{"id":"0e204037-c58b-4515-a78f-f30c5ae08178","arxiv_id":"2505.09998","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A VR-sketch-conditioned diffusion model, trained in three stages with curriculum learning and a new 969-pair dataset, generates plausible 3D garments from freehand 3D sketches.","lead":"This paper presents a system that turns rough 3D sketches, drawn in the air with a VR controller, into complete 3D digital garments, using a diffusion model trained on a new dataset of 969 sketch-garment pairs collected from amateur users. The work targets a practical gap: letting non-designers create virtual clothing for avatars, virtual try-on, and metaverse applications.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'significantly outperforms' claim rests on a 1.5% CD margin and a worse IoU, with no error bars or participant-disjoint split; style leakage could explain both the margin and the user-study gap.","rationale":"The paper does provide a new dataset and a plausible three-stage pipeline, and the ablation in Table IV shows internal value for pretraining and curriculum learning. However, the headline comparison to prior work is the weakest link: the numbers in Table I are close and contradictory (better CD, worse IoU), and no uncertainty quantification is offered. The reader's concern about participant style leakage is real and likely explains how such a large user-study gap could arise from such small metric differences. I therefore agree with the CONDITIONAL verdict, but I would frame the load-bearing issue slightly differently: even without leakage, the reported metrics do not by themselves support 'significantly outperforms' without significance testing. The condition should be that the comparison is re-run on a participant-disjoint split with repeated seeds and significance tests, and that the dataset and code are released for verification.","tokens_in":13622,"tokens_out":3382,"duration_ms":34347,"concrete_test":"Redo the Table I evaluation with a participant-disjoint split: hold out all sketch-shape pairs from at least two of the ten participants, train on the rest, and run 5 seeds. Report mean and 95% CI for CD and IoU, and a paired significance test (e.g., Wilcoxon signed-rank) over per-test-sample metrics. Also state whether any test sketch was drawn by a participant seen during training. If the CD margin disappears or IoU remains lower, the headline outperformance claim should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim ('significantly outperforms existing baselines in both fidelity and usability') is not established by the evidence in Table I. Ours achieves CD 0.0597 vs Deep3DVRSketch 0.0606, a 1.5% relative margin, while its IoU (0.3190) is below Deep3DVRSketch (0.3252). No standard deviations, confidence intervals, significance tests, or per-sample paired comparisons are reported, so the CD margin could be within run-to-run noise. The split described in Section V.A ('8:2 ratio') is not stated to be participant-disjoint or garment-disjoint; Section III reports sketches from only 10 participants. If test sketches were drawn by the same 10 participants whose style appears in training, the model can memorize a few personal stroke habits rather than generalize to arbitrary users. The user-study gap (MOS 4.59 vs 3.28) may then reflect style memorization rather than superior general fidelity. Thus the load-bearing condition — that the observed outperformance is real and generalizable — is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a sketch-driven 3D garment generation framework for AR/VR users. The authors introduce KO3DClothes, a dataset of 969 paired 3D garment meshes and human-drawn 3D VR sketches collected from 10 non-professional participants, and a three-stage training strategy: pre-training a conditional diffusion model on point-cloud shapes, training a Transformer sketch encoder in a shared latent space, and jointly fine-tuning both with an adaptive curriculum learning schedule. The method is evaluated against 3DSketch2Shape and Deep3DVRSketch using voxel IoU and Chamfer Distance, together with two user studies. The central claim is that the system interprets imprecise free-hand input, produces realistic personalized garments, and significantly outperforms existing baselines in both fidelity and usability.","tokens_in":13831,"tokens_out":3326,"duration_ms":35068,"significance":"If the claims were fully supported, the paper would make a useful contribution: it addresses a practical data-scarcity problem in sketch-based garment generation, contributes a new paired dataset of real user VR sketches and garment meshes, and combines a point-cloud diffusion prior, shared-latent alignment, and curriculum learning in a way that could be reused by other sketch-to-shape tasks. The ablation study is a strength because it isolates the role of the point-cloud prior and curriculum learning, and the comparison against 3DSketch2Shape and Deep3DVRSketch is methodologically sensible. However, the headline quantitative claim is not supported by the numbers in Table I: on IoU the proposed method is worse than Deep3DVRSketch, and on Chamfer Distance the improvement is only 0.0009 with no error bars or significance tests. The data split is not documented as participant-disjoint, which threatens the generalization claim, and the user studies are not statistically grounded. These are load-bearing issues for the paper's main conclusion, but they appear addressable with additional analysis, statistical reporting, and a more carefully controlled evaluation protocol.","major_comments":[{"comment":"The abstract claims that the method 'significantly outperforms existing baselines in both fidelity and usability,' but Table I shows IoU of 0.3190 for the proposed method versus 0.3252 for Deep3DVRSketch, and a Chamfer Distance improvement of only 0.0009 (0.0597 versus 0.0606). No standard deviations, confidence intervals, significance tests, or per-sample paired comparisons are reported. On the reported evidence, the method is not better than Deep3DVRSketch on IoU, and the CD margin is within the range that could arise from run-to-run noise. This unsupported 'significant' claim needs to be either substantiated with statistical evidence or removed.","section":"V.B, Table I"},{"comment":"The train/test split is described only as an 8:2 ratio of the KO3DClothes dataset, while Section III states that all sketches were drawn by only 10 participants. The paper does not state that test sketches were drawn by held-out participants, nor that the split is garment-disjoint or style-stratified. If the same participants' drawing styles appear in both training and test sets, the reported metrics, including the user-study preference, may reflect memorization of a small number of personal stroke habits rather than generalization to arbitrary users. The authors should report a participant-disjoint evaluation or otherwise demonstrate that drawing style does not leak across the split.","section":"V.A and III"},{"comment":"The user study is described as 15 designers rating 12 generated results, but the paper does not report how many tasks were rated per method, whether the raters were the same 10 participants who created the training sketches, whether the raters were blinded to the method, or any inter-rater agreement. No significance or effect-size statistics are given for the MOS differences. The reported MOS gap (4.59 versus 3.28 for fidelity) could be affected by rater bias, task difficulty imbalance, or style leakage from the training set. A controlled protocol with independent raters, balanced tasks, and paired statistical tests is required to support the usability claim.","section":"V.C, Tables II and III"}],"minor_comments":[{"comment":"The phrase 'significantly outperforms' appears in the abstract and introduction, but the quantitative results in Table I do not support a significance claim; please rephrase these statements to match the evidence actually presented.","section":"Abstract and I"},{"comment":"The section heading 'The Proposed Multi-stage Training Steategy' contains a typo; it should read 'Strategy.'","section":"IV.B"},{"comment":"The sentence 'Just as beginners in sketching typically start to learn with with simple and flexible shapes' contains a duplicated 'with.'","section":"I"},{"comment":"The figure label 'DeepVRSketch+' is inconsistent with the method name 'Deep3DVRSketch+' used in the rest of the paper; please unify the terminology.","section":"Fig. 1"},{"comment":"The text refers to a 'KO3DClothes+ dataset' but the dataset is introduced as 'KO3DClothes' in Section III; please use one name consistently.","section":"V.A"},{"comment":"In Equation (2), the definition of y as 'the SDF value corresponding to the hand-drawn sketch' is unclear because the input is a point cloud; the authors should explain how SDF values are assigned to sketch points and how the difficulty score is computed from them.","section":"IV.C"}],"recommendation":"major_revision","confidential_remarks":"The strongest baseline, Deep3DVRSketch, is the authors' own prior work, and the proposed method is named Deep3DVRSketch+. This is a legitimate comparison but should be clearly disclosed as an extension of the baseline in the paper itself. The main issue is that the empirical evidence for the headline claim is weak: one table with a mixed result and no statistics, a split that may leak drawing style, and user studies without significance testing. These can be addressed with re-analysis and additional experiments, so I see this as a major-revision rather than a reject, provided the authors either produce the missing statistical evidence or substantially temper the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The dataset is the real contribution here, and the application is sensible: freehand 3D VR sketches as a condition for garment generation. KO3DClothes, 969 paired sketches and garments drawn by 10 non-experts, is genuinely useful for a niche that has almost no public data. The three-stage training idea—point-cloud prior, sketch encoder, joint fine-tuning, plus curriculum learning—is a reasonable way to stretch a small dataset, and the ablation suggests curriculum learning matters.\n\nThat said, the central performance claim does not hold up. Table I shows your method gets IoU 0.3190 against Deep3DVRSketch's 0.3252—you are worse on that metric—and the Chamfer distance advantage is 0.0597 versus 0.0606, a 1.5% relative difference with no error bars, no significance test, and a single train/test split. Calling that 'significantly outperforms' in the abstract is not defensible. The stress-test note is correct: if the same ten users' drawing styles appear in both training and test, the model could memorize personal stroke habits rather than generalize. The paper does not state that the split is participant-disjoint, and that omission is load-bearing.\n\nThere are also smaller but real issues. The dataset is built from DeepFashion3D, but reference [11] is the PAMIR paper, not the DeepFashion3D paper. The dataset name flips between KO3DClothes and KO3DClothes+; the figure caption uses DeepVRSketch+ while the method is Deep3DVRSketch+. These are easy fixes but suggest a rushed manuscript.\n\nA word on what is not wrong: the method itself is a coherent combination of known components, not a fitted story. The user study, while gated on the same leakage concern, at least reports standard deviations. The baselines are few—only two, one being the authors' own prior—but for a new dataset that is an acceptable starting point.\n\nWho benefits: researchers working on sketch-based 3D modeling or garment generation will want the dataset. The paper deserves a serious referee because the dataset and application are valuable, and the training strategy is worth discussing even if the quantitative claims need an overhaul. I would recommend peer review, but with a clear request: report the numbers honestly, add error bars and a participant-disjoint split, and release the data or provide a detailed protocol. Without that, the empirical contribution remains plausible rather than demonstrated.","headline":"The dataset and the input modality are the real contributions; the abstract's 'significantly outperforms' claim is not supported by the paper's own Table I.","tokens_in":768,"tokens_out":911,"would_cite":false,"duration_ms":26133,"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":"Drawing a garment in the air with a VR controller can yield a realistic 3D model, according to this paper.","keywords":["3D garment generation","VR sketching","sketch-to-3D generation","diffusion models","curriculum learning","point cloud latent space","KO3DClothes dataset","digital fashion"],"falsifier":"Re-run the evaluation with a leave-participants-out split of KO3DClothes—no sketcher and no garment shape shared between training and test—and check whether the Chamfer-distance margin over the strongest prior VR-sketch baseline persists.","tokens_in":13354,"feed_emoji":"👗","tokens_out":6719,"duration_ms":63084,"temperature":0.7,"pith_summary":"This paper tries to establish that ordinary people without design training can create personalized 3D clothing by sketching in AR/VR, with an AI model filling in the realism. It argues that imprecise freehand strokes should be treated as a loose condition for a generative model rather than a precise input to regress from, and that a three-stage training scheme—point-cloud diffusion prior, sketch encoder into the same latent space, and joint fine-tuning with adaptive curriculum learning—lets a small paired dataset support this. If true, it matters because it would lower the main barrier to user-generated virtual fashion: instead of modeling software, users only need to draw roughly in the air. The paper's own numbers support the usability claim most strongly: users rated its outputs 4.59/5 for fidelity and 4.62/5 for quality, well above the prior VR-sketch method's 3.28 and 3.39.","feed_headline":"VR sketches become wearable 3D garments","feed_subtitle":"A diffusion pipeline turns loose strokes into usable digital clothes and beats prior sketch-to-shape tools.","key_machinery":"The load-bearing object is a shared 3D latent space: a 1024-dimensional feature embedding produced by a pre-trained point-cloud encoder for garment point clouds and by a Transformer-based sketch encoder for VR sketches, with the conditional diffusion model accepting these embeddings through multi-head cross-attention. The second mechanism is adaptive curriculum learning: a difficulty score per sample based on signed-distance-field prediction errors, a pacing function that feeds easy samples first, and an update rule that re-scores the sample pool during training. Together, these mechanisms let a small paired dataset of 969 sketch–garment pairs drive a high-resolution garment generator.","core_discovery":"On its own terms, the paper's central claim is that user-created 3D VR sketches—imprecise, incomplete, and stylistically varied—carry enough information to condition a generative model into producing realistic 3D garments. The system, which the paper calls Deep3DVRSketch+, achieves this by pre-training a coarse-to-fine conditional diffusion model on garment point clouds, freezing it while training a Transformer-based sketch encoder into the same latent space, then jointly fine-tuning both stages with an adaptive curriculum that orders samples from easy to hard. On the introduced KO3DClothes dataset of 969 paired sketches and garments, the method reports a Chamfer distance of 0.0597, marginally better than the strongest prior VR-sketch baseline (0.0606) and clearly better than a normalizing-flow baseline (0.0820), while IoU is comparable (0.3190 vs 0.3252 and 0.3188). In a 15-designer user study, fidelity and quality mean opinion scores are well above the prior VR-sketch method, and controllability and usefulness beat a 2D-sketch baseline by large margins. The paper's framing is that the system targets overall garment shape—sleeves, length, silhouette—while leaving wrinkles and dynamic folds to downstream clothing simulation.","pith_inferences":["A direct extension would be leave-one-participant-out evaluation: the paper's 8:2 split is not stated to separate sketching styles, so holding out entire users would test whether the model truly generalizes to unfamiliar stroke habits.","Because the sketch is treated as a loose condition rather than a precise contour, the same two-stage latent-alignment recipe could transfer to other shape domains by swapping the pre-trained point-cloud prior, such as furniture or avatars.","The reported margin over the strongest prior baseline is small on Chamfer distance; an implication is that the usability advantage, not geometric accuracy, may be the more robust selling point worth reproducing.","A testable extension is to measure per-sketch diversity: if the loose condition allows several plausible garments per sketch, the method could support interactive design exploration rather than single-output reconstruction."],"forward_implications":["A non-designer wearing a VR headset can produce a plausible personalized 3D garment from rough freehand strokes, removing the main technical skill barrier to virtual fashion creation.","A small paired dataset of 969 sketch–garment samples is sufficient to train a usable sketch-conditioned garment generator when a point-cloud diffusion prior and curriculum learning are used.","The system targets overall garment shape—silhouette, sleeve presence, and sleeve length—rather than fine wrinkles, on the expectation that downstream cloth simulation adds realistic dynamic folds.","Compared against two prior sketch-to-shape baselines, the method reports the lowest Chamfer distance and the best user ratings for fidelity and quality, supporting the viability of VR sketches as a design input.","The new KO3DClothes dataset of paired user-created VR sketches and 3D garments provides a benchmark resource for future sketch-based garment generation work."],"supporting_citations":[{"why":"Source of the DeepFashion3D garment models from which the KO3DClothes sketch–shape pairs are created, as the paper cites it.","marker":"[11]"},{"why":"The VR-sketch baseline the paper compares against, and also the source of the data-collection protocol and evaluation metrics.","marker":"[50]"},{"why":"Earlier VR-sketch dataset protocol that the data collection procedure is adapted from.","marker":"[51]"},{"why":"The coarse-to-fine self-conditioned SDF diffusion architecture that the two-stage generator is built on.","marker":"[52]"},{"why":"Provides the denoising training objective used to train both diffusion UNets.","marker":"[53]"},{"why":"Classifier-free guidance, used to keep generation faithful to the sketch while preserving diversity.","marker":"[54]"},{"why":"The pre-trained point-cloud encoder that defines the shared latent space the sketch encoder is trained to align with.","marker":"[55]"},{"why":"The adaptive curriculum learning method whose difficulty scoring and pacing scheme are adopted.","marker":"[58]"},{"why":"The normalizing-flow sketch-to-shape baseline that the method outperforms in quantitative and user evaluations.","marker":"[61]"}],"fun_headline_variants":["VR doodles become tailored digital outfits","Sketch in 3D, get a custom virtual garment","Diffusion turns rough VR sketches into 3D clothes","From VR scribbles to wearable digital fashion","AI designs custom garments from freehand 3D sketches"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the 8:2 training/test split prevents the ten participants' individual drawing styles from appearing in both halves; if test sketches come from the same people who produced training sketches, the reported gains may reflect memorized stroke habits rather than general understanding of freehand input.","fun_headline_variants_meta":{"raw":{"variants":["VR doodles become tailored digital outfits","Sketch in 3D, get a custom virtual garment","Diffusion turns rough VR sketches into 3D clothes","From VR scribbles to wearable digital fashion","AI designs custom garments from freehand 3D sketches"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000657,"raw_usage":{"total_tokens":3032,"prompt_tokens":997,"completion_tokens":2035,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":1960}},"tokens_in":613,"tokens_out":2035,"duration_ms":14965,"temperature":1.0,"reasoning_tokens":1960,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:19:00.682974+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the evaluation with a leave-participants-out split of KO3DClothes—no sketcher and no garment shape shared between training and test—and check whether the Chamfer-distance margin over the strongest prior VR-sketch baseline persists.","supporting_citations":[{"cited_title":"Pamir: Parametric model- conditioned implicit representation for image-based human reconstruc- tion,","cited_arxiv_id":null,"evidence_quote":"Source of the DeepFashion3D garment models from which the KO3DClothes sketch–shape pairs are created, as the paper cites it."},{"cited_title":"Rapid 3d model generation with intuitive 3d input,","cited_arxiv_id":null,"evidence_quote":"The VR-sketch baseline the paper compares against, and also the source of the data-collection protocol and evaluation metrics."},{"cited_title":"Fine- grained vr sketching: Dataset and insights,","cited_arxiv_id":null,"evidence_quote":"Earlier VR-sketch dataset protocol that the data collection procedure is adapted from."},{"cited_title":"Adaptive curriculum learning,","cited_arxiv_id":null,"evidence_quote":"The adaptive curriculum learning method whose difficulty scoring and pacing scheme are adopted."}],"review_version":1}