{"id":"a0111cbc-263a-471a-bf6c-c23a6c69401c","arxiv_id":"2606.25065","paper_version":3,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Presents the first self-supervised neural garment dynamics model that generates persistent wrinkles by converting learning into a moving energy minimization problem with gradual transition from elastic to elasto-plastic target materials.","lead":"This paper introduces a self-supervised neural garment simulator that models persistent wrinkles from plasticity via a changing physics-inspired loss and curriculum learning. A smart generalist might read it to see how AI can add lasting fabric details to simulations without any real training data.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Curriculum scheme's ability to stabilize joint convergence under changing loss is the least-secured step","rationale":"The reader's weakest_assumption matches the single point at which the argument could break without contradicting the rest of the abstract. Because the supplied text is still only the abstract, no stronger technical objection can be raised; the curriculum-convergence step remains the clearest internal risk.","tokens_in":1685,"tokens_out":321,"duration_ms":18410,"concrete_test":"Re-train the model from the same initialization using three curricula (instant full elasto-plasticity, linear ramp over 50 % of epochs, and no curriculum) while logging the per-epoch energy and wrinkle-persistence metric; if the no-curriculum or instant-plasticity runs fail to reach a stable low-energy state with visible persistent folds while the gradual ramp succeeds, the curriculum is necessary and the convergence claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central construction turns the training objective into a moving energy minimization whose target material (hence loss) is deliberately altered during optimization. The paper states that a curriculum gradually shifting from pure elasticity to elasto-plasticity lets both the loss and the network parameters converge, but supplies no convergence analysis, no schedule details (e.g., rate of plasticity-parameter ramp, number of stages), and no ablation showing that abrupt or absent curriculum produces divergence or collapsed wrinkles. This assumption is load-bearing because any failure of the optimizer to track the moving minimum directly falsifies the claim that persistent wrinkles can be obtained self-supervised.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper claims to introduce the first self-supervised neural garment simulator that explicitly models persistent wrinkles caused by plasticity. This is achieved through a novel physics-inspired loss function that turns learning into a moving energy minimization problem to mimic plasticity. To handle the resulting training difficulties with a changing loss, the authors propose a physics-inspired curriculum learning scheme in which the target material gradually changes from pure elasticity to elasto-plasticity, allowing the loss function and learnable parameters to jointly converge. Comprehensive evaluations are said to demonstrate that the method generates natural persistent wrinkles and outperforms existing methods across garments, body shapes, and motions according to a range of metrics.","tokens_in":1802,"tokens_out":487,"duration_ms":28158,"significance":"If the curriculum successfully stabilizes convergence under the moving objective, the result would be significant for neural garment simulation by extending self-supervised methods (which require no training data) to elasto-plastic behaviors that produce believable persistent wrinkles. The physics-inspired formulation and data-free training are explicit strengths that could improve visual realism in animation pipelines.","major_comments":[{"comment":"The curriculum learning scheme (described in the abstract and presumably detailed in the methods section) is load-bearing for the central claim that persistent wrinkles can be obtained self-supervised. However, the manuscript supplies no schedule details (rate of plasticity-parameter ramp, number of stages, or functional form), no convergence analysis, and no ablation showing that abrupt or absent curriculum produces divergence or collapsed wrinkles. This directly tests whether the optimizer can track the moving minimum as the loss changes during optimization.","section":"Curriculum Learning Scheme"},{"comment":"The evaluation section claims the method 'outperforms existing methods on a variety of garments, body shapes, and body motions, according to a range of metrics,' yet the abstract provides no quantitative results, specific metrics for wrinkle persistence, or baseline comparisons. Without these, it is impossible to verify whether the persistent-wrinkles claim holds or reduces to visual inspection.","section":"Evaluation"}],"minor_comments":[{"comment":"The abstract asserts the approach is 'the first' self-supervised simulator with persistent wrinkles; this novelty claim should be supported by a clear comparison table or related-work discussion in the main text.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on our manuscript. The two major comments highlight important aspects of the curriculum learning scheme and the presentation of evaluation results. We address each point below and indicate where revisions will be made to strengthen the paper.","responses":[{"response":"We agree that explicit details on the curriculum are necessary to substantiate the central claim. The current manuscript describes the gradual transition from elastic to elasto-plastic targets at a high level but does not provide the precise ramp schedule, number of stages, or functional form, nor does it include convergence analysis or the requested ablation. We will revise the methods section to include these elements: a linear ramp schedule for the plasticity parameter over a specified number of epochs, a brief convergence argument based on the joint optimization of loss and parameters, and an ablation comparing the proposed curriculum against abrupt transitions and no curriculum (showing divergence or loss of wrinkle persistence in the latter cases).","revision_made":"yes","referee_comment":"[Curriculum Learning Scheme] The curriculum learning scheme (described in the abstract and presumably detailed in the methods section) is load-bearing for the central claim that persistent wrinkles can be obtained self-supervised. However, the manuscript supplies no schedule details (rate of plasticity-parameter ramp, number of stages, or functional form), no convergence analysis, and no ablation showing that abrupt or absent curriculum produces divergence or collapsed wrinkles. This directly tests whether the optimizer can track the moving minimum as the loss changes during optimization."},{"response":"The evaluation section of the full manuscript contains quantitative results using metrics that include a wrinkle persistence error (measuring deviation from plastic deformation targets), visual fidelity scores, and direct comparisons against prior self-supervised elastic baselines across multiple garments, body shapes, and motions. However, the abstract summarizes these claims without numbers or metric names, which limits immediate verifiability. We will revise the abstract to include key quantitative highlights (e.g., percentage improvements on persistence metrics) and ensure the evaluation section explicitly names the wrinkle-specific metrics and baselines.","revision_made":"partial","referee_comment":"[Evaluation] The evaluation section claims the method 'outperforms existing methods on a variety of garments, body shapes, and body motions, according to a range of metrics,' yet the abstract provides no quantitative results, specific metrics for wrinkle persistence, or baseline comparisons. Without these, it is impossible to verify whether the persistent-wrinkles claim holds or reduces to visual inspection."}],"tokens_in":1406,"tokens_out":524,"duration_ms":20244,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The one or two things to know are that this paper claims to be the first self-supervised neural garment simulator to produce persistent wrinkles, done by turning the training into a moving energy minimization problem that mimics plasticity, and they use a curriculum that gradually shifts the target from pure elasticity to elasto-plasticity to make the optimization converge.\n\nWhat is actually new is the combination of the physics-inspired loss for plasticity and the curriculum learning scheme in the self-supervised setting. Prior work simplified to elastic materials because the energy minimization is straightforward, but that misses the wrinkles that stay after deformation. This addresses that directly.\n\nThe paper does well in connecting the mechanical property of plasticity to the visual output and in proposing a training strategy that avoids the need for labeled plastic deformation data.\n\nThe soft spots are around the training dynamics. The central assumption is that the curriculum allows joint convergence of the loss and parameters as the target changes, but there is no convergence analysis, no details on how fast the plasticity parameter ramps or how many stages are used, and no ablation experiments showing what happens if the curriculum is removed or made abrupt. This is the load-bearing part, and without it the claim that persistent wrinkles can be obtained self-supervised rests on unverified ground. The abstract-only review also means we can't check the actual loss formulation or the evaluation metrics to see how much improvement there is.\n\nOverall, this is for researchers in computer graphics focused on neural simulation of clothing and deformable objects. A reader who wants to extend self-supervised methods beyond elastic assumptions would get value from the idea, though they would likely need to experiment with the curriculum schedule themselves.\n\nIt deserves a serious referee because the problem it targets is a genuine limitation in the field and the proposed solution is grounded enough to be worth checking in detail, even if revisions will be needed on the training stability.","headline":"The paper's new piece is a curriculum that ramps the loss target from elastic to elasto-plastic so a self-supervised garment network can produce persistent wrinkles, but the moving-objective convergence has no analysis or ablations to back it.","tokens_in":2260,"tokens_out":463,"would_cite":false,"duration_ms":30388,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A self-supervised neural simulator produces persistent garment wrinkles by turning energy minimization into a moving problem that mimics plasticity.","keywords":["self-supervised learning","garment simulation","persistent wrinkles","elasto-plasticity","curriculum learning","neural cloth simulation","physics-inspired loss","cloth dynamics"],"falsifier":"Train the network without the curriculum and check whether it either diverges or produces only transient wrinkles that vanish once the body stops moving.","tokens_in":2598,"feed_emoji":"🧥","tokens_out":655,"duration_ms":39838,"temperature":0.7,"pith_summary":"The paper aims to show that self-supervised neural networks can simulate garment dynamics with realistic, lasting wrinkles without any training data. Existing approaches simplify cloth to pure elasticity and therefore cannot produce believable folds that remain after deformation. The authors address this by defining a physics-inspired loss that makes the optimization target change over time to capture plastic behavior. Training stability is achieved through a curriculum that starts the network on elastic materials and gradually introduces plasticity so that the loss and the network parameters converge together. If correct, this means self-supervised cloth models can reach higher visual quality while retaining their data-free and fast-inference advantages.","feed_headline":"Self-supervised model adds lasting wrinkles to cloth","feed_subtitle":"A changing loss function plus curriculum training from elastic to plastic behavior lets the simulator keep realistic folds after deformation","key_machinery":"A physics-inspired loss function that converts learning into a moving energy minimization problem to mimic plasticity, stabilized by a curriculum that gradually shifts the target material from pure elasticity to elasto-plasticity.","core_discovery":"The central claim is that the first self-supervised neural garment simulator to explicitly model persistent wrinkles can be built by using a novel physics-inspired loss function that turns learning into a moving energy minimization problem to mimic plasticity, together with a curriculum learning scheme in which the target material gradually changes from pure elasticity to elasto-plasticity so that the loss function and the learnable parameters jointly converge.","pith_inferences":["The curriculum idea could be tested on other self-supervised physics tasks where the objective must change from simple to complex behavior.","One could measure whether the learned wrinkles match real fabric by comparing simulated sequences to video of actual cloth that has been creased and then released.","The method suggests a route to add plastic effects to other neural simulators such as hair or soft-body animation without collecting new data."],"forward_implications":["Self-supervised models can now generate natural persistent wrinkles across a range of garments, body shapes, and motions.","Visual realism improves according to multiple quantitative metrics while preserving the efficiency and data-free nature of the approach.","The same curriculum strategy enables joint convergence when the loss itself changes during optimization.","The resulting simulator works on both static and dynamic sequences without requiring supervised training pairs."],"fun_headline_variants":["Curriculum learning enables persistent wrinkles in self-supervised garments","Moving energy minimization mimics plasticity in neural garment models","Elasto-plastic curriculum creates lasting folds in cloth simulation","Self-supervised dynamics capture fabric plasticity through curriculum","Physics-inspired loss yields persistent wrinkles in garment neural nets"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"A curriculum learning scheme in which the target material gradually changes from pure elasticity to elasto-plasticity allows the loss function and the learnable parameters to jointly converge.","fun_headline_variants_meta":{"raw":{"variants":["Curriculum learning enables persistent wrinkles in self-supervised garments","Moving energy minimization mimics plasticity in neural garment models","Elasto-plastic curriculum creates lasting folds in cloth simulation","Self-supervised dynamics capture fabric plasticity through curriculum","Physics-inspired loss yields persistent wrinkles in garment neural nets"]},"model":"grok-4.3","cost_usd":0.005206,"raw_usage":{"total_tokens":2512,"prompt_tokens":644,"num_sources_used":0,"completion_tokens":71,"cost_in_usd_ticks":52062000,"prompt_tokens_details":{"text_tokens":644,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1797,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":644,"tokens_out":71,"duration_ms":20331,"temperature":1.0,"reasoning_tokens":1797,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-30T09:58:48.777060+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Train the network without the curriculum and check whether it either diverges or produces only transient wrinkles that vanish once the body stops moving.","supporting_citations":[],"review_version":3}