{"id":"1fe25dad-3291-4a8f-a618-21df972dee16","arxiv_id":"2412.06201","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A mask deformation network conditioned on physical size parameters performs size-variable virtual try-on, supported by a new paired-image dataset and a size evaluation metric.","lead":"This paper introduces size-variable virtual try-on, where a neural network changes the size of clothes in a photo to match a user-specified physical size. It also contributes a new dataset of paired photos and a new metric for measuring size changes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The mask deformation network never receives the try-on clothes image CT, so the claimed size-variable virtual try-on is unverified for target garments whose silhouette differs from the reference garment.","rationale":"The paper presents a plausible size-conditioned mask deformation network, and the reported results do show size changes in the paired same-garment setting. However, the strongest assertion is broader than what the architecture can support. The MDN is explicitly a function of PR, MR, SR, and ST only; CT appears only in the texture-warping stage (TPS/CFN). This means MD is a deformation of the reference person's current garment mask, not a mask estimated from the target garment's shape. The training and test protocol never ask the model to predict a mask for a garment type different from the one in PR, so the central claim 'fitting any size of clothes to a reference person' is not established for arbitrary try-on clothes. The reader's weakest assumption concerns the availability of SR at inference; that is a legitimate deployment concern but not the deepest gap. A model with SR available still cannot handle a target garment whose silhouette differs from the reference garment. The appropriate verdict remains CONDITIONAL: the paper should either explicitly scope its claim to same-garment size changes or extend the method to incorporate CT (e.g., a standard garment-agnostic mask from an existing parser) and evaluate cross-garment performance. No rejection is warranted because within its stated dataset the method demonstrably produces size-variable masks and the reported SEM/LPIPS/FID improvements are internally consistent.","tokens_in":11006,"tokens_out":10312,"duration_ms":115211,"concrete_test":"Run the full pipeline twice with the same PR, SR, ST but two different CT images of different garment categories (e.g., a t-shirt and a coat) that share the same size parameters. Because MDN ignores CT, MD will be bitwise identical. If the resulting PT images show the coat squeezed into the t-shirt-shaped mask or otherwise fail to reproduce the coat silhouette, the method is confined to same-garment resizing. Alternatively, build a small cross-garment evaluation set with the same subject and pose, where PR wears garment A and CT is a different garment B of the intended size, and measure SEM and LPIPS against same-garment pairs; a large degradation would confirm the scope limitation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 (Fig. 6) and the Inference paragraph specify that MDN takes only PR, MR, SR, and ST; the try-on clothes image CT is not used for mask estimation. Consequently the predicted deformed mask MD is independent of the target garment's shape. The dataset (Sec. 3.1) is built from paired images of the same garment type in different sizes, and all experiments, including the 24 new-clothes test items, are of this same-garment form. The central claim in the abstract ('fitting any size of clothes to a reference person') and in Sec. 4.3 ('can deform the mask according to the physical size of the clothes') is therefore supported only for the narrow case in which the reference person already wears the same garment as the try-on clothes. For a different garment type (e.g., PR shows a t-shirt and CT is a coat), the model has no mechanism to infer the coat's silhouette; it can only resize the t-shirt-like reference mask. This is a scope gap in the central claim, not just a deployment detail. The SR-availability issue raised by the reader is real but secondary: even with SR known, the absent CT input limits what the method can claim about general size-variable virtual try-on.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper defines size-variable virtual try-on as a new problem in which the garment worn by a reference person is resized in the output image according to the physical size parameters of the target garment. The proposed pipeline first estimates a deformed clothes mask MD from the reference person image PR, the reference segmentation mask MR, and the 5D physical-size vectors of the reference and target garments (SR and ST); afterward, the existing ACGPN components TPS and CFN warp the target clothes image CT into the deformed mask and fuse it with the person. The authors collect a paired dataset of 26 subjects wearing 14 garments in different sizes, using posture matching and projective alignment, and propose a Size Evaluation Metric (SEM) based on hem and sleeve area differences. Experiments compare the method with ACGPN and StyleGAN and include ablations of the residual mask path, the mask refiner, the reference image input, and several loss choices.","tokens_in":11233,"tokens_out":6647,"duration_ms":69522,"significance":"The problem setup is timely, and the dataset is a potentially useful asset for a largely unexplored task. The idea of learning a residual between the reference mask and the target-size mask, conditioned on physical size vectors, is clean and pragmatic, and the modular use of existing warping and fusion components is sensible. The authors also explicitly acknowledge the dataset's limited scale in the conclusion. However, the current evidence does not support the broad claim of fitting 'any size of clothes to a reference person': the mask estimator never sees the target garment image, the dataset and test set contain only same-garment pairs, and the quantitative evaluation relies on a self-proposed metric with no error bars and no public code or data release. These limitations are within the authors' control and should be addressed before publication.","major_comments":[{"comment":"The Mask Deformation Network takes PR, MR, SR, and ST as inputs and does not receive the try-on clothes image CT. The predicted deformed mask MD is therefore independent of the silhouette of the target garment. The dataset in Sec. 3.1 contains only pairs of the same garment type in different sizes, and the 24 new-clothes test items are also of this same-garment form. This supports a claim about size-conditioned deformation of the reference garment mask, but not the abstract's 'fitting any size of clothes to a reference person' or Sec. 4.3's 'can deform the mask according to the physical size of the clothes' when the target garment is a different type from the reference garment. The authors should either re-scope the claims to the same-garment setting or modify the architecture to condition on CT and evaluate on cross-garment pairs.","section":"Sec. 3.2, Fig. 6, and the Inference paragraph"},{"comment":"All quantitative results are single point estimates computed on 96 test pairs, with no error bars, confidence intervals, or significance tests. FID is known to be unstable at such small sample sizes, and the paper does not state how many samples were used for each FID computation. LPIPS is essentially identical (0.44) across nearly all methods and ablations, so the reported differences are not demonstrably meaningful. The authors should report multiple runs or seeds, per-pair distributions, and significance tests, and they should disclose the exact sample size for each metric.","section":"Sec. 4.2 and Tables 2-5"},{"comment":"SEM is proposed by the same authors and directly measures area differences in the hem and sleeve regions between the estimated mask MD and the ground-truth mask MG. MG is generated by the same Graphonomy segmentation estimator used to produce MR, and it is the exact training target of the mask losses LW, LD, and LA in Eq. (1). SEM therefore tracks the training objective, and no evidence is provided that SEM correlates with human judgments of correct garment size. To support the central quantitative claim, the authors should validate SEM against human ratings or an independent segmentation/reference procedure, and they should report SEM for baseline methods with the same post-processing.","section":"Sec. 3.3, Eqs. (2)-(4)"},{"comment":"The method requires the physical size SR of the clothes in the reference person image to be provided at inference time, but no estimator for SR from an arbitrary photograph is proposed. In the dataset this quantity is known by construction, so the experiments do not address the deployment scenario implied by the title. If SR is intended as user-supplied metadata, this should be stated explicitly; otherwise, a size-estimation module or a robustness experiment with noisy SR values is needed.","section":"Sec. 3.2, Inference"}],"minor_comments":[{"comment":"The word 'closes' should be 'clothes'.","section":"Abstract"},{"comment":"'ReLe activation' should be 'ReLU activation', and 'The encoder and decode consist' should be 'the encoder and decoder consist'.","section":"Sec. 4.1"},{"comment":"The sentence 'This gap can be suppressed by rescaled/normalizing an input image' mixes participle and gerund forms; it should read 'by rescaling/normalizing an input image'.","section":"Sec. 3.1"},{"comment":"References [3] and [4] appear to be duplicates of the same 'Multi-garment net' paper; the duplicate should be removed.","section":"References"},{"comment":"The paper does not mention any plan to release code or data, which weakens reproducibility; even a commitment to release the dataset would strengthen the contribution.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you work in virtual try-on. The paper defines a new task, builds a paired dataset with physical size labels, and proposes a residual mask deformation network plus a size-specific metric. That is real work, and the dataset collection with posture matching is a solid contribution on its own.\n\nThe stress-test point is the one that matters most. The mask deformation network never receives the try-on clothes image CT. It takes only the reference person, the reference mask, and the two size vectors, so it can only resize the reference garment's silhouette. All experiments use the same garment type in different sizes. The abstract's claim that the method fits any size of clothes to a reference person is therefore only supported for resizing the same garment. That is a scope gap in the central claim, not a deployment detail. The SR-availability issue the reader flagged is secondary but real: at inference you need to know the reference person's clothing size, and the paper only gestures at a rescaling heuristic.\n\nThe evaluation soft spots are proportionate but real. The test set is 96 pairs, there are no error bars, and SEM is self-proposed, with ground-truth masks generated by the same segmentation estimator used in the pipeline, so the metric risks circularity. The comparison against ACGPN is acknowledged as unfair. The ablations are thorough and the loss analysis is honest, which is to the authors' credit.\n\nWhat the paper does well: the residual formulation makes sense, the failure of the straightforward size-conditioned warping is plausible, and the qualitative results do show size changes. LPIPS and FID are comparable to the no-size variant, which suggests the size deformation does not hurt image quality.\n\nFor whom: researchers in image-based virtual try-on, especially those looking for a benchmark or dataset for size-conditioned generation. It deserves a serious referee because the problem and dataset are new, but the authors should either narrow the claim to same-garment size resizing or extend the architecture to take CT into account, and they should release at least part of the data. Send it to review, with the expectation of major revision.","headline":"A genuinely new task and dataset, but the central claim of size-variable virtual try-on is only demonstrated for the same garment type, and the evaluation is too thin to verify the rest.","tokens_in":11764,"tokens_out":1510,"would_cite":true,"duration_ms":17174,"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":"A mask-deformation network turns physical size labels into visible garment-size changes in virtual try-on images.","keywords":["virtual try-on","size-variable garment fitting","mask deformation network","physical clothes size","size evaluation metric","try-on dataset","image warping","human parsing"],"falsifier":"Fix a reference image with a known reference size and generate deformed masks for target sizes S, M, L, and XL; if the hem and sleeve areas of the deformed mask do not increase monotonically with the labeled target size, or if swapping the reference and target sizes does not reverse the size ordering, the network is not actually conditioning on physical size. A concrete test is to train only on certain garment types and then evaluate on a never-seen garment while holding the person fixed: if the Size Evaluation Metric on that held-out garment is close to chance, the size control does not generalize to new garments.","tokens_in":10767,"feed_emoji":"👕","tokens_out":6393,"duration_ms":61456,"temperature":0.7,"pith_summary":"The paper introduces size-variable virtual try-on, a problem setting in which the garment in a synthetic try-on image changes its on-body size according to the physical size labels of the clothes, not just the person's body shape. The authors claim that a size-variable mask deformation network can produce a clothing mask whose hem and sleeve regions match the garment's physical size when conditioned on the five-dimensional size parameters of both the reference garment and the target garment. To train and evaluate this, they collect a paired-image dataset of 1,524 images from 26 subjects wearing the same garments in different sizes, and they propose a Size Evaluation Metric that measures hem and sleeve area differences. Their experiments report that this approach beats general virtual try-on baselines on the size metric while staying comparable on perceptual quality. If this holds, users could specify a garment size in advance and see a realistic preview of how that size would fit.","feed_headline":"Virtual try-on now resizes garments to their physical labels","feed_subtitle":"A mask-deformation network takes reference and target clothes sizes and extends hems and sleeves accordingly.","key_machinery":"The load-bearing component is the size-variable Mask Deformation Network, built from a Residual Mask Deformation Network and a Mask Refiner. The residual encoder-decoder encodes the reference person image and its segmentation mask into a mask feature map, while a Size Feature Extractor turns the reference and target size vectors into a size feature map; the two are fused by elementwise multiplication. The decoder estimates the residual mask between the reference mask and the target mask, and that residual is added to the reference mask before a two-convolution-layer refiner produces the final deformed mask. By modelling the residual rather than the full mask, the network can concentrate on the small hem and sleeve displacements that encode garment size. The paired dataset and the Size Evaluation Metric are auxiliary machinery that make training and evaluation of this size-conditioned deformation possible.","core_discovery":"The central claim is that the physical size of a garment can be propagated through an image-based virtual try-on pipeline by learning to deform the clothing mask, rather than by warping pixels directly or by adjusting generative style vectors. The proposed mask deformation network takes the reference person image, the reference garment mask, and the size vectors for the reference and target garments, fuses the mask features with the size features by elementwise multiplication, and predicts the residual between the reference mask and the ground-truth try-on mask. Adding that residual to the reference mask and passing it through a mask refiner yields the deformed mask used for subsequent Thin Plate Spline warping and content fusion. The paper reports that this residual formulation is what lets the network extend or shrink the torso hem and sleeve areas in line with the physical size labels, and that the resulting try-on images preserve visual quality comparable to non-size-aware baselines.","pith_inferences":["Inference: a practical deployment would need to estimate the reference garment size from a single photograph, since the method assumes that size is given at inference time and the paper does not provide such an estimator.","Inference: the residual-mask mechanism could be applied to other garment attributes beyond size, such as length or looseness, by replacing the physical-size feature vector with a different conditioning signal.","Inference: the proposed Size Evaluation Metric could be reused to evaluate body-garment alignment in related tasks such as size recommendation and fit prediction, not only virtual try-on.","Inference: the paper's suggestion to normalize input images by pixel-to-physical-size ratio could in principle support cross-camera use, but the authors do not test that scenario."],"forward_implications":["A user can choose an arbitrary target size label for a garment shown in a different size, and the predicted try-on mask will extend or shrink the hem and sleeves before the clothes texture is warped onto the person.","The residual-mask formulation makes size adjustment robust to small differences between reference and target silhouettes, since the network only needs to predict the delta between them rather than a whole new mask.","The new dataset pairs images of the same person in the same garment across sizes with synchronized poses, providing a training signal that previous try-on datasets lacked because they had no size annotations.","The Size Evaluation Metric gives a quantitative handle on size correctness, measuring torso-hem and sleeve area changes separately instead of relying only on full-image perceptual similarity.","Because the mask network is modular, it can be plugged into existing warping-based virtual try-on systems in place of their mask-estimation stage."],"supporting_citations":[{"why":"Supplies the Thin Plate Spline warping and content fusion modules used to produce the final try-on image, and serves as the main baseline without size control.","marker":"[39]"},{"why":"Provides the standard virtual try-on dataset that motivates the new size-annotated dataset because it lacks physical size labels.","marker":"[16]"},{"why":"Densepose is used in the Size Evaluation Metric to identify torso and sleeve regions for the hem and sleeve area comparison.","marker":"[14]"},{"why":"Provides the pose and keypoint estimator used in posture matching to build the paired dataset.","marker":"[5]"},{"why":"Graphonomy is used as the segmentation estimator that produces the reference mask and the ground-truth try-on mask.","marker":"[13]"},{"why":"Provides the Dice loss applied to the residual mask, focusing training on the small differences between reference and target masks.","marker":"[30]"},{"why":"Supplies the adversarial loss used to make the deformed mask boundary look realistic.","marker":"[20]"},{"why":"StyleGAN is the image-style-transfer baseline in the comparison, with style parameters manually adjusted to try to change clothes size.","marker":"[23]"}],"fun_headline_variants":["Size-aware virtual try-on resizes hems and sleeves","Virtual try-on now respects physical clothing sizes","Mask deformation adjusts garment size in try-on","Try-on images that honor labeled size differences","Physical size labels now guide virtual try-on fit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the physical size of the clothes worn by the person in the reference image is known at inference time, because the size feature extractor needs that reference size alongside the target size; nothing in the paper estimates it from an arbitrary photo.","fun_headline_variants_meta":{"raw":{"variants":["Size-aware virtual try-on resizes hems and sleeves","Virtual try-on now respects physical clothing sizes","Mask deformation adjusts garment size in try-on","Try-on images that honor labeled size differences","Physical size labels now guide virtual try-on fit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1607,"prompt_tokens":898,"completion_tokens":709,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":639}},"tokens_in":514,"tokens_out":709,"duration_ms":7089,"temperature":1.0,"reasoning_tokens":639,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:53:44.688301+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a reference image with a known reference size and generate deformed masks for target sizes S, M, L, and XL; if the hem and sleeve areas of the deformed mask do not increase monotonically with the labeled target size, or if swapping the reference and target sizes does not reverse the size ordering, the network is not actually conditioning on physical size. A concrete test is to train only on certain garment types and then evaluate on a never-seen garment while holding the person fixed: if the Size Evaluation Metric on that held-out garment is close to chance, the size control does not generalize to new garments.","supporting_citations":[{"cited_title":"Towards photo-realistic virtual try-on by adaptively generating↔preserving image content","cited_arxiv_id":null,"evidence_quote":"Supplies the Thin Plate Spline warping and content fusion modules used to produce the final try-on image, and serves as the main baseline without size control."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the standard virtual try-on dataset that motivates the new size-annotated dataset because it lacks physical size labels."},{"cited_title":"Densepose: Dense human pose estimation in the wild","cited_arxiv_id":null,"evidence_quote":"Densepose is used in the Size Evaluation Metric to identify torso and sleeve regions for the hem and sleeve area comparison."},{"cited_title":"Realtime multi-person 2d pose estimation using part affinity fields","cited_arxiv_id":null,"evidence_quote":"Provides the pose and keypoint estimator used in posture matching to build the paired dataset."},{"cited_title":"Graphonomy: Universal human pars- ing via graph transfer learning","cited_arxiv_id":null,"evidence_quote":"Graphonomy is used as the segmentation estimator that produces the reference mask and the ground-truth try-on mask."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the adversarial loss used to make the deformed mask boundary look realistic."}],"review_version":1}