{"id":"ff22ed38-2553-4347-aefb-8350707f42f7","arxiv_id":"2412.06285","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A learning-based method reconstructs high-resolution garment geometry from low-resolution simulation inputs by predicting coarse shape corrections and fine wrinkle residuals.","lead":"This paper presents a neural network that turns low-resolution garment simulations into high-resolution ones with added wrinkle detail, trained on just a few examples. It could make realistic clothing animation cheaper for games, apps, and virtual try-on.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Generalization across unseen garment types and body shapes is only shown qualitatively; the single quantitative table covers training garment types, and roll-out training uses ground-truth history while test uses self-predicted history.","rationale":"The reader's weakest assumption identifies the small-data generalization premise, and I agree that the evidence for cross-garment and cross-body-shape transfer is qualitative only. I add a distinct technical concern: the roll-out evaluation is teacher-forced in training but self-feeding at test time, and the only long-horizon metric is deformation energy rather than fidelity to the HR reference. Both concerns point to the same conclusion: the central claim is plausible and well-motivated, with released code and a topology-agnostic design as genuine supporting evidence, but it is not yet quantitatively verified along the dimensions the abstract promises. The paper deserves conditional acceptance with additional experiments, which matches the reader's verdict; no verdict change is needed.","tokens_in":18412,"tokens_out":3824,"duration_ms":41776,"concrete_test":"Run the released code to compute SSIM and mean Chamfer distance to HR ground truth for garment types (i)-(v) and both unseen body shapes, across 300+ frames and 3 training seeds; report means and standard deviations for GDSR, DDE, and PhysGraph(-). If unseen-garment SSIM is more than about 0.05 below the seen-garment numbers in Table 3, or overlaps the baselines, the generalization claim is unsupported. In the same run, measure per-frame SSIM/Chamfer at roll-out lengths 50, 100, 1000, and 1500; a monotone degradation while Table 2 energies stay flat would confirm that teacher-forced training masks error accumulation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5 and Appendix C report SSIM (Table 3) only for training garments (A), (C), and (E) on an unseen motion; no quantitative result is given for the five unseen garment types (i)-(v) or the two unseen body shapes shown in Figures 6-7. Since the central claim is precisely about transfer to unseen body shapes, motions, and garment topologies, the current evidence cannot distinguish genuine generalization from interpolation among similar training outfits. A second, compounding issue is roll-out: the paper states (Section 5) that history states in training are derived from high-resolution ground truth, but at inference the network consumes its own predictions; the only long-rollout metric (Table 2) is stretching/shearing energy, which does not measure geometric or perceptual fidelity. Thus the load-bearing premise, that a small, homogeneous training set yields a transferable SR map that remains stable under self-feeding, is not quantitatively tested on the claimed axes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces GDSR, a learning-based method that takes low-resolution garment simulations and body motion as input and produces high-resolution garment geometry with fine wrinkle details. The architecture combines a mesh-graph-net that extracts super-resolution features from coarse garment dynamics and garment-body interactions, a decoder that corrects the coarse garment shape, and a hyper-net that constructs per-triangle implicit functions for wrinkle residuals. The method is trained on two dance sequences with five garment types and one fixed body shape, and the authors claim generalization to unseen motions, body shapes, and garment types, demonstrating results qualitatively and reporting SSIM comparisons on training garment types. The paper also reports a lightweight 65MB network, timing measurements, an explicit collision handling step, and releases code and data.","tokens_in":18618,"tokens_out":6296,"duration_ms":60069,"significance":"If the generalization claims are fully borne out, GDSR would be a practically valuable tool for low-budget high-resolution garment simulation: it is lightweight, fast (about 0.115s per frame including low-resolution simulation), and the released code/data would facilitate adoption. The method's architectural choices — graph-based feature extraction, explicit shape correction, and continuous per-triangle upsampling — are well motivated, and the qualitative results, especially for pleated and laced garments, are visually compelling. However, the current quantitative evidence covers only training garment types on an unseen motion, and the paper's own Section 6 acknowledges limitations such as single-material training and unresolved collisions. Consequently, the strongest claim in the abstract — robust generalization to unseen body shapes, motions, and garment types — is not yet supported by the reported experiments, which tempers the significance of the contribution at this stage.","major_comments":[{"comment":"The quantitative evaluation (SSIM on normal maps) is computed only for training garment types (A), (C), and (E), tested on an unseen motion sequence. No quantitative result is provided for the five unseen garment types (i)-(v) shown in Figure 6 or for the unseen body shapes in Figure 7, although the abstract and Section 1 explicitly claim generalization along these axes. Since the qualitative figures alone cannot distinguish genuine transfer from interpolation across similar training outfits, the central claim is not yet quantitatively established. The paper itself notes in Section 6 that it is trained on a single fabric material and that friction-dependent details are missed; these are honest limitations, but they make the absence of quantitative cross-garment evidence more conspicuous. I recommend reporting SSIM (or a geometric metric) on the unseen garments and body shapes, with error bars across frames or motion sequences, and ideally over multiple roll-out lengths.","section":"Section 5 and Appendix C, Table 3"},{"comment":"The training procedure uses history states derived from high-resolution ground truth (Section 5, first paragraph), while at inference the network consumes its own predictions in a roll-out. The only quantitative long-roll-out evaluation is stretching/shearing energy (Appendix B, Table 2), which measures deformation stability but not geometric or perceptual fidelity to the high-resolution reference. This gap matters because the self-feeding regime is exactly where accumulation errors would appear. I recommend adding a fidelity metric (SSIM or point-wise error) for roll-outs of increasing length (e.g., 1, 50, 100, 1000, and 1500 frames) and comparing against a one-step oracle that uses ground-truth history.","section":"Section 5 and Appendix B, Table 2"},{"comment":"The comparison with DDE and PhysGraph is also restricted to training garment types; no quantitative comparison is reported for unseen garments or body shapes. Given that the paper claims 'significant improvements over state-of-the-art alternatives' and the baselines are trained on the same or similar data, adding such comparisons on the unseen test garments would strengthen the central claim. This can be addressed together with the additional experiments suggested in Major Comment 1.","section":"Section 5.1 and Table 3"}],"minor_comments":[{"comment":"The phrase 'referre to' appears in the descriptions of the Decoder E and Mesh-Hyper-Net A blocks; it should read 'refer to'.","section":"Section 3"},{"comment":"The citation 'Later et al. [2018]' should be 'Lahner et al. [2018]' (the DeepWrinkles paper).","section":"Related Work"},{"comment":"The equation defining L_c_def contains mismatched norm bars, e.g., '|| || w_u(Ĉ_t|| - || w_u(Ĉ*_t)|| ||_1'; please fix the parentheses and norm-bar pairing for readability.","section":"Section 4.5"},{"comment":"The notation '|f|' is used both for a triangle and for the number of adjacent triangles; consider using a symbol such as deg(v) or N(v) to avoid ambiguity.","section":"Section 4.3"},{"comment":"The caption states that SSIM is reported for 'a t-shirt (A), a pleated short skirt (C), and a triple-lace long skirt (E)' but does not say these are training garment types; since the main text's generalization claim concerns unseen garments, the caption should explicitly clarify this point.","section":"Table 3 caption"},{"comment":"The timing sentence 'including 0.083 seconds for low-resolution simulation in Marvelous Designer (MD)' could clarify that this 0.083s is the coarse simulation time and that the GDSR network itself accounts for 0.032s (as in Table 4), to avoid confusion about what the 0.115s total includes.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"One note for the editor: the manuscript appears to be the camera-ready version of a SIGGRAPH Asia 2024 paper (ACM SA Conference Papers '24), and this report is based on the arXiv version, which matches the published format. The main concern is the gap between the strong generalization claim and the quantitative evidence, which currently covers only training garment types. The paper is otherwise competently written, and the code/data release is a positive factor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The bottom line: this is a solid SIGGRAPH Asia paper with a genuinely useful architecture for garment dynamic super-resolution, and the stress-test note gets the main weakness right - the quantitative evidence is thinner than the claims. The method itself is coherent and worth engaging with.\n\nWhat's new: GDSR casts low-to-high-resolution garment detail synthesis as a mesh-space super-resolution problem, combining a MeshGraphNet-style feature extractor, a hyper-network that generates per-triangle implicit wrinkle fields, and an explicit decoder that corrects the coarse shape before adding details. The components are borrowed - MeshGraphNet, WIRE, hyper-networks - but the joint shape-correction-plus-detail-synthesis design, with iterative roll-out for temporal coherence, is a real combination. The writing is clear, the figures make the qualitative results easy to read, and the code and data are released, which is real evidence. The baselines (DDE and PhysGraph) are handled fairly, including a PhysGraph(-) variant trained on downsampled ground truth. The 0.115 s/frame figure, including the low-res simulation, makes the practical claim credible. The limitations section is honest about friction and collision failures.\n\nThe soft spots are real but they are evaluation gaps, not flaws in the method. Table 3's SSIM numbers are the only geometric/perceptual quantitative evidence against baselines, and they cover just three training garment types on an unseen motion, with no error bars. The headline claim - generalization to unseen garment types and body shapes - rests mainly on Figures 6 and 7. The roll-out evaluation (Table 2) reports stretching and shearing energies, which is a stability check (and it does include one unseen garment), but it does not measure how far self-predicted geometry drifts from ground truth. Teacher-forced training with ground-truth history versus self-predicted history at inference is a known autoregressive gap, and the paper does not quantify the resulting error accumulation. Nothing about the architecture or the small-data setup is unsound on its own terms.\n\nThis paper is for graphics researchers working on cloth simulation and learned detail synthesis. It deserves a serious referee. My recommendation: send it to review, and expect the authors to add error bars, quantitative results on at least some unseen garment types, and an error-accumulation curve for long roll-out against HR ground truth.","headline":"Solid, well-written garment super-resolution method with released code; the generalization claims outrun the quantitative evidence, but it deserves a serious referee.","tokens_in":19072,"tokens_out":4486,"would_cite":true,"duration_ms":43382,"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 single lightweight network upscales low-resolution cloth simulation into detailed, wrinkled garments.","keywords":["garment simulation","dynamic super-resolution","wrinkle synthesis","graph neural networks","hyper-networks","implicit neural fields","cloth animation","generalization"],"falsifier":"Run the released network, without fine-tuning, on a low-resolution simulation of a garment made of a very different fabric (e.g., stiff denim) in an unseen motion, and compare the normal maps of the synthesized high-resolution geometry against a full high-resolution simulation; if the SSIM drops to the level of simple up-sampling from the coarse mesh, the learned mapping does not transfer across materials, which the paper's claims require it to do.","tokens_in":18199,"feed_emoji":"👗","tokens_out":7809,"duration_ms":70636,"temperature":0.7,"pith_summary":"This paper claims that fine-grained garment wrinkles can be synthesized from cheap, low-resolution cloth simulation by a single lightweight network, turning garment detail enhancement into a dynamic super-resolution problem. The method, GDSR, learns local 'super-resolution features' from coarse mesh dynamics and garment-body contacts, then uses a hyper-network to build a continuous implicit field of wrinkle residuals for each coarse triangle while simultaneously correcting the coarse shape. The authors report that, despite training on only two dance sequences and five outfits with one body shape and fabric, the network generalizes to unseen motions, body shapes, and garment types, and that iterative roll-out prediction keeps long 1500-frame sequences stable. If correct, this would make high-fidelity garment animation practical on smartphones and cut the cost of detailed cloth content.","feed_headline":"Lightweight network adds high-res wrinkles to low-res cloth sims","feed_subtitle":"Trained on five outfits, it generalizes to unseen motions, body shapes, and garments for real-time virtual try-on.","key_machinery":"The central object is the per-triangle implicit wrinkle field built by a hyper-network: for each coarse mesh triangle, the hyper-network takes the graph-learned super-resolution features at the triangle's three vertices and outputs the weights of a small MLP, whose Gabor-wavelet ('WIRE') activation is chosen to represent high-frequency detail; that MLP maps barycentric coordinates to a wrinkle residual vector. The mesh-graph-net that supplies the features, and the decoder that corrects the coarse shape, are supporting mechanisms that make the field reflect dynamics and garment-body interaction.","core_discovery":"GDSR reconstructs high-resolution garment geometry from low-resolution simulation by predicting two things at once: a corrected coarse garment shape and a per-triangle residual field that adds high-frequency wrinkles. The super-resolution features come from a mesh graph network operating on the coarse garment vertices, edges, and inter-layer edges, augmented with garment-body interaction vectors. A decoder maps part of these features to vertex displacements that fix the coarse shape, and a hyper-network maps the remaining features to the weights of a small implicit MLP per triangle; evaluating that MLP at barycentric coordinates yields wrinkle residuals that are added to the up-sampled corrected mesh. The architecture's key bet is that these local, topology-agnostic features and continuous fields let the same trained network handle garment types, body shapes, and motions it never saw, while the roll-out loop keeps details temporally coherent.","pith_inferences":["If the learned detail function is truly topology-agnostic, the same recipe (graph features plus hyper-network implicit fields) could be applied to other deformable surfaces such as hair, skin, or soft tissue; the paper does not test this.","A natural compression application follows implicitly: store only the low-resolution simulation and reconstruct high-resolution details at playback, cutting storage and transmission for character animation; this is not stated as a result.","Because the paper trains on one fabric and lists fabric friction as a failure mode, a material-conditioned variant (e.g., conditioning the hyper-network on fabric parameters) would be a direct extension to test whether wrinkle residuals are material-agnostic.","The quantitative claims rely on SSIM of normal maps over 300 frames on unseen motion; adding the same metric for unseen body shapes and garment types would directly probe the generalization claim's boundaries."],"forward_implications":["High-resolution garment animation could run on low-budget devices: simulate a coarse mesh and synthesize details with the 65MB network at about 0.115 s per frame (including coarse simulation), instead of running an expensive fine simulation.","The continuous per-triangle field decouples output resolution from the network, so one trained model can produce geometry at any target mesh density.","Roll-out prediction makes the network usable autoregressively; stretching and shearing energies stay within a small range over 1500 frames, suggesting long sequences do not drift.","The generalization results imply that new garments can be added to a pipeline with only a low-resolution garment simulation and no retraining, as long as the coarse topology semantics match.","Quantitatively, on an unseen motion the method beats the DDE and PhysGraph baselines on normal-map SSIM for three garment types, supporting the claim that it recovers high-frequency detail structure."],"supporting_citations":[{"why":"Supplies the message-passing graph network that the Mesh-Graph-Net is built on, used to propagate features over the coarse garment mesh.","marker":"[Pfaff et al. 2020]"},{"why":"Provides hierarchical message passing that accelerates feature propagation and is adopted to extend the graph process.","marker":"[Grigorev et al. 2023]"},{"why":"Introduces the WIRE/Gabor wavelet activation used in the implicit wrinkle MLP to represent high-frequency details.","marker":"[Saragadam et al. 2023]"},{"why":"PhysGraph, a state-of-the-art mesh-based graph baseline that the paper compares against for detail synthesis.","marker":"[Halimi et al. 2023]"},{"why":"DDE, the normal-map style-transfer baseline the method improves upon and the main alternative approach.","marker":"[Zhang et al. 2021a]"},{"why":"SMPL body model used to extract body shapes and generate motions for the training and test data.","marker":"[Loper et al. 2015]"},{"why":"Defines the deformation mapping matrix used in the coarse-correction loss to preserve stretching and shearing behavior.","marker":"[Baraff and Witkin 2023]"}],"fun_headline_variants":["Lightweight AI adds lifelike wrinkles to low-res clothing","Smart cloth upscaling: coarse sims get fine wrinkles fast","Generalizable network boosts cloth sims into high-res detail","One net, any motion: low-res cloth to high-res wrinkles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that a mapping from low-resolution to high-resolution garment geometry can be learned from a small, homogeneous training set (two dance sequences, five outfits, one body shape, one silk fabric) and still transfers to unseen motions, body shapes, and garment types.","fun_headline_variants_meta":{"raw":{"variants":["Lightweight AI adds lifelike wrinkles to low-res clothing","Smart cloth upscaling: coarse sims get fine wrinkles fast","Generalizable network boosts cloth sims into high-res detail","One net, any motion: low-res cloth to high-res wrinkles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1667,"prompt_tokens":956,"completion_tokens":711,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":640}},"tokens_in":572,"tokens_out":711,"duration_ms":7828,"temperature":1.0,"reasoning_tokens":640,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:48:55.866265+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released network, without fine-tuning, on a low-resolution simulation of a garment made of a very different fabric (e.g., stiff denim) in an unseen motion, and compare the normal maps of the synthesized high-resolution geometry against a full high-resolution simulation; if the SSIM drops to the level of simple up-sampling from the coarse mesh, the learned mapping does not transfer across materials, which the paper's claims require it to do.","supporting_citations":[],"review_version":1}