acceptodds
Under review as a conference paper at ICLR 2027

Inverse Garment Design with Visual Agents

Abstract

We study the problem of inverse garment design, reconstructing simulation-ready 3D outfits with arbitrary layers from a single image, with visual agents. Unlike differentiable optimization that is time-consuming, or feed-forward methods limited by training data, we introduce DressAgent, a hierarchical agentic system that recovers garment sewing patterns through planning and closed-loop refinement grounded in physical simulation. To this end, we first present an enhanced representation based on GarmentCode for diverse styles and arbitrary dressing layers, and build the first multi-layer sewing pattern dataset with over 15K outfits for supervision and benchmarking. Inverse design then proceeds coarse to fine: In layer planning and coarse estimation, agents infer layer order and relations, peel each layer into its own view, and estimate coarse patterns. In fine-grained refinement, an agentic trio of Tailor, Critic, and Orchestrator refines patterns in a closed loop of editing, simulating, and judging against the reference with visual and measured evidence. Carefully designed tools and an evolving memory make this refinement converge on the target geometry. We also design an agentic generation method for high-fidelity textures. DressAgent outperforms feed-forward methods by a large margin, surpasses or matches differentiable optimization in a fraction of the time, and generalizes to complex layering, diverse poses, and body shapes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D garment generation, sewing patterns, agentic systems

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_u176OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents DressAgent, a hierarchical agentic framework for recovering simulation-ready garments from a single in-the-wild image. The method first introduces GarmentCode-XL, which extends the original garment program representation with richer garment constructions, wider design ranges, and explicit multi-layer relations. DressAgent then decomposes an outfit into ordered layers, generates canonical and layer-specific views, and estimates an initial garment program for each layer. A Tailor-Critic-Orchestrator trio subsequently refines the programs through iterative cloth simulation and measured visual feedback. The framework additionally supports texture reconstruction and dynamic simulation, enabling the recovered garments to be reposed and simulated after reconstruction

Critical strengths
  • GarmentCode-XL broadens the garment program space with richer constructions and explicit multi-layer relations, enabling structured reconstruction of complex outfits.
  • The Tailor–Critic–Orchestrator design provides a well-structured closed-loop refinement process, with extensive ablations supporting its effectiveness and stability.
Critical weaknesses
  • The paper's main contributions largely center on GarmentCodeData-XL and an agentic pipeline heavily dependent on pretrained VLMs and existing tools, raising concerns about its methodological novelty for ICLR. Furthermore, while Tables 3, 11, and 14 support the benefits of refinement and role decomposition, the experiments do not sufficiently isolate the incremental benefit of adding GarmentCodeData-XL to decoder training, leaving the contribution of this additional training data unclear.
  • The claimed generalization to diverse real-world garments is not convincingly demonstrated. GarmentCodeData-XL is generated within the predefined GarmentCode-XL design space, inherently favoring the proposed representation. Although real-world evaluations are included, the paper lacks systematic quantitative and qualitative evaluations on structurally novel or difficult-to-represent garments, leaving its practical generalization capabilities and limitations unclear.
Questions for authors
  • What is the key methodological novelty of DressAgent beyond combining pretrained VLMs with garment-specific tools?
  • Could you provide comparisons with existing methods on per-layer sewing-pattern accuracy and final multi-layer geometry using GarmentCodeData-XL, beyond layer-count and relation accuracy?
  • How is structural consistency verified across the synthesized canonical front, generated back, and peeled views, and how do inconsistencies among these views affect reconstruction quality?
  • Could you provide quantitative and qualitative evaluations on unseen garment structures and garments beyond GarmentCode-XL's design space, including representative failure cases?
  • Since the representation ablation in Table 3 is partly evaluated on GarmentCodeData-XL, could you provide controlled comparisons on real-scan datasets, including refinement of existing methods (e.g., ChatGarment (Bianetal.,2025))?
1 comment

Thank you for your constructive feedback! We address your questions and concerns as follows:

  • (A) Methodological Novelty

    To our knowledge, DressAgent is the first work to harness coding agents for multi-layer sewing pattern reconstruction through garment-specific skills and tools that extend their capabilities to physical design. Our complete workflow enables (1) intelligent planning of layers and structural relations, (2) closed-loop refinement where agents autonomously coordinate subagents and select physical tools, and (3) agentic texture generation and extended garment simulation.

  • (B) Benefit of Decoder Training using GarmentCodeData-XL

    Our goal of Stage 1 is to write semantically aligned GarmentCode-XL programs from images. A straightforward approach is asking a VLM to read the source code of GarmentCode-XL and write programs directly from images, but its performance varies substantially across VLM backbones. To improve consistency and robustness of our framework, we introduce a more robust solution (Lines 211–255): predefine visual questions from the source code, let VLMs answer them from images, and train a decoder to map the answers back to executable parameters, which helps narrow the knowledge gap and improves consistency for our specific task.

    To demonstrate this, we compare this design with direct program writing with source code across different backbones on cases of Fig. 8, with Stage 1 results reported below:

    BackboneVariantPanel ↑Edge ↑Stitch ↑IoU ↑
    Qwen3.6-27BDirect writing37.535.837.70.519
    Qwen3.6-27BOurs42.544.543.50.631
    GPT-5.6 LunaDirect writing50.063.468.40.618
    GPT-5.6 LunaOurs52.564.472.80.623
    GPT-6 AstraDirect writing65.068.865.70.715
    GPT-6 AstraOurs70.072.976.50.728

    Compared with direct program writing, our decoder-based approach yields consistent gains for the closed-source backbones and more substantial gains for open-weight Qwen (+0.112 IoU), narrowing the performance gap and improving the capability of backbone models.

  • (C) Per-Layer Pattern Accuracy and Multi-Layer Geometry

    We extend Table 9 by reporting per-layer sewing-pattern accuracy and final multi-layer geometry on GarmentCodeData-XL, evaluated only on cases where each method predicts the correct number of layers.

    MethodCount Acc. ↑Per-Layer Garment IoU ↑Full CD ↓Full IoU ↑
    DressCode34.0%0.4213.150.662
    ISP50.0%0.6331.780.789
    GarmentRecovery49.5%0.6451.740.791
    Ours95.5%0.6941.430.840
  • (D) Structural Consistency across Front, Back, and Peeled Views

    During generation, we use strict prompts to instruct the image model to preserve the outfit and keep the human pose consistent across views (similar to [1]). We then use least-squares camera fitting to align each view with a shared 3D body, mitigating minor differences in position and scale. We measure our body-landmark alignment differences on 4D-Dress and report that all errors remain below 3% of body height (see table below). We’ll provide more details in the revised version.

    Error (% body height)Head topShouldersElbowsWristsHipsKneesFoot bottom
    Mean0.840.430.611.040.260.441.39
    Maximum2.870.981.642.530.731.502.65

    [1] Dress-1-to-3, SIGGRAPH 2025.

  • (E) Evaluations on Generalization and Representative Failure Cases

    Appendix Fig. 23 demonstrates broad coverage of everyday outfits, while our user study (Table 2) on DeepFashion, Pinterest, and user-uploaded photos, covering general garment structures beyond the synthetic training distribution, shows substantially higher preference than baselines. These results demonstrate applicability to everyday garments. We also discuss limitations in the appendix and show a failure case in Fig. 22 (a free-hanging sash, beyond our current scope).

  • (F) Representation Ablation on Real Scans, and Refining ChatGarment

    We report the representation ablation and the refinement of ChatGarment on the real scans of CloSe and 4D-Dress (same cases as Table 1) with a refinement budget of T = 6:

    VariantCloSe CD ↓CloSe IoU ↑4D-Dress CD ↓4D-Dress IoU ↑
    ChatGarment2.390.5801.940.763
    ChatGarment + our refinement1.380.7481.250.861
    Ours w/o GarmentCode-XL1.310.7521.240.863
    Ours Full1.200.7641.210.880

    We also provide visual comparisons on the ChatGarment dataset: https://imgur.com/a/JFvVlpR. ChatGarment generates standard-size garments on an SMPL body without layering, fine-grained component control, or texturing, while our refinement reduces these discrepancies and provides substantial improvements.

Reviewer_oCDJOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an agentic system to reconstruct layer-based 3D garments from a single image, grounded in physical simulation. The system leverages an enhanced garment representation based on GarmentCode, and takes a coarse to fine approach for its generation: layers and patterns are first estimated in the coarse-grained stage, and then refined in a closed-loop simulation and editing in the fine-grained stage. The system is demonstrated to outperform feed-forward methods and differentiable optimization methods.

Critical strengths

Novel idea and interesting system for using agents to solve inverse garment design. The paper is a timely contribution to the research area of agentic systems for computer visions.

The paper is very well written. The presentation is very clear and easy to follow, hinting a publication-ready manuscript.

The visual results of the paper are very nice, demonstrating the strong performance of the proposed system.

Critical weaknesses

Reproducibility: This is the weakest part of the paper. The system relies on GPT-5.6 Luna to design its agents, and GPT-Image-2 for image editing. Such closed-source system significantly hinders reproducibility.

Missing baseline: [A] Efficient 3D Garment Generation with Geometry Image Representation, CVPR 2026. This is a recent feedforward method that needs comparison.

Technical contribution: the proposed system glues an array of existing techniques. I can identify the system itself as a technical contribution. Template generation and program decoding could be interesting technical details as well.

Ablation: improvement from the enhanced representation, data from GarmentCodeData-XL or the agentic system should be demonstrated separately. The current ablation study (Table 3) is performed on GarmentCodeData-XL, making it difficult to understand how the ablated system performs compared to the baselines. The ablation could have been done on the original GarmentCodeData.

Layered design: the definition of layer is not clear or does not relfect layering in practice. For example, taking Fig 2b as an example. How many layers are there in the example? When the top is tucked into the skirt, it is counted as two layers, so the total layers in the example is 3? And when the top is not tucked into the skirt, it is counted as a single layer? To avoid confusion, please annotate each example with the number of layers.

For real-world examples, Fig 7 only shows results from single-layer garments. Based on the definition of layers in the real world, the paper seems not showing results of more than two layers. What is the maximum number of layers supported in the proposed dataset?

GarmentCodeData-XL: does the proposed dataset include all-new 15K outfits, or there are overlaps / reuse from data samples in GarmentCodeData? How many percents of the data is overlapped? How is the dataset size compared between GarmentCodeData-XL and GarmentCodeData?

AI use statement: the paper claims "We have not used them for any other research task", but this is clearly a system of agentic agents. It is important to mention the role of agents in the statement, as I am not convinced that this claim is accurate for now.

Questions for authors
  1. Please discuss how reproducibility for the entire system can be achieved. Could the authors confirm that even with source code release, all results in the paper can be reproduced exactly?

  2. Please run ablations on the original GarmentCodeData to justify how the proposed components improve over the baselines.

  3. Please clarify the layer definition, and the proposed dataset when positioned with original GarmentCode data.

--- Post-rebuttal recommendation ---

I raised my score given that the rebuttal addressed well my concern. Please make sure that the paper is reproducible upon publication.

1 comment

Thank you for your constructive feedback! We address your questions and concerns as follows:

  • (A) Reproducibility and Dependence on Closed-Source Models

    We include the source code in the supplementary material. To enable reproduction without proprietary APIs, we additionally evaluate an open-weight configuration of Table 1, using Qwen3.6-27B (thinking off) for agents and FLUX 3 Image Klein (seed = 0) for image tools.

    MethodPanel ↑Edge ↑Stitch ↑CloSe CD ↓CloSe IoU ↑4D-Dress CD ↓4D-Dress IoU ↑Avg. Time (min) ↓
    Ours (gpt-5.6-luna + GPT-Image-2)87.3676.5688.901.200.7641.210.8808.6
    Ours (Qwen3.6-27B + FLUX 3 Image Klein)77.8073.8379.701.350.7441.280.8764.9

    This open-weight configuration achieves competitive performance and shows advantages over baselines.

  • (B) Comparison with the SwiftTailor (CVPR 2026) Baseline

    As SwiftTailor’s code is not yet publicly released, we provide a qualitative comparison of DressAgent against it on the same input images from its paper. See: https://imgur.com/a/BI7TAAu.

    Additionally, DressAgent supports multi-layer reconstruction and agentic texture generation, which are not available in SwiftTailor.

  • (C) Component Ablations on Original GarmentCodeData

    We evaluate on 200 cases randomly selected from the GarmentCodeData test set. All refinement variants use T = 6. We report structure scores (%) and silhouette IoU using the evaluation protocols of Table 3:

    VariantPanel ↑Edge ↑Stitch ↑IoU ↑
    w/o Refinement65.158.659.10.612
    w/o Critic85.377.574.20.740
    w/o Measure Tools84.876.771.00.693
    w/o Rollback86.277.282.60.738
    w/o GarmentCode-XL87.878.385.10.771
    Full (T = 6)88.078.685.20.772

    These results further support the effectiveness of our complete system design beyond construction representation.

  • (D) Layer Definition, Annotations, and Maximum Supported Layers

    As stated in Lines 182–184, we empirically define layers from inner to outer garments, pairing one upper and one lower piece in each layer (with relations such as tucking between them) or including only a single upper/lower piece if the other type runs out. Fig. 2(b) therefore has two layers. Fig. 6 (top right) and Appendix Fig. 23 (bottom) show three-layer in-the-wild outfits, and our dataset covers up to four layers. We will clarify this in the revised version and annotate the number of layers as needed.

  • (E) Dataset Size and Relation with GarmentCodeData

    GarmentCodeData-XL contains 15,205 new outfits with no overlap with GarmentCodeData. Every outfit introduces structural or layering extensions unsupported by the original GarmentCode. Each example includes executable programs, layer structures and relations, sewing patterns, draped 3D meshes, and rendered views.

  • (F) AI Use Statement

    Thanks for pointing it out. We will revise this statement and explicitly disclose the roles of generative models within DressAgent in the revised version.

Reviewer_4NQBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses inverse garment design, recovering simulation-ready, possibly multi-layer sewing patterns from a single in-the-wild photo. It introduces GarmentCode-XL (an extended, layered GarmentCode with a matching simulator) and a 15K-outfit multi-layer dataset. Its agentic system, DressAgent, works in two stages: image-editing tools canonicalize the photo and peel off each layer, and a questionnaire-driven Visual Reader plus a learned decoder produce coarse programs. A Tailor/Critic/Orchestrator loop then refines them with simulation and measurement tools. A tool-based texture module completes the outfit. Validation includes structure accuracy on GarmentCodeData, Chamfer distance and IoU on CloSe and 4D-Dress, a user study and set of ablations. The paper claims large gains over feed-forward methods and parity with differentiable optimization at a fraction of the time.

Critical strengths
  1. Well engineered and controlled refinement. The Tailor/Critic/Orchestrator split, the Measure and Fit tools, and rollback are properly ablated.

    • Removing the Critic drops GarmentCodeData-XL IoU from 0.839 to 0.822.
    • Removing rollback drops structure accuracy from 73.59 to 58.70 and quadruples self-intersections.
    • The plain T-shirt stress test (Table 11) shows the loop can recover construction from a bad start.
    • The single-agent TextGrad comparison with a self-selection variant (Table 14) is a careful way to separate self-bias from the evidence the agent sees.
  2. Arbitrary layering is handled explicitly. The layered program, the dressing order, relations, and the simulator extension are mentioned. GarmentCodeData-XL (15K outfits, up to four layers) is a useful resource. Layer count accuracy of 95.5% and 96.7%, against at most 65% for ISP, GarmentRecovery and DressCode (Table 9), is a large and practically relevant gap.

Critical weaknesses
  1. The loop is fitted to generated images, and their faithfulness is only checked semantically. The image-generation tools do most of the seeing:

    • g_canonical re-poses the person.
    • g_back invents the back view.
    • g_peel invents the occluded inner layers.

    The Critic and the Measure tool then compare the drape against these generated views, in percent of body height. Table 10 validates the views with CLIP-I, DINO and ΔE, which are appearance metrics. They cannot tell whether a hem moved 5 cm or a sleeve got shorter during re-posing. So "precise fitting" may partly mean precise fitting to GPT-Image-2's interpretation of the outfit.

    The quantitative geometry benchmarks also use front renderings of scans. CloSe is close to A-pose already (Appendix G.1 says so), so the in-the-wild claim effectively rests on the user study. A geometric faithfulness check is needed, for example hem, sleeve-end and neck positions in the canonical view against the 4D-Dress scan rendered in the same pose.

  2. The evaluation leans towards the representation, and reproducibility depends on unreleased proprietary models.

    • Structure accuracy (Table 1) counts matching panels, edges and stitches on the GarmentCodeData test split. The decoder Ψ is trained on the GarmentCodeData-v2 training split, and the output space is GarmentCode. SewFormer, DressCode and GarmentGPT use different panel vocabularies, so 87.36 vs 18.67 largely measures agreement with the GarmentCode representation.
    • Table 3, "w/o GarmentCode-XL" is evaluated on a test set generated from GarmentCode-XL styles, so that drop is partly by construction. On DeepFashion2 the same variant (0.650) is better than "w/o Refinement" (0.562), which contradicts the text's claim that it "costs the most".
    • Proprietary models and variance. Every agent is GPT-5.6-luna and every image tool is GPT-Image-2. No run-to-run variance is given for Table 1, although the agents are stochastic. Fig. 8 shows the open Qwen3.6-27B backbone is much weaker, so it is unclear what an outside group could reproduce.
Questions for authors

Answers to the following queries could change my assessment.

  1. Is there any geometric check of the canonical, back and peeled views against ground truth, for example landmark positions as a percentage of body height on 4D-Dress?

  2. How many runs is Table 1 based on, and what is the spread? What are the CD units and the number of test cases for CloSe and 4D-Dress?

  3. How do the default budget and the Table 14 budget relate?

    • Fig. 8 says the loop converges after about 6 iterations.
    • Table 14 reports 0.846 for the trio at 10 simulations.
    • At the default 6 simulations, Fig. 16 gives 0.783 (4D-Dress) and 0.831 (CloSe).

    How does the trio at T = 6 compare with the single agent at 10 simulations (0.820)? The answer matters for the cost argument.

  4. How are panel, edge and stitch counts matched across methods with different pattern vocabularies?

  5. Report Table 1 with an open-weight backbone end to end (agents and image tools), so the community has a reproducible reference point.

4 comments

Thank you for your constructive feedback! We address your questions and concerns as follows:

  • (A) Geometric Validation across Views and Poses

    We evaluate garment landmarks across views and poses on 4D-Dress (same cases in Table 1), demonstrating only minor alignment differences. Peeling is not evaluated here because 4D-Dress lacks matched peeled references. Values are mean differences as percentages of body height:

    Evaluated outputViewTop hemBottom hemSleeve endNecklineWaist
    Draped A-pose vs canonical A-pose imageFront0.970.540.860.800.60
    Draped A-pose vs canonical A-pose imageBack1.091.791.090.661.29
    Reposed simulation vs dataset poseFront1.772.371.600.751.62
    Reposed simulation vs dataset poseBack1.923.151.760.722.35

    We also provide some visualizations here: https://imgur.com/a/R95BL8c.

  • (B) Details about Table 1

    In Table 1, we use the official GarmentCodeData-v2 test split (6,265 cases). Following [1], we use 600 scans from CloSe (100 subjects × 6 poses) and 128 cases from 4D-Dress (32 subjects × 2 outfits × 2 frames per sequence). CD is reported in centimeters after body-scale calibration (same as [1]). We’ll include the details in the revised version.

  • (C) Default Budget vs Table 14 Budget

    Table 14 uses T = 10 for both methods for fair comparison. We additionally compare it against our default budget T = 6 as follows. Our default setting achieves higher accuracy at lower cost, demonstrating it already suffices for general use cases.

    MethodIoU ↑Output tokens (k) ↓Refinement Cost ($/case) ↓VLM time (s/case) ↓
    Single agent + T=100.82038.70.145748
    Agentic trio + T=6 (ours)0.8229.20.096265
  • (D) Pattern Matching across Different Panel Vocabularies

    Following [2], we map each method's panels to the role-based panel classes of NeuralTailor [3] and pair panels of the same class. We then compare panel counts, the edge counts of paired panels, and stitches as (panel class, edge index) pairs. Methods with implicit sewing patterns (ISP, GarmentRecovery) have no explicit panels to pair, so we do not report structure accuracy for them.

  • (E) Reproducing Table 1 with Open-Weight Agents and Image Tools

    We replace the agents with Qwen3.6-27B (thinking off) and image tools with FLUX 3 Image Klein (seed = 0), keeping all other settings unchanged. It achieves competitive performance with lower runtime, still better qualities compared to the strongest baselines in Table 1 (e.g., GarmentRecovery).

    MethodPanel ↑Edge ↑Stitch ↑CloSe CD ↓CloSe IoU ↑4D-Dress CD ↓4D-Dress IoU ↑Avg. Time (min) ↓
    Ours (gpt-5.6-luna + GPT-Image-2)87.3676.5688.901.200.7641.210.8808.6
    Ours (Qwen3.6-27B + FLUX 3 Image Klein)77.8073.8379.701.350.7441.280.8764.9
  • (F) Addressing Representation Bias in Table 3

    To address the concern that evaluation on GarmentCodeData-XL favors our representation, we conduct ablation on real scans from CloSe and 4D-Dress using the protocol of Table 1.

    VariantCloSe CD ↓CloSe IoU ↑4D-Dress CD ↓4D-Dress IoU ↑
    Ours w/o GarmentCode-XL1.310.7521.240.863
    Ours Full1.200.7641.210.880

    GarmentCode-XL improves both metrics on both real-scan datasets. We will also correct the “costs the most” statement.

  • (G) Details Of Ethics Concerns

    Thank you for highlighting these. We obtained informed consent from user-study participants and will include a detailed ethics statement in the revised version.

Refs:

[1] Dress-1-to-3, SIGGRAPH 2025.

[2] GarmentGPT, ICLR 2026.

[3] NeuralTailor, SIGGRAPH 2022.

Thanks authors for addressing the concerns. Some of the doubts are cleared with results reported in the reply. However as we can see in all your replies (A-G). This require substantial additions to the main paper, main in terms of the new results, corrected statements or rephrasing etc. Therefore, I am keeping my initial rating.

Thank you for the follow-up. We have incorporated all additions into the revised PDF (blue in the PDF).

  • Sec. 3.1: layer definition.
  • Sec. 4.1: open-weight configuration.
  • Sec. 4.2: test-set sizes, CD units, and panel matching; open-weight row in Table 1
  • Sec. 4.3: revised "costs the most" statement.
  • AI use and ethics statements: roles of generative models; participant consent.
  • Appendix F.2: evaluation protocol and panel matching across vocabularies.
  • Appendix F.4: per-layer accuracy and multi-layer geometry.
  • Appendix F.5: open-weight configuration.
  • Appendix F.6: geometric faithfulness of synthesized views.
  • Appendix G.3: ablations on the original GarmentCodeData and on real scans.
  • Appendix G.4: questionnaire and decoder vs. direct program writing across backbones.
  • Appendix G.5: trio at T=6T = 6 vs. single agent at T=10T = 10.

Thanks to authors for pointing this out. Somehow these changes were not rendered in the pdf. I can see these changes are visible in the revised pdf now. I am keeping my score with an inclination to accept.

Comments on the paper

1 comment

We thank all reviewers and the AC for their constructive feedback. We are encouraged by the reviewers’ recognition of our idea as “novel” and “a timely contribution to agentic systems for computer vision” (oCDJ), our refinement process as “well engineered” and “properly ablated” (4NQB, u176), GarmentCodeData-XL as “a useful resource” (4NQB), and our paper as “very well written” (oCDJ).

As the discussion period draws to a close, we would like to take this opportunity to summarize and reiterate our main contributions:

(1) GarmentCode-XL and GarmentCodeData-XL, the first multi-layer sewing pattern representation and dataset (15K outfits with up to four layers).

(2) DressAgent, the first complete agentic system for inverse garment design, encompassing layer structure planning, a closed-loop multi-agent workflow grounded in physical simulation, and agentic texture extrapolation.

(3) Strong zero-shot generalization to in-the-wild photos of multi-layer outfits, with our results clearly preferred by users over those of all baselines.

We have addressed the reviewers’ main concerns in the revised PDF (with changes highlighted in blue):

  • Reproducibility: an open-weight configuration still outperforms all feed-forward baselines (Table 1, Appendix F.5).
  • Representation bias: every component remains effective on both the original GarmentCodeData and real scans (Appendix G.3).
  • View faithfulness: additional evaluation using garment landmarks on 4D-Dress (Appendix F.6).
  • Multi-layer accuracy: additional evaluation of per-layer and multi-layer geometry, with our method outperforming all baselines across all metrics (Appendix F.4).
  • Clarity: clarifications of the evaluation protocol and layer definition, as well as the AI use and ethics statements (Sec. 3.1, Sec. 4.2, Appendix F.2).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.