{"id":"34160fd2-0402-4931-b91d-b8cace415eeb","arxiv_id":"2506.12348","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A per-garment virtual try-on method for loose-fitting garments uses a garment-invariant pose representation and a recurrent ConvLSTM synthesis network to achieve temporally smoother try-on video at about 10 fps.","lead":"This paper presents a per-garment virtual try-on system for loose-fitting clothing that combines a garment-invariant body representation with a recurrent video-synthesis network. It targets two known problems in per-garment try-on: unreliable body segmentation under baggy garments and flickering video output.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantitative support for the temporal-consistency claim is not established: VFID is computed from unsynchronized videos, gains over the baseline are tiny, and no direct temporal metric is reported.","rationale":"The reader's weakest_assumption focused on the garment-invariant representation, but the more load-bearing problem is the lack of valid quantitative evidence for the temporal-consistency part of the central claim. Even if the garment-invariant representation is perfectly invariant, the headline claim of temporal superiority over the baseline rests on a metric that is not designed to measure frame-to-frame coherence, is computed on non-synchronized videos, and shows only marginal differences without error bars. This is a correctness risk in the evaluation, not just a missing value. The reader did mention the small and inconsistent gains in the rationale, so there is partial agreement, but the specific selection of weakest assumption differs. The verdict remains CONDITIONAL because the issue is addressable with a proper temporal metric and statistical testing, rather than a fundamental flaw in the method's design.","tokens_in":14194,"tokens_out":3859,"duration_ms":48001,"concrete_test":"Compute a direct frame-to-frame temporal jitter metric, such as the mean absolute difference between consecutive synthesized frames in garment-covered regions under static or slowly moving poses, for 'Ours', 'Ours w/o ConvLSTM', and Baseline [57], repeated over 10 random seeds. If 'Ours' is not significantly lower than Baseline or 'Ours w/o ConvLSTM', the temporal-consistency claim should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the method 'outperforms existing approaches in both image quality and temporal consistency' is not quantitatively established for temporal consistency. In Section 4.3, VFID is computed between the generated try-on video and a video of the same person wearing the actual loose-fitting garment, but the authors state these videos 'are not strictly synchronized.' VFID measures the distribution of I3D features over the whole video; it does not require frame-to-frame correspondence and is not a direct measure of temporal coherence or jitter. A video with strong frame-to-frame jitter can still have a feature distribution similar to a smooth reference if the clip-level statistics are close. Moreover, Table 1 shows VFID improvements over the baseline [57] of only 0.002 to 0.012 across the four garments, and no error bars or significance tests are reported, so these differences are within plausible noise. The only ablation of the ConvLSTM module (Section 4.6) is a qualitative difference-map visualization, which does not quantify temporal consistency. Thus the load-bearing claim of temporal superiority is not supported by the evidence presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a per-garment virtual try-on method for loose-fitting garments. It introduces a 'garment-invariant representation' that concatenates a rendered SMPL mesh with zeroed body-shape parameters (a virtual measurement garment) and grouped joint heatmaps with hip/knee joints removed. This representation is mapped to a simplified DensePose-style semantic map by a per-person BodyMap network, which is then used to generate per-garment training data for a recurrent garment synthesis network (pix2pixHD extended with a ConvLSTM). At inference, the user wears a tight-fitting garment, from which the virtual measurement garment and a DensePose semantic map are extracted and fed into the recurrent network, producing temporally coherent try-on results at about 10.5 fps on an RTX 4090. The paper reports qualitative and quantitative comparisons against OOTDiffusion, ViViD, and the authors' prior method [57], along with ablation studies of the garment-invariant representation and the ConvLSTM module.","tokens_in":14436,"tokens_out":8333,"duration_ms":95520,"significance":"If the claimed performance were fully established, this would be a practical contribution to per-garment virtual try-on: it addresses a known weakness of prior per-garment methods on loose-fitting garments and adds temporal consistency in a lightweight, real-time architecture. The idea of using a body representation that is relatively insensitive to garment shape is well motivated, and the ablations comparing DensePose, different heatmap variants, the virtual measurement garment, and their combination are informative. The paper is clearly written, and the limitations are discussed honestly. However, the quantitative evidence for the central claims—particularly temporal superiority—is currently weak, and the invariance of the proposed representation is not directly measured. The contribution is therefore plausible but not yet convincingly demonstrated.","major_comments":[{"comment":"The central temporal-consistency claim is not quantitatively established. VFID is computed on videos that the authors state are 'not strictly synchronized,' and no error bars, confidence intervals, or significance tests are reported. In Table 1, the VFID improvements of 'Ours' over 'Baseline [57]' are 0.016, 0.003, 0.012, and 0.012 across the four garments, which are small relative to typical VFID variability. VFID is a clip-level distributional metric based on I3D features and does not directly penalize frame-to-frame jitter; a video with strong jitter can have feature statistics close to a smooth reference. The only ConvLSTM ablation (Section 4.6, Figure 16) is a qualitative difference map. The authors should report a direct temporal-coherence metric on synchronized sequences, such as mean absolute difference or LPIPS between consecutive frames, a temporal warping error, or a user study, together with variance estimates, and should clarify why the per-frame loss in Eq. (4), which contains no explicit temporal term, is expected to enforce coherence.","section":"Section 4.3, Table 1; Section 4.6"},{"comment":"The paper's central assumption that the proposed representation is garment-invariant is not directly tested. Figure 5 shows that joint positions and confidence scores for hips and knees shift with garment type, motivating their removal, but no quantitative evaluation compares the extracted representation for the same person in tight-fitting versus loose-fitting garments under matched poses, nor measures downstream BodyMap accuracy. Since BEV pose estimation, like DensePose, may degrade when the body is occluded by loose fabric, the BodyMap network could receive inputs from a different distribution than its training data. The qualitative ablation in Figure 15 is insufficient to support the invariance claim, and the paper explicitly acknowledges in Section 4.5 that no quantitative evaluation of semantic-map accuracy is provided.","section":"Section 3.2, Figure 15"},{"comment":"The inference pipeline does not specify whether the body-shape parameters beta are fixed to zero when extracting the virtual measurement garment I_vm from the user's input frame. In Section 3.2, beta is fixed to zero for the garment-invariant representation used to train the BodyMap network and to generate per-garment datasets. If the same zero-beta I_vm is used at inference, this should be stated explicitly; if not, the I_vm component of the hybrid representation at inference has a different distribution from the training data, recreating the very mismatch the method aims to avoid. In either case, the effect of beta on the hybrid representation and on generalization to unseen body shapes needs to be discussed and, ideally, tested.","section":"Section 3.6"},{"comment":"The comparison with OOTDiffusion and ViViD is acknowledged to be unfair because these methods cannot be retrained on the per-garment datasets; the large metric gaps in Table 1 therefore mostly show that off-the-shelf universal models underperform on rarely seen garment types. For the closest baseline [57], the image-quality results are mixed: in Table 1, 'Ours' has slightly worse KID and FID than 'Baseline' for the dress (0.073 vs 0.071 and 67.26 vs 66.75) and slightly worse FID for the hanbok (27.01 vs 26.92). Thus the abstract's claim that the method 'outperforms existing approaches in both image quality and temporal coherence' is stronger than the evidence currently supports. Please qualify the claim and, if possible, add a fairer comparison or a statistical significance analysis.","section":"Section 4.1, Table 1, Abstract"}],"minor_comments":[{"comment":"The introduction states 'approximately 10 frames per second on a standard PC,' but the frame rates in Table 2 are measured on an NVIDIA RTX 4090, which is not a typical 'standard PC' GPU. Please rephrase, or report performance on a more representative device.","section":"Section 1, Table 2"},{"comment":"The text says the baseline 'occasionally surpasses our method in certain image quality metrics,' but Table 1 also shows that 'Ours w/o ConvLSTM' sometimes beats 'Ours' (e.g., FID for hanbok and dress). Clarify whether the recurrent module is intended to trade a small amount of image quality for temporal consistency, and if so state this trade-off explicitly.","section":"Section 4.3, Table 1"},{"comment":"The ablation study on the garment-invariant representation is entirely qualitative. Consider reporting a proxy quantitative metric, such as the accuracy of the estimated semantic map on tight-fitting frames of the same person where DensePose can serve as a reference, or a measure of body-part detection rate under loose garments.","section":"Section 4.5"},{"comment":"Several claims about high memory usage and dataset bias in image-based virtual try-on methods are not substantiated with specific citations. Please add references or soften these statements.","section":"Section 2.2"},{"comment":"The caption contains the typo 'ReGarSynt' instead of 'ReGarSyn'. Please correct it.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"This is a sensible system paper with a plausible pipeline and honest limitations, but the quantitative evidence for the advertised temporal-consistency advantage is thin. The main baseline is the authors' own prior method [57], and the claimed improvements over it in Table 1 are small and without variance estimates; the invariance of the proposed representation is also only argued qualitatively. These issues are addressable with additional experiments, so I recommend major revision rather than rejection. The authors should also clarify the beta handling at inference, as this bears directly on the central distribution-mismatch argument."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece here is the two-stage BodyMap network: instead of feeding DensePose a loose-fitting garment image and hoping for the best, the authors extract a six-channel representation (joint heatmaps with hip/knee dropped, plus a rendered SMPL mesh with shape parameters zeroed) and train a per-person pix2pixHD network to map it to a body semantic map. That is a sensible way to decouple garment appearance from body estimation, and the qualitative ablation in Figure 15 does show that the representation fixes the missing-lower-body artifacts DensePose produces on a hanbok skirt. The second contribution, a ConvLSTM inserted into pix2pixHD for recurrent garment synthesis, is simpler but reasonably motivated: per-frame synthesis jitters on loose garments, and feeding temporal state at 10.5 fps with constant memory is a practical answer. The writing is clear, the limitations section is honest about scalability, the tight-fitting inference assumption, and the lack of physical dynamics modeling.\n\nThe soft spot is the evaluation, and it is the load-bearing one. The abstract claims superiority in both image quality and temporal consistency. Image quality is at best a wash: in Table 1 the baseline beats the proposed method on KID for the hanbok and dress, and on FID for the hanbok, and the differences are small. Temporal consistency rests on VFID computed on videos the authors explicitly say are not strictly synchronized. VFID is a distributional metric on clip-level I3D features; it does not measure frame-to-frame coherence, and a jittery video can have VFID close to a smooth one. The VFID gains over the baseline are 0.002 to 0.012, with no error bars or significance tests. The only direct temporal evidence is the qualitative difference map in Figure 16, which is suggestive but not quantified. The garment-invariant representation is also validated only qualitatively; invariance is asserted by construction, not measured. No code or data are provided, so the numbers cannot be checked.\n\nThese are fixable problems rather than fatal ones. The method is plausible, the engineering is careful, and the problem is relevant for e-commerce try-on. But the paper currently asks the reader to take the central temporal-consistency claim on faith. A serious referee should push for a direct temporal metric (e.g., jitter or warp error on synchronized or roughly aligned sequences), error bars over multiple runs or clips, and ideally code/data release. I would not cite it in its current form, but I would send it to peer review: the core idea deserves scrutiny and a revision.","headline":"A coherent per-garment try-on system with a clever garment-invariant body-map estimator and recurrent synthesis, but the temporal-consistency claim is under-supported by the quantitative evidence.","tokens_in":14954,"tokens_out":2175,"would_cite":false,"duration_ms":24490,"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 garment-invariant pose representation plus a recurrent synthesis network lets per-garment virtual try-on handle loose clothing in real time while suppressing frame-to-frame jitter.","keywords":["virtual try-on","loose-fitting garments","garment-invariant representation","body semantic maps","temporal consistency","recurrent synthesis","ConvLSTM","real-time"],"falsifier":"A direct test would compare the garment-invariant representation for the same person in the same pose wearing a tight garment versus each loose garment; if the remaining joint heatmaps or the rendered SMPL image shift by more than a few pixels in the loose condition, the invariance assumption fails. A second check would measure the BodyMap network's semantic-map output against manually annotated or depth-sensor ground truth under loose garments, rather than relying on the qualitative images in the paper's Figure 15.","tokens_in":14010,"feed_emoji":"👗","tokens_out":5965,"duration_ms":64235,"temperature":0.7,"pith_summary":"This paper claims that per-garment virtual try-on for loose-fitting clothing can be made visually credible and temporally stable by replacing the usual DensePose body map with a two-stage semantic-map estimator and by making the garment synthesizer recurrent. Existing per-garment methods train on tight-fitting data and then fail when the target garment hides the body contour; the proposed representation is designed to be robust to that occlusion. The paper also claims that adding a lightweight convolutional recurrent module keeps successive frames coherent without sacrificing real-time speed, reporting about 10.5 frames per second. If correct, this removes two known failure modes of per-garment try-on and makes live virtual try-on for loose garments practical.","feed_headline":"Loose-garment try-on hits 10.5 fps with jitter-free video","feed_subtitle":"A garment-invariant pose map plus a recurrent synthesis network keeps loose skirts and sleeves stable across frames.","key_machinery":"The load-bearing object is the garment-invariant representation, a six-channel image formed by concatenating a rendered SMPL mesh with a grid-pattern texture and body shape parameters fixed to zero with a three-channel RGB encoding of joint heatmaps that excludes the hip and knee joints. Its job is to carry enough pose information to reconstruct the hidden body semantic map while being insensitive to whether the person is photographed in tight or loose clothing. The second mechanism is the recurrent synthesis block, a lightweight ConvLSTM inserted into the pix2pixHD generator, which carries cell and hidden states across frames. Together they support the paper's two fixes: reliable semantic-map estimation for dataset generation and frame-to-frame coherence during inference.","core_discovery":"The core claim is that a per-garment virtual try-on system can handle loose-fitting garments if it separates the problem into two learned mappings. First, a per-person BodyMap network translates a garment-invariant representation characterized by a rendered SMPL virtual measurement garment with body shape parameters zeroed, concatenated with joint heatmaps from which hip and knee joints have been removed, into a simplified human body semantic map. This map is robust enough to train a garment-specific synthesis network from footage of a person wearing the loose garment. Second, a recurrent garment synthesis network inserts a ConvLSTM into a pix2pixHD-style generator so that synthesis at each frame depends on previous cell and hidden states, enforcing temporal coherence. The paper reports that this pipeline gives better image quality and temporal consistency than general image-based and video-based try-on methods and than the prior per-garment baseline, with temporal coherence measured by Video FID and image quality by KID and FID.","pith_inferences":["If the invariance claim holds quantitatively, the same representation could enable body semantic estimation for users wearing loose clothing during inference, removing the tight-fitting-garment assumption that the paper lists as a limitation.","The fixed-size recurrent state suggests the method could run as an online streaming loop with bounded latency, extending beyond pre-recorded clips to live camera try-on.","A direct numerical test of invariance, measuring how much the remaining joint heatmaps shift between tight and loose versions of the same pose, would turn the qualitative ablation into a predictive failure analysis.","Because the per-person BodyMap network can be reused across all garments worn by the same model, the framework points naturally toward per-garment-type training to cut the cost of collecting a dedicated dataset for every individual garment."],"forward_implications":["Per-garment datasets for loose garments can be generated from ordinary video of one human model without DensePose failures by using the BodyMap network to estimate semantic maps.","Try-on results for loose garments become temporally stable, with lower Video FID than per-frame baselines, reducing jitter at hems and sleeves.","Inference remains real-time at about 10.5 frames per second and runs with a constant memory footprint on input sequences of arbitrary length because the recurrent state is fixed-size.","The trained ReGarSyn network generalizes to unseen body shapes despite being trained on a single human model.","The expensive garment-invariant representation is needed only while building per-garment training data; at inference the BodyMap network runs once per person and DensePose is still used for tight-clothed users."],"supporting_citations":[{"why":"Supplies the per-garment dataset collection pipeline and the baseline method whose DensePose-based semantic maps fail on loose garments.","marker":"[57]"},{"why":"Provides DensePose body semantic maps, which the paper argues degrade under loose garments and uses as the direct comparison for semantic-map estimation.","marker":"[19]"},{"why":"Provides the pix2pixHD conditional GAN backbone used for both the per-person BodyMap network and the ReGarSyn garment synthesis network.","marker":"[56]"},{"why":"Supplies the ConvLSTM module whose cell and hidden states give the recurrent garment synthesis network its temporal memory.","marker":"[49]"},{"why":"Provides the monocular 3D pose regression that produces the SMPL mesh used to render the virtual measurement garment.","marker":"[51]"},{"why":"Provides the joint heatmaps used in the garment-invariant representation; the paper discards hip and knee heatmaps from them.","marker":"[30]"},{"why":"Introduces the virtual measurement garment and the per-garment intermediate representation that this paper extends to a full-body template.","marker":"[58]"},{"why":"Diffusion-based image virtual try-on baseline used in comparisons to show that universal networks fail on loose garments.","marker":"[60]"},{"why":"Video virtual try-on baseline used in comparisons for both image quality and temporal consistency.","marker":"[14]"}],"fun_headline_variants":["Real-time try-on for loose garments with no jitter","Steady loose-fit try-on: recurrent synthesis kills jitter","Garment-invariant maps plus RNN make loose try-on stable","Per-garment try-on that handles loose clothes in real time","Loose-garment virtual try-on stays smooth across frames"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the six-channel pose representation is truly invariant to what a person is wearing; the paper supports this only with a qualitative ablation, not a measured invariance test. If the pose estimator drifts when loose fabric hides the body, the BodyMap network receives inputs outside its training distribution and the estimated semantic maps degrade.","fun_headline_variants_meta":{"raw":{"variants":["Real-time try-on for loose garments with no jitter","Steady loose-fit try-on: recurrent synthesis kills jitter","Garment-invariant maps plus RNN make loose try-on stable","Per-garment try-on that handles loose clothes in real time","Loose-garment virtual try-on stays smooth across frames"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000824,"raw_usage":{"total_tokens":3609,"prompt_tokens":956,"completion_tokens":2653,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":2576}},"tokens_in":572,"tokens_out":2653,"duration_ms":23657,"temperature":1.0,"reasoning_tokens":2576,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:53:08.962546+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would compare the garment-invariant representation for the same person in the same pose wearing a tight garment versus each loose garment; if the remaining joint heatmaps or the rendered SMPL image shift by more than a few pixels in the loose condition, the invariance assumption fails. A second check would measure the BodyMap network's semantic-map output against manually annotated or depth-sensor ground truth under loose garments, rather than relying on the qualitative images in the paper's Figure 15.","supporting_citations":[{"cited_title":"Low-Barrier Dataset Collection with Real Human Body for Interactive Per-Garment Virtual Try-On","cited_arxiv_id":"2506.10468","evidence_quote":"Supplies the per-garment dataset collection pipeline and the baseline method whose DensePose-based semantic maps fail on loose garments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides DensePose body semantic maps, which the paper argues degrade under loose garments and uses as the direct comparison for semantic-map estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the pix2pixHD conditional GAN backbone used for both the per-person BodyMap network and the ReGarSyn garment synthesis network."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ConvLSTM module whose cell and hidden states give the recurrent garment synthesis network its temporal memory."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the monocular 3D pose regression that produces the SMPL mesh used to render the virtual measurement garment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the joint heatmaps used in the garment-invariant representation; the paper discards hip and knee heatmaps from them."}],"review_version":1}