{"id":"3b07299d-1cd2-48a3-b893-15fc9040c9ba","arxiv_id":"2504.20409","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"An image-conditioned masked autoregressive model predicts GarmentCode-compatible garment parameters, then decodes them into valid, simulation-ready 3D garments without pattern self-intersections.","lead":"GarmentX turns a single photo of clothing into an editable, ready-to-simulate 3D garment by predicting high-level sewing pattern parameters instead of raw pattern geometry. The method pairs a masked autoregressive model with 378,682 synthetic garment image-parameter pairs, and reports faster, more wearable results than prior single-image garment generators.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'always valid' guarantee lacks a stated constraint on diffusion outputs, so decoded patterns may not stay in GarmentCode's validated parameter domain.","rationale":"Good-faith reading: the paper's main novelty is replacing unconstrained pattern prediction with a parametric representation that guarantees valid garments. This is attractive and partially supported: the GarmentX dataset is large, the autoregressive+diffusion design is sensible, the ablations (DiT comparison, CFG) support choices, and the zero failure on CLOTH3D is a positive signal. However, the guarantee is end-to-end only if every output of the network is mapped back into the valid parameter domain. The paper describes normalization in Section 3.1 but is silent on the inverse side for out-of-range and off-midpoint values. This is not an internal contradiction, but it is the least secure link in the central claim: everything downstream (valid patterns, simulation-ready garments, editability) depends on it. The reader's domain-gap concern is real but secondary; it affects real-photo alignment and is acknowledged in Section 5. My proposed check would settle the validity question with a small code-level experiment, and the paper should be accepted only after that step is documented.","tokens_in":13292,"tokens_out":8379,"duration_ms":92495,"concrete_test":"Check the inference path from raw diffusion outputs to N^{-1} in the released code (or, if unavailable, ask the authors for the exact post-processing). Run the model on 1,000 CLOTH3D validation images and record the raw normalized parameter vectors before any post-processing. Compute the fraction of continuous entries outside [-1,1] and, for discrete entries, the distance to the nearest category midpoint; then feed the decoded patterns to GarmentCode and count simulation failures both with and without clamping/rounding. If raw outputs are always in-domain and discrete outputs always land on midpoints, the concern is resolved; otherwise the paper must report the constraint step and its effect on the Table 1 numbers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 normalizes parameters with Eq. 1 and maps discrete attributes to category midpoints via Eq. 2, and Section 3.3 credits 'strictly constrained parameter boundaries' for the zero-failure result. The paper, however, never specifies how raw diffusion outputs (Eq. 3) are constrained: no clamping to [-1,1] and no rounding/quantization to nearest discrete midpoint is described before the inverse transform N^{-1}. A denoised sample from a diffusion model has no built-in guarantee of staying inside the training support, so predicted parameters can fall outside [p_min,p_max] or between category midpoints. Without an explicit constraint, GarmentCode is being asked to interpret values it was not designed for, and the asserted guarantee that decoded patterns 'always form valid, simulation-ready 3D garments' is unsupported. If such a constraint exists, it is a central part of the method and must be stated and evaluated; if it does not, the zero-failure claim is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"GarmentX proposes a single-image 3D garment generation framework that predicts a structured parametric representation compatible with GarmentCode, rather than directly predicting sewing patterns. A masked autoregressive model with a diffusion loss generates the parameters from DINOv2 image tokens; the parameters are then decoded into sewing patterns and simulated into 3D garments. The paper also introduces a 378,682-pair synthetic dataset built by rendering GarmentCodeData garments and applying canny-conditioned ControlNet, and reports state-of-the-art Chamfer distance (4.92) and point-to-surface distance (4.96) on CLOTH3D with zero simulation failures, along with ablations over the autoregressive backbone, CFG scale, and data construction choices.","tokens_in":13355,"tokens_out":6204,"duration_ms":59257,"significance":"The underlying idea is promising: constraining generation to a validated parametric space could eliminate the pattern failures of direct edge/stitch predictors while retaining editability. If the validity guarantee and the quantitative results hold, the 60x runtime improvement over deformation-based baselines and the large dataset would be valuable community contributions. The paper also provides substantial architectural detail and includes ablations over model choice, CFG scale, and data construction. However, the key guarantee is not yet substantiated because the post-processing of diffusion outputs is unspecified, and the evaluation lacks the statistical and protocol details needed to judge the state-of-the-art claim. The zero-failure and runtime claims are concrete and testable, and evaluation on an external benchmark (CLOTH3D) is a strength; the work is significant in potential but requires a careful revision.","major_comments":[{"comment":"The central guarantee that decoded sewing patterns 'always form valid, simulation-ready 3D garments' is not supported by the inference procedure as described. The reverse diffusion process produces continuous denoised samples, but the manuscript never specifies how these samples are constrained before the inverse transform N^{-1}: no clamping to [-1,1] for continuous parameters and no rounding/quantization to the nearest discrete midpoint introduced in Eq. (2) is reported. Without an explicit constraint, denoised outputs can lie outside the normalized range or between category midpoints, and GarmentCode would receive values outside the validated domain. Please state the exact post-processing (e.g., clamp plus nearest-midpoint quantization), evaluate its impact, or revise the 'always valid' claim.","section":"§3.3, Eq. (3)"},{"comment":"The conversion from GarmentCodeData's low-level attributes to the GarmentX representation is described only as a transformation; the paper does not provide the attribute inventory, the mapping rules from panel/stitch data to the parameter vector p, the determination of p_min/p_max for Eq. (1), or a validation procedure that filters parameter vectors that GarmentCode cannot decode. This underspecification is load-bearing because the dataset construction and the claimed guarantee that every sampled parameter vector yields a valid garment rest on it. Please provide the mapping details or an explicit reference to the GarmentCode API and any filtering steps.","section":"§3.2"},{"comment":"The quantitative evaluation lacks essential context: the CLOTH3D test set size is not reported, no variance or error bars are given, and the protocol for producing input images from CLOTH3D (viewpoint, resolution, renderer, and whether images are original or newly rendered) is not documented. DressCode is mentioned as a baseline in §4.2 but does not appear in Table 1, and the 'Wearable' column is undefined. Please add these details and state how the input-image protocol affects the generalization-to-real-photos claim.","section":"§4.1, Table 1"},{"comment":"The data construction pipeline generates ControlNet-enhanced images from the same parametric representation that serves as prediction targets, so any geometric distortion introduced by the realism-enhancement step (e.g., altered folds, hemlines, collars) can bias the learned image-to-parameter mapping. The paper provides no check that ControlNet images are geometrically faithful to the original renders. A quantitative comparison of parameters predicted from ControlNet images versus the original renders would substantiate the claim that these synthetic pairs are faithful proxies for real photos; without this, the input-image alignment claim on real images is only weakly supported.","section":"§3.2, Fig. 2"}],"minor_comments":[{"comment":"Typos 'realistc' and 'compatability' should be corrected.","section":"§3.2"},{"comment":"The notation omega_s = s * lambda_cfg / S is ambiguous; clarify that s is the auto-regressive step index and S is the total number of steps.","section":"§3.3"},{"comment":"The caption contains the typo 'compartion' and should read 'comparison'.","section":"Table 3"},{"comment":"The relationship between the 32 auto-regressive steps and the token count N is unclear; please clarify how the cosine schedule in Eq. (4) is applied when N differs from 32.","section":"Supplementary §6"},{"comment":"DressCode is compared only qualitatively; either report its quantitative numbers in Table 1 or justify its exclusion from the quantitative comparison.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"Dear Editor, the manuscript addresses an important problem and the core idea is sound, but the 'always valid' guarantee needs to be pinned down and the evaluation needs more statistical substance. I would be willing to look at a revised version that adds the boundary-constraint specification, the GarmentCodeData-to-GarmentX mapping details, and proper evaluation statistics. I also suggest the authors clarify how their approach differs from DressCode and AIpparel, since both are cited but not included in the quantitative comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the paper is a genuine step toward practical image-to-garment generation, and the new dataset is the kind of asset the field needs. But the central guarantee—that decoded patterns always form valid, simulation-ready garments—is not actually supported as written, because the paper never specifies how diffusion outputs are kept inside the valid parameter domain.\n\nWhat's new: the recipe of predicting GarmentCode parameters with a masked autoregressive transformer plus diffusion loss is a real combination, and the 378k parameter-image dataset created via ControlNet from rendered meshes is a useful resource. The ablations are informative: the autoregressor beats a DiT baseline with fewer parameters, the CFG sweep is sensible, and the data-construction comparison is honest. Evaluation on CLOTH3D, which was not used for training, gives the numbers some independence. The paper also includes failure cases and a clear limitation statement about human-free inputs.\n\nWhere it's soft: the stress-test concern is real. Section 3.1 defines normalization, and Section 3.3 credits 'strictly constrained parameter boundaries' for zero failures. But the inference text never describes clamping to [-1,1] or rounding discrete attributes to the nearest midpoint after denoising. A diffusion sample can land anywhere. If such a constraint exists, it belongs in the method section; if not, the 'always' guarantee is overstated. This is the load-bearing flaw in the wearability claim.\n\nBeyond that, the usual arXiv-competition gaps: no error bars or test-set sizes in Tables 1–3, no code or weights release yet, no numeric comparison with AIpparel, and no quantitative evaluation on real photos. Tuning CFG on the same CLOTH3D benchmark used for the headline result is a minor overfitting risk. The dataset construction's reliance on ControlNet for realism is acknowledged, but the domain-gap question remains unanswered.\n\nWho it's for: anyone working on garment generation, virtual try-on, or asset pipelines will want to read this and probably use the dataset once it's out. It deserves a serious referee. My recommendation: send it to review, with the explicit request that the authors either document the parameter-domain constraint or soften the guarantee, and release the dataset/code for reproducibility.","headline":"A useful new dataset and a sensible parametric recipe, but the 'always valid' guarantee is under-specified and needs a fix before publication.","tokens_in":14012,"tokens_out":2671,"would_cite":true,"duration_ms":24834,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Predicting garment parameters, not pattern edges, yields wearable 3D garments from a single image.","keywords":["3D garment generation","sewing pattern prediction","masked autoregressive model","parametric garment representation","image-to-garment","diffusion model","garment dataset","cloth simulation"],"falsifier":"Take a set of real garment photos with known ground-truth sewing patterns (or high-quality scans of the same garments), run GarmentX on them, and compare predicted parameters and simulated meshes against ground truth; if Chamfer or point-to-surface errors on real photos are much larger than on synthetic pairs, or if simulation failures reappear, the claimed validity guarantee and input-image alignment fail on real input. A cheaper check: uniformly sample random parameter vectors and verify that every one decodes to a valid pattern; any decoded self-intersection would refute the 'always valid' claim.","tokens_in":12984,"feed_emoji":"👗","tokens_out":5669,"duration_ms":53085,"temperature":0.7,"pith_summary":"GarmentX claims that image-to-garment methods fail because they predict low-level sewing-pattern geometry, an unconstrained space full of invalid configurations. The paper instead predicts a structured vector of high-level, editable garment parameters — neckline type, sleeve shape, length, flare — that is guaranteed to decode through GarmentCode into a valid, simulation-ready sewing pattern. On the CLOTH3D benchmark the method reports the best Chamfer distance (4.92) and point-to-surface distance (4.96), zero simulation failures, and 15-second inference, and it supports direct editing by changing parameters. A dataset of 378,682 parameter-image pairs, with photorealistic images synthesized from rendered garments, provides the training substrate. If these results hold, a single image becomes a practical entry point to production-grade editable digital clothing.","feed_headline":"One image to a wearable 3D garment, with zero simulation failures","feed_subtitle":"A masked autoregressive model predicts editable garment parameters, beating pattern-based methods.","key_machinery":"The load-bearing object is the GarmentX parameter representation: a fixed-length vector of semantically meaningful garment attributes, normalized to $[-1,1]$ (continuous values linearly rescaled; $K$ discrete choices mapped to midpoints of $K$ equal intervals). Because this vector is a valid GarmentCode program, the difficult problem of predicting panel edges, stitches, and connectivity is removed from the network output. The predictor is a masked autoregressive model: DINOv2 extracts image tokens, an MAE-style encoder-decoder fuses them with partially masked parameter tokens, and a diffusion MLP denoises each token to its ground-truth parameter value; tokens are generated in a cosine-scheduled order and classifier-free guidance is applied. The data-construction pipeline — rendering multi-view images in Blender and photorealism-enhancing them with canny-conditioned ControlNet — supplies the paired supervision that makes the image-to-parameter map learnable.","core_discovery":"The paper's central discovery is that constraining generation to a parametric garment space converts an ill-posed pattern-assembly problem into a tractable sequence-prediction problem. Each garment is represented by a normalized parameter vector, with continuous attributes linearly rescaled to $[-1,1]$ and discrete choices mapped to interval midpoints; because every vector in this space corresponds to a valid GarmentCode program, the decoded 2D sewing patterns always stitch up into open, single-layer, wearable 3D garments. A masked autoregressive transformer, conditioned on DINOv2 image tokens and using a diffusion denoiser for the continuous parameter values, predicts the vector; a cosine generation schedule and classifier-free guidance balance quality and condition adherence. Trained on the GarmentX dataset and evaluated on CLOTH3D, the model achieves Chamfer distance 4.92 and point-to-surface distance 4.96 with no simulation failures, outperforming pattern-based, deformation-based, and general 3D-generation baselines, and runs in 15 seconds.","pith_inferences":["The dataset's reliance on canny-conditioned ControlNet as a realism bridge means the method's real-world ceiling is set by how faithfully that model preserves geometric cues; a testable extension is training on a small set of real photos with paired parameters to measure and close the gap.","A natural next step is text-conditioned generation: replacing the image-token condition with text embeddings would let the same parametric autoregressor produce garments from descriptions while retaining the validity guarantee.","The recipe of restricting the output space to a closed parametric manifold may transfer to other structured assets, such as furniture, buildings, or mechanical parts, where decoded outputs must satisfy hard physical constraints.","If the parameter space is truly closed under GarmentCode decoding, then random edits within normalized bounds should always stay valid; a simple stress test would verify this central promise."],"forward_implications":["Direct pattern predictors like SewFormer fail in 76.3% of simulations; GarmentX-style parameter prediction brings the failure rate to zero on the benchmark, making simulation robustness a property of the representation rather than of network luck.","Generated garments are open-structure and single-layer, so they can be draped onto arbitrary human bodies and animated, unlike closed or double-layered meshes produced by general image-to-3D models.","Because the output is a parameter vector, users can edit sleeves, collars, hoods, and even garment category after generation; editing is geometry-aware rather than a mesh deformation.","Autoregressive token generation with diffusion denoising reaches 15-second inference, about 60 times faster than deformation-based Garment3DGen, making interactive use plausible.","The model generalizes beyond its training distribution to sketches and uncommon garment types such as asymmetric one-shoulder dresses."],"supporting_citations":[{"why":"GarmentCode decodes the predicted parameter vector into valid, simulation-ready sewing patterns; the entire validity guarantee rests on this.","marker":"[29]"},{"why":"GarmentCodeData supplies the raw parametric garments whose attributes are converted into the GarmentX representation, grounding the dataset.","marker":"[30]"},{"why":"Canny-conditioned ControlNet turns rendered garment images into photorealistic training pairs, bridging the synthetic-to-real gap.","marker":"[66]"},{"why":"MaskGIT supplies the masked autoregressive generation scheme and cosine token schedule used by the predictor.","marker":"[6]"},{"why":"DINOv2 extracts the image condition tokens that guide parameter prediction.","marker":"[44]"},{"why":"The MAE encoder-decoder architecture fuses image tokens with masked parameter tokens for conditional generation.","marker":"[21]"},{"why":"CLOTH3D is the evaluation benchmark whose Chamfer distance and point-to-surface metrics quantify the reported improvements.","marker":"[4]"},{"why":"SewFormer is the direct pattern-prediction baseline whose 76.3% failure rate motivates the parametric approach.","marker":"[36]"},{"why":"Garment3DGen is the deformation-based baseline compared for fidelity and runtime.","marker":"[48]"}],"fun_headline_variants":["One image to simulation-ready 3D garment, zero failures","Parametric autoregression beats pattern-based garment generation","GarmentX turns single photos into editable, simulation-ready 3D garments","Autoregressive model ensures valid 3D garments from any input image","Predict parameters, not seams: autoregressive 3D garment generation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that canny-conditioned ControlNet images made from rendered 3D garments preserve the geometric cues (folds, hemlines, collars, proportions) that the model needs, so the image-to-parameter mapping learned on synthetic pairs transfers to real photographs.","fun_headline_variants_meta":{"raw":{"variants":["One image to simulation-ready 3D garment, zero failures","Parametric autoregression beats pattern-based garment generation","GarmentX turns single photos into editable, simulation-ready 3D garments","Autoregressive model ensures valid 3D garments from any input image","Predict parameters, not seams: autoregressive 3D garment generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001304,"raw_usage":{"total_tokens":5323,"prompt_tokens":957,"completion_tokens":4366,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":4277}},"tokens_in":573,"tokens_out":4366,"duration_ms":26000,"temperature":1.0,"reasoning_tokens":4277,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:29:31.781062+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of real garment photos with known ground-truth sewing patterns (or high-quality scans of the same garments), run GarmentX on them, and compare predicted parameters and simulated meshes against ground truth; if Chamfer or point-to-surface errors on real photos are much larger than on synthetic pairs, or if simulation failures reappear, the claimed validity guarantee and input-image alignment fail on real input. A cheaper check: uniformly sample random parameter vectors and verify that every one decodes to a valid pattern; any decoded self-intersection would refute the 'always valid' claim.","supporting_citations":[{"cited_title":"Garment- Code: Programming Parametric Sewing Patterns","cited_arxiv_id":null,"evidence_quote":"GarmentCode decodes the predicted parameter vector into valid, simulation-ready sewing patterns; the entire validity guarantee rests on this."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MaskGIT supplies the masked autoregressive generation scheme and cosine token schedule used by the predictor."},{"cited_title":"Masked Autoencoders are Scal- able Vision Learners","cited_arxiv_id":null,"evidence_quote":"The MAE encoder-decoder architecture fuses image tokens with masked parameter tokens for conditional generation."},{"cited_title":"CLOTH3D: Clothed 3D Humans","cited_arxiv_id":null,"evidence_quote":"CLOTH3D is the evaluation benchmark whose Chamfer distance and point-to-surface metrics quantify the reported improvements."},{"cited_title":"Towards Garment Sewing Pattern Recon- struction from a Single Image","cited_arxiv_id":null,"evidence_quote":"SewFormer is the direct pattern-prediction baseline whose 76.3% failure rate motivates the parametric approach."}],"review_version":1}