript (Sec. 1, line 79), GRPO is simply the RL framework we adopt to optimize these rewards, rather than our main contribution.

The continuous geometry is generated by flow matching, but the paper does not explain how the policy ratio is computed for it.

A3:

We introduce stochastic Gaussian SDE transitions to compute the flow-matching policy ratio. Specifically, geometry generation uses 12 flow steps, with a randomly selected two-step stochastic window (noise scale η=0.2\eta=0.2); the remaining 10 steps use deterministic Euler integration. For each stochastic transition, the policy ratio is computed as:

ρi,k=exp⁡(ℓθ,i,k−ℓθold,i,k).\rho_{i,k} = \exp\left(\ell_{\theta,i,k} - \ell_{\theta_{\mathrm{old}},i,k}\right).

Here, ℓθ,i,k\ell_{\theta,i,k} denotes the dimension-normalized Gaussian transition log probability evaluated on the same sampled transition under the current policy. The ratio is subsequently used in the transition-level GRPO objective with PPO clipping.

We have supplemented a more detailed description in the appendix of the revised manuscript. (Appendix. B.1, line 715)

The claim to be "the first RL framework that improves downstream simulation success through dedicated reward design" should also be limited to the domain of sewing patterns. DSO (Li et al., 2025) already fine-tunes 3D generators with physics simulation feedback.

A4:

Thank you for pointing this out, and we apologize for the confusion. We have explicitly limited our novelty claim to the domain of sewing pattern generation in the revised manuscript. We have revised this statement in the manuscript accordingly. (Sec. 1, line 81)

AuthorsOpenReview ↗

Pointwise Response to Reviewer 9SBw (Part 2/3)

Response to 'Critical Weaknesses':

Simple baselines are missing. The ablation in Table 2 suggests that most of the gain comes from two terms. Without the stitch-count term, SSR drops from 76.13 to 73.63, and without the mesh term, it drops to 73.54. Removing the whole geometry reward gives 73.49, so the size and corner terms add little beyond the mesh term. The mesh term is zero unless simulation succeeds, so it acts largely as a success reward. Please compare with (a) a success-only reward, (b) best-of-N sampling from SewFusion with the simulator as a filter, at the same number of simulator calls, and cc fine-tuning on successful samples only.

A5:

Thank you for your constructive suggestions. Following your advice, we evaluated three additional baselines on the GCD-MM dataset under the image-conditioned setting.

  • Success-only Reward. We replace our dense simulation-aware rewards with a binary simulation-success reward while keeping the GRPO training configuration unchanged.
  • Best-of-8 Sampling. We generate multiple candidates using SewFusion and select simulation-successful samples, matching the total number of simulator calls.
  • Successful-only SFT. We further evaluated supervised fine-tuning on high-quality generated samples selected through simulation-aware filtering (best-of-8 selection), without GRPO-based policy optimization.
MethodPanel Acc. (%) ↑Edge Acc. (%) ↑Stitch Acc. (%) ↑Vertex L2 ↓SSR (%) ↑
Success-only Reward91.0585.3682.034.236472.76
Best-of-8 Sampling92.7586.3182.354.080572.20
Successful-only SFT92.7686.1982.434.030871.91
SimSew (Ours)93.5287.2283.873.797676.13

The results show that SimSew achieves the highest SSR, demonstrating that our fine-grained simulation-aware rewards provide more effective learning signals than binary success feedback alone, while GRPO-based optimization offers substantial benefits beyond inference-time filtering and supervised fine-tuning on successful samples.

The paper reads as unfinished. Appendix B is empty and has the same title as Appendix C. The reward ablation in Table 2 is never discussed, and a sentence in Section 5.3 is cut off (“reported in.”). Appendix D.1 refers to “Fig. ??”. “Meshgen” in Section 4.2 is not defined. The introduction cites BCNet, GarmentNets, GarmentTracking, CLOTH4D, and DeClotH as methods trained to match annotated sewing patterns, but none of these works predict sewing patterns, and CLOTH4D is a dataset.

A6:

Thank you for pointing out these errors. We have carefully revised the manuscript to address all the issues raised.

  • Removed the empty Appendix B and corrected the appendix numbering.
  • Added a detailed discussion of the reward ablation results of Table 2 (line 514).
  • Completed the truncated sentence in Sec. 5.3 (line 531) and corrected the missing figure reference to Fig. 5 (line 883).
  • Clarified the definition of Meshgen in Sec. 4.2 (line 315).
  • Corrected the inappropriate citations in the Sec. 1 (line 44), distinguishing sewing pattern generation from garment reconstruction and related datasets.
AuthorsOpenReview ↗

Pointwise Response to Reviewer 9SBw (Part 3/3)

Response to 'Questions For Authors':

How is the likelihood ratio computed for the flow-matching geometry, and which sampling steps are stochastic?

A7:

  • About Likelihood ratio. Please refer to A3 for a detailed explanation of the likelihood ratio computation for flow-matching geometry.
  • About Stochastic sampling. Flow matching generation uses 12 flow steps, with stochastic sampling applied within a randomly selected contiguous two-step window in the first half of the trajectory (noise scale 0.2). The remaining 10 steps use deterministic Euler integration.

We have further clarified these details in the revised manuscript. (Appendix. B.1, line 715)

How is SSR defined, which simulator and settings does it use, and how long does one reward evaluation take?

A8:

  • About SSR Definition. We define simulation success based on three criteria: (1) the generated sewing pattern can be successfully parsed and instantiated, with valid panel and stitch references; (2) all stitching constraints can be successfully established without triggering a StitchingError; and (3) all panels can be successfully triangulated without triggering a DegenerateTrianglesError.
  • About Simulator and Setting. We adopt the NVIDIA Warp XPBD backend provided by the official GCD pipeline for garment simulation. The simulation runs for up to 2,400 frames at 60 FPS (40 seconds of simulated time). The static threshold is set to 0.03, with a maximum non-static vertex percentage of 1.5%. The maximum numbers of body-cloth collisions and cloth self-collisions are set to 35 and 300, respectively.
  • About Reward Evaluation Time. A complete simulation-aware reward evaluation takes an average of 1.87 seconds per sample.

We have clarified SSR definition and simulation details in the revised manuscript. (Sec.5, line 335)

What are λ, ω and γ, and how were they chosen?

A9:

  • Parameter Definitions. The coefficients λ control the relative contributions of the Pattern-Format, Stitch-Aware, and Geometry-Aware rewards (Eq. 5), while ω balances the size, corner, and mesh-quality terms within the Geometry-Aware Reward (Eq. 15). The parameter γ is a shared scale multiplier used to calibrate the exponential distance rewards (Appendix C.4).
  • Selection of Reward weights λ and ω. Since exhaustively searching all weight combinations is computationally impractical, we empirically assign equal weights to both λ and ω to balance the reward components.
  • Selection of Calibrate Coefficient γ. Since the reward metrics have different units and numerical ranges, we calibrate each metric using the median of its positive distances, estimated from a validation subset. It controls how each metric occupies the range. With (γ=1), the median positive distance of every metric is mapped to a reward of 0.5, yielding comparable reward sensitivity across metrics with different units and numerical scales.

Which dataset and input setting do Table 2, Table 3 and Table 4 use? Why does the full model in Table 2 differ from Table 1?

A10:

  • All experiments in Table 2, 3, 4 are conducted under the image-conditioned setting on the GCD-MM dataset.
  • The difference arises from different experimental configurations. Specifically, the main experiments in Table 1 used 1,000 training steps and the full dataset, whereas the ablation experiments were conducted with 500 training steps using 50% of the dataset for training and evaluation.

We apologize for the confusion and have clarified these experimental settings in the revised manuscript to improve clarity. (Sec. 5.3, line 509)

Reviewer_ynLDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SimSew proposes a RL method to improve the simulation success rate of a supervised multimodal sewing-pattern generator, specifically SewFusion. Its reward combines reference-based panel and edge counts, stitch-pair consistency, seam-length compatibility, panel geometry, and lower-tail triangle quality.

The motivation is that reference reconstruction accuracy does not necessarily imply successful downstream garment construction or simulation. Experiments on SewFactory and GCD-MM cover several conditioning settings. The main image-conditioned GCD-MM result increases simulation success from 71.89% to 76.92%. Reward ablations and sampling studies provide additional evidence about the proposed formulation.

I find the problem worth studying, and the reported improvements in simulation success are encouraging. However, I am not yet convinced that the proposed reward is a sufficiently general measure of sewing validity.

Critical strengths

The paper addresses a failure mode that matters in practice: a predicted pattern may resemble the reference but still fail during stitching or triangulation. The reported increase in GCD-MM simulation success from 71.89% to 76.92% is therefore a useful result. It evaluates whether the generated patterns can actually pass through the downstream pipeline, rather than relying entirely on reconstruction accuracy.

The reward design also gives failed candidates more informative feedback than a binary success indicator would. Panel and edge counts, reference stitch-pair consistency, and panel geometry can distinguish candidates even when none successfully completes simulation. This is a sensible choice for GRPO, which depends on reward differences between samples generated for the same input.

The mesh-quality term is the most convincing part of the reward formulation. Using the fifth percentile of triangle quality focuses the reward on poorly shaped regions that an average could conceal. Table 2 provides some evidence that this matters: removing the mesh term lowers simulation success from 76.13% to 73.54%. Although this result needs uncertainty estimates, it supports including feedback from the constructed mesh.

However, I'd like to mention that the technical novelty is mainly in combining these rewards for sewing-pattern generation, and post-training a supervised generator with GRPO and geometric rewards already has precedents e.g. CAD-Coder. SimSew’s contribution would be more compelling if the authors showed that its particular reward choices remain effective beyond SewFusion (e.g. AIpperal, SewingLDM etc.) and the synthetic GarmentCodeData dataset (e.g. GarmageSet).

Critical weaknesses

My primary concern is that the reward design in SimSew may be tailored to several inherit assumptions within the synthetic GCD dataset. For example, the stitches may not be one-to-one on a multi-layer flared-dress. Previous work GarmageNet also discussed arbitrary edge subdivision and many-to-many, partial-edge sewing relationships in production patterns, and subdividing a curve can change the reward without changing the physical garment. Also, the simulation failures themselves could stem from limitations in the simulator.

Moreover, Section 5.1 initializes every proposed post-training experiment from SewFusion. Table 1 includes AIpparel and SewingLDM as competing generators, but does not apply the proposed post-training to either. And I'd like to encourage the author to specify the simulator, simulation horizon, acceptance thresholds, and treatment of failures.

Finally, the full model obtains 76.92% SSR in the main image-conditioned result in Table 1 but 76.13% in Table 2. It would be better if the author could document the configuration diffeences or evaluation subsets.

Questions for authors
  1. Does the proposed post-training improve another initialization, such as AIpparel? Which parts of the implementation depend specifically on SewFusion?

  2. Do the gains persist against matched-budget continued or filtered supervised training, candidate selection, and explicit validity repair?

  3. Are the trade-offs in Tables 1–2 stable across seeds and reward weights? What explains the increase in extreme seam-length imbalance, and how does it affect actual assembly?

  4. What exactly constitutes simulation success?

3 comments
AuthorsOpenReview ↗

Pointwise Response to Reviewer ynLD (Part 1/3)

We sincerely thank the reviewer for the constructive feedback and address each concern in detail below. Unless otherwise specified, the additional experiments on GCD-MM are conducted under the image setting, using 500 training steps and 50% of the dataset for both training and evaluation.

Response to 'Critical Weaknesses':

My primary concern is that the reward design in SimSew may be tailored to several inherit assumptions within the synthetic GCD dataset. For example, the stitches may not be one-to-one on a multi-layer flared-dress. Previous work GarmageNet also discussed arbitrary edge subdivision and many-to-many, partial-edge sewing relationships in production patterns, and subdividing a curve can change the reward without changing the physical garment.

A1:

Our reward design is not limited to GCD-specific assumptions. To further validate this, we conduct additional experiments on the GarmageSet dataset mentioned by the reviewer. It is noteworthy that:

  • To adapt to the point-to-point stitching representation in GarmageSet, we group consecutive matched point pairs into seam-level correspondences, making the stitch-aware reward largely insensitive to edge subdivisions and sampling density.
MethodPanel Acc. (%) ↑Panel-IoU (%) ↑Stitch-F1 (%) ↑SSR (%) ↑
GarmageNet (5000)15.3571.5258.7173.66
GarmageNet (6000)16.1372.6859.0673.94
GarmageNet (5000) + SimSew Rewards (1000)16.7973.7561.3476.31

These results demonstrate that our simulation-aware reward design can accommodate more flexible stitching representations, rather than relying exclusively on GCD-specific structural assumptions.

Also, the simulation failures themselves could stem from limitations in the simulator.

A2:

We agree with you. However, since it is relatively hard to modify the simulator itself, our main efforts lie in maximumly ensure the validity of sewing pattern, e.g., stitch references and panel triangulation, before sending it to simulator. Experiments show that after improving the validity of sewing pattern, simulation success rate can be improved.

Moreover, Section 5.1 initializes every proposed post-training experiment from SewFusion. Table 1 includes AIpparel and SewingLDM as competing generators, but does not apply the proposed post-training to either.

A3:

Thanks for your constructive suggestions. We supplement experiments as your request, applying our proposed post-training to AIpparel and SewingLDM with the same configuration as SimSew.

MethodPanel Acc. (%) ↑Edge Acc. (%) ↑Stitch Acc. (%) ↑Vertex L2 ↓SSR (%) ↑
SewingLDM76.2372.6571.945.940856.33
SewingLDM + SimSew Rewards77.9173.8473.975.579260.71
AIpparel86.3480.9573.495.460566.07
AIpparel + SimSew Rewards87.0582.7275.134.965770.12
SimSew (Ours)93.5287.2283.873.797676.13

The results demonstrate that our reward design consistently improve simulation success rates across different sewing pattern generators, supporting the generalizability of our framework beyond SewFusion.

We have also been incorporated into Table 1 of the revised manuscript. (Sec. 5.2, line 387)

And I'd like to encourage the author to specify the simulator, simulation horizon, acceptance thresholds, and treatment of failures.

A4:

  • Simulator. We adopt the NVIDIA Warp XPBD backend provided by the official GCD pipeline as our simulator.
  • Simulation Horizon. The simulation runs for up to 2,400 frames at 60 FPS (40 seconds of simulated time).
  • Acceptance Thresholds. The static threshold is set to 0.03, with a maximum non-static vertex percentage of 1.5%. The body-cloth collision and cloth self-collision limits are set to 35 and 300, respectively.
  • Treatment of Failures. If we understand your concern correctly, as far as we know, currently simulators are not able to correct failure cases. It would be appreciate if u could provide further details.

We have explicitly clarified the simulator and the definition of simulation success in the revised manuscript. (Sec. 5.1, line 335)

Finally, the full model obtains 76.92% SSR in the main image-conditioned result in Table 1 but 76.13% in Table 2. It would be better if the author could document the configuration differences or evaluation subsets.

A5:

Sorry for the missing configuration details. Specifically, the ablation experiments in Table 2 were conducted with 500 training steps using 50% of the dataset for training and evaluation, whereas the main experiments in Table 1 used 1,000 training steps and the full dataset. We have clarified these experimental settings in the revised manuscript to avoid confusion and improve reproducibility. (Sec. 5.3, line 509)

AuthorsOpenReview ↗

Pointwise Response to Reviewer ynLD (Part 2/3)

Response to 'Questions For Authors':

Does the proposed post-training improve another initialization, such as AIpparel?

A6:

Yes. We further applied our reward design to SewingLDM and AIpparel. Please refer to A3 for detailed experimental results.

Which parts of the implementation depend specifically on SewFusion?

A7:

  • Model-Agnostic Reward Design. Our main contribution, the simulation-aware reward design, is model-agnostic and can be adapted to different sewing pattern generators, as proved by A3.
  • Model-Dependent Policy Optimization. The policy ratios must be adapted to each generator's probabilistic formulation. For example, in our SewFusion-based implementation, discrete topology is optimized using autoregressive token-level policy ratios, while continuous flow-matching geometry is optimized using stochastic SDE transition-level policy ratios.

Do the gains persist against matched-budget continued or filtered supervised training, candidate selection, and explicit validity repair?

A8:

We conducted three additional controlled experiments under image setting on the GCD-MM datasets.

  • Matched-Budget Continued SFT. Starting from the same SewFusion checkpoint, we compared SimSew with continued SFT under identical GPU-hours.
  • Matched-Budget Filtered SFT. We further evaluated supervised fine-tuning on high-quality generated samples selected through simulation-aware filtering (best-of-8 selection), without GRPO-based policy optimization.
  • Candidate Selection. At inference time, we generate eight candidate sewing patterns for each input using the matched-budget continued SFT model and select the candidate with the highest reward score. Only ground-truth-free reward components are used for candidate selection to ensure a fair evaluation.
  • Explicit Validity Repair. If we understand correctly, explicit validity repair refers to rule-based post-processing of generated sewing patterns. Inspired by PatternGSL’s [1] deterministic validity handling, we merge only unstitched short edges with adjacent nearly collinear edges, then reindex all edges and update the stitch references using an old-to-new edge index mapping.
MethodPanel Acc. (%) ↑Edge Acc. (%) ↑Stitch Acc. (%) ↑Vertex L2 ↓SSR (%) ↑
SewFusion (SFT)92.3385.6481.744.125171.89
Matched-Budget Continued SFT92.6986.1282.214.065771.96
Matched-Budget Filtered SFT92.7686.1982.434.030871.91
Candidate Selection92.7586.3182.354.080572.20
Explicit Validity Repair92.4081.2479.074.890572.67
SimSew (Ours)93.5287.2283.873.797676.13

The results show that SimSew achieves the highest SSR, outperforming all compared baselines. These findings suggest that improving individual pattern-level metrics or applying rule-based repairs alone is insufficient to ensure simulation success, highlighting the effectiveness of our simulation-aware reward design and RL-based optimization. The experimental results have been incorporated into the revised manuscript. (Sec. 5.2, line 394)

Are the trade-offs in Tables 1–2 stable across seeds and reward weights?

Reference

[1] Zhenyang Li, Lutao Jiang, Yizhou Zhao, et al. PatternGSL: A Structured Specification Language for Template-Free and Simulation-Ready 3D Garments. ACM Transactions on Graphics (SIGGRAPH), 2026.

AuthorsOpenReview ↗

Pointwise Response to Reviewer ynLD (Part 3/3)

Response to 'Questions For Authors':

Do the gains persist against matched-budget continued or filtered supervised training, candidate selection, and explicit validity repair?

A9:

  • We conducted additional experiments using three random seeds while keeping the reward weights and all other training configurations unchanged.
MethodPanel Acc. (%) ↑Edge Acc. (%) ↑Stitch Acc. (%) ↑Vertex L2 ↓SSR (%) ↑
SewFusion94.0589.7485.183.568773.24
SimSew (Seed 1)93.8388.6084.303.177177.59
SimSew (Seed 2)94.2789.1584.703.794377.24
SimSew (Seed 3)93.9988.7784.443.585077.49
Mean+-Std94.03+-0.0588.84+-0.0884.48+-0.043.5188+-0.098577.44+-0.18

Across three random seeds, SimSew consistently improves SSR despite minor trade-offs in certain pattern-level metrics, demonstrating the robustness of our approach.

  • We further conducted a reward weights sensitivity analysis under the same random seed, systematically varying the weights of the Pattern-Format, Stitch-Aware, and Geometry-Aware rewards while keeping the remaining training configurations unchanged.
Methodλfmt\lambda_{\mathrm{fmt}}λstitch\lambda_{\mathrm{stitch}}λgeom\lambda_{\mathrm{geom}}Panel Acc. (%) ↑Edge Acc. (%) ↑Stitch Acc. (%) ↑Vertex L2 ↓SSR (%) ↑
Format-low0.100.450.4593.1687.1583.723.730876.03
Format-high0.600.200.2093.2187.1783.703.747275.97
Stitch-low0.450.100.4593.0087.1983.503.795776.09
Stitch-high0.200.600.2093.1786.9183.713.731975.95
Geometry-low0.450.450.1093.2087.1383.793.752976.05
Geometry-high0.200.200.6093.5987.0283.753.763776.01
Default1/31/31/393.5287.2283.873.797676.13

The topology and geometry metrics also exhibit relatively small variations, despite minor trade-offs. These results demonstrate that SimSew is robust to reward weight variations.

What explains the increase in extreme seam-length imbalance, and how does it affect actual assembly?

A10:

  • About Increase Seam-Length Imbalance. The increased seam-length imbalance (from 0.504 to 0.649) indicates closer to GT (0.713).
  • About Impact on Garment Assembly. In complex sewing patterns, seam-length imbalances are often intentionally introduced to create gathers, ruffles, or easing effects. Therefore, seam-length imbalance does not necessarily indicate assembly failure.

To improve clarity and avoid potential confusion, we have revised Figure 4 accordingly. (Sec. 4.2, line 369)

What exactly constitutes simulation success?

A11:

We define simulation success based on three criteria: (1) the generated sewing pattern can be successfully parsed and instantiated, with valid panel and stitch references; (2) all stitching constraints can be successfully established without triggering a StitchingError; and (3) all panels can be successfully triangulated without triggering a DegenerateTrianglesError. We have added this explicit definition of simulation success to the revised manuscript. (Sec.5, line 335)

Comments on the paper

1 comment

Dear Reviewers,

The discussion period ends on November 18 (AoE). If you have not yet replied to the author rebuttal, please do so as soon as possible, indicating which of your concerns are resolved and which remain.

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.