{"id":"a0dd9bdd-06fa-465f-99b7-3af5136da718","arxiv_id":"2411.16216","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SMGDiff generates real-time, user-controllable soccer animations with an autoregressive diffusion model plus a contact guidance module, trained on a new 1.08-million-frame soccer motion dataset.","lead":"This paper presents a two-stage AI system that turns simple user controls, such as direction and skill choice, into realistic soccer player animations in real time, and introduces a new dataset of over 1.08 million frames of soccer motion. It matters because video games and VR/AR need diverse, responsive character animation, and soccer has been underserved by data-driven motion generation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim lacks comparison against the closest existing method, CAMDM, on which SMGDiff is directly built; Tab. 1 only includes older non-diffusion controllers.","rationale":"The reader's weakest assumption focuses on the contact-guidance threshold (Eq. 10) and its potential to distort motions. That is a legitimate concern about a novel component, but it is not the single most load-bearing risk to the central claim. The central claim is about outperforming existing methods; the paper's own related work identifies CAMDM as the closest real-time diffusion-based character controller, and SMGDiff adopts CAMDM's core mechanisms (autoregressive diffusion, trajectory blending, skill-accuracy evaluation). Omitting CAMDM from the comparison leaves the SOTA claim unsupported against the strongest relevant competitor. A conditional verdict is appropriate: the method is coherent and the dataset is a contribution, but the paper must add the missing baseline (and ideally release code/data) before the headline claim can be accepted. I disagree with the reader in the sense that I would prioritize the evaluation gap over the threshold issue, though both should be addressed.","tokens_in":14482,"tokens_out":2891,"duration_ms":42286,"concrete_test":"Train CAMDM from reference [5] on Soccer-X using the same 9:1 train/test split, the same per-frame representation (human state h, ball state b, contact label c), the same condition inputs (skill label S, past motion X_P, future trajectory T_F), and the same 8 denoise steps; then evaluate with the exact Tab. 1 metrics. If SMGDiff still improves FID and Skill Acc. over CAMDM by comparable margins at similar runtime, the SOTA claim is supported; if not, it should be downgraded.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim—'significantly outperforms existing methods'—is established only against LMP, MANN-DP, and CM, all non-diffusion real-time controllers from before 2024. SMGDiff explicitly builds on CAMDM [5] for its autoregressive diffusion backbone, trajectory blending, and skill-accuracy evaluation protocol (Secs. 3.1, 3.2, 5). Yet CAMDM is never included as a baseline in Tab. 1. This is the most load-bearing gap: without a CAMDM baseline trained on Soccer-X with the same representation and condition inputs, the reported FID 0.1813 and Skill Acc. 93.3% do not establish superiority over the actual state of the art in real-time controllable diffusion-based character motion. The chosen baselines are weaker and not designed for soccer, so 'same motion representation' does not compensate for the architectural mismatch. If CAMDM, adapted to the same task, achieves comparable or better metrics, the paper's central SOTA claim fails, even though the individual modules (trajectory generation, contact guidance) may still work. This concern is about evaluation completeness, not internal inconsistency: the method itself appears coherent, and the contact-threshold issue identified by the reader is real but secondary, since it affects one module and is partially supported by the CGM ablation.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"SMGDiff is a two-stage framework for real-time, user-controllable soccer motion generation. The first stage converts coarse user controls (skill label, target point, and past trajectory) into smooth global trajectories using a single-step diffusion model with temporal blending borrowed from CAMDM. The second stage trains a transformer-based autoregressive diffusion model to generate future human, ball, and contact states conditioned on trajectory, skill, and past motion; during inference, a contact guidance module applies a custom loss in the final denoising steps to improve ball-foot contact. The authors contribute Soccer-X, a 1.08M-frame motion capture dataset with six soccer skill categories, and report experiments on their own test set against LMP, MANN-DP, and CM, achieving FID 0.1813 and Skill Accuracy 93.3%, with ablations and runtime analysis showing that 8 denoise steps and contact guidance on the last two steps are preferred.","tokens_in":14790,"tokens_out":6224,"duration_ms":198241,"significance":"If validated, the paper would make a useful applied contribution: a large soccer motion dataset, a real-time interactive pipeline, and a concrete way to inject ball-foot contact into autoregressive diffusion. The dataset size and the live Unity demo are tangible strengths, and the use of standard diffusion losses and a single-step trajectory generator is technically sound. The core concern is that the central SOTA claim is not yet supported: the closest diffusion-based real-time controller, CAMDM, is absent from the comparison, and the contact-guidance module's target is inferred from the very motion it is supposed to correct. These are fixable with additional experiments and validation, but they are load-bearing.","major_comments":[{"comment":"The central claim that SMGDiff 'significantly outperforms existing methods' is not established because CAMDM [5], the method on which SMGDiff directly builds, is never included as a baseline. SMGDiff's second stage is described as 'a transformer-based autoregressive diffusion model [5]' (Sec. 3.2), its trajectory blending uses CAMDM's HFTE strategy (Sec. 3.1), and its skill-accuracy evaluation follows [5] (Sec. 5). Yet Table 1 compares only LMP, MANN-DP, and CM, all non-diffusion controllers predating CAMDM. Since CAMDM is the relevant state of the art for real-time controllable diffusion-based character motion, the reported FID 0.1813 and Skill Acc. 93.3% do not demonstrate superiority over the architecture on which SMGDiff is built. A CAMDM baseline trained on Soccer-X with the same representation and condition inputs must be added, or the 'outperforms existing methods' claim must be scaled back.","section":"Sec. 5.1, Table 1"},{"comment":"The contact guidance target is self-referential and unvalidated. Eq. (10) infers the contact indicator ĉb from the generated ball acceleration exceeding a hand-set threshold τa=2 m/s², and Eq. (12) then uses that inferred contact to guide the same generated motion. Because the guidance signal is derived from the variable being optimized, it cannot provide independent evidence of contact; if the acceleration heuristic is inaccurate, guidance will actively distort motion around non-contact frames. The dataset already contains ground-truth contact labels c (defined in Sec. 3), so the heuristic can be validated or replaced by a learned contact predictor. The CGM ablation in Sec. 5.2 is therefore inconclusive as evidence of improved physical contact.","section":"Sec. 3.3, Eqs. (10)-(12)"},{"comment":"The empirical support for 'significantly outperforms' is statistically weak. No error bars, confidence intervals, or significance tests are reported; the numbers appear to be single-run point estimates. In addition, evaluation is confined to the authors' own Soccer-X test set, and for the Shoot category the ball trajectories were reconstructed in Unity rather than captured (Sec. 4), which may bias contact-related metrics in that category. The paper should report per-category metrics with variance across seeds, perform a significance test, and quantify the effect of the Unity-based Shoot reconstruction on contact accuracy.","section":"Sec. 5, Tables 1-4"}],"minor_comments":[{"comment":"The text attributes the simple objective to 'Go et al. [15]', but reference [15] is Ho et al.; please correct the citation.","section":"Sec. 3.2"},{"comment":"The Diversity and FID values for 'Ours' differ substantially between Table 1 (Div. 0.6177, FID 0.1813) and Table 2 (Div. 2.6925, FID 0.3580) under the same model. The paper should explain why the evaluation protocol differs; as written, the ablation gains are hard to interpret.","section":"Tables 1 and 2"},{"comment":"The contribution bullet says 'Our data and model will be disseminated to the community' and Sec. 4 calls Soccer-X a 'large-scale open-source soccer dataset', but no dataset release URL or licensing information is provided; please include these details.","section":"Sec. 4"},{"comment":"References [15] and [16] are the same paper (Ho et al., DDPM) with different numbering cues; please deduplicate.","section":"References"},{"comment":"The pipeline figure uses 'SMDM' without defining the acronym in the text or caption; please define it at first use.","section":"Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The missing CAMDM baseline is the most important issue for the SOTA claim. If the authors can include that baseline and report error bars or significance tests, the paper could become a solid contribution; otherwise the claims should be substantially softened. I would also encourage the editor to verify that the promised dataset and code release actually materialize, since the dataset is a major claimed contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the Soccer-X dataset is the real contribution; the generation pipeline is competent but the claimed SOTA is not established, because the comparison omits CAMDM, the very method the paper builds on.\n\nWhat's actually new: 1.08M frames of motion-captured soccer data across six skills with ball and contact labels; a joint human+ball+contact representation; a two-stage real-time pipeline with a single-step trajectory diffusion and an autoregressive motion diffusion; and a contact-guidance heuristic that improves ball-foot contact. The runtime engineering is thoughtful—training with 8 denoise steps and restricting guidance to the last two steps is a sensible trade-off, and the ablations give the modules some support.\n\nThe soft spots, in order. First, the load-bearing one: Table 1 compares only against LMP, MANN-DP, and CM—all pre-diffusion, non-diffusion controllers. SMGDiff uses CAMDM's autoregressive diffusion, its trajectory blending, and its evaluation protocol, but CAMDM is never included as a baseline. Without that comparison, the \"significantly outperforms existing methods\" claim is unsupported. It may still be true that SMGDiff is the best soccer-specific controller, but the paper doesn't show it. Second, the evaluation is entirely on their own dataset and no error bars or significance tests are reported. Differences like FID 0.18 vs 0.25 are within a single split and could be noise. The Shoot category's ball trajectories are reconstructed in Unity rather than captured, which weakens that category. Third, the contact-guidance threshold τa is hand-picked and never validated against ground-truth contact labels. The guidance is also self-referential—the contact indicator comes from the model's own predicted ball acceleration (Eq. 10) and then steers that same prediction (Eq. 12). It's a reasonable heuristic and the ablation suggests some benefit, but it is not a validated mechanism. Finally, no code or data is released despite the abstract promising dissemination; for a dataset paper that should be a condition of acceptance. On the plus side, the limitations section honestly names the physics, foot-only, and single-player gaps.\n\nWho this is for: anyone in character animation, especially game/VR controllers, will care about the dataset and the contact heuristic. The paper deserves a serious referee, but it needs to add a CAMDM baseline, report error bars, and release artifacts before I'd trust the SOTA claim.","headline":"The dataset and pipeline are solid contributions, but the SOTA claim rests on a baseline set that omits CAMDM, the method the paper directly builds on.","tokens_in":15305,"tokens_out":3588,"would_cite":true,"duration_ms":29730,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SMGDiff, a two-stage diffusion framework, synthesizes real-time, user-controllable soccer motions and reports better motion quality and condition alignment than the compared controllers.","keywords":["soccer motion generation","diffusion probabilistic models","real-time character control","human-object interaction","ball-foot contact","motion capture dataset","autoregressive diffusion","trajectory conditioning"],"falsifier":"Compute, on the held-out split of Soccer-X, precision and recall of the Eq. (10) contact detector ($\\|b_a\\| > 2\\,\\mathrm{m/s^2}$) against the recorded binary ball-foot contact labels; if many true contacts have acceleration below threshold or many non-contacts exceed it, the contact-guidance signal is driven by an unvalidated proxy, and the FID improvement attributed to the module needs a different explanation.","tokens_in":14298,"feed_emoji":"⚽","tokens_out":12363,"duration_ms":103045,"temperature":0.7,"pith_summary":"SMGDiff's central claim is that realistic soccer animation can be generated in real time from coarse user controls by splitting the task into two stages: a lightweight trajectory generator that turns keyboard-style inputs into a smooth future path, and a transformer-based autoregressive diffusion model that produces the body, ball, and contact states along that path. The paper adds a contact-guidance pass at inference to sharpen ball-foot interactions, and contributes Soccer-X, a mocap dataset of over 1.08 million frames covering six soccer skill categories from 30 players. On its own test set the system reports the lowest FID and foot-sliding distance and the highest skill accuracy among the compared real-time controllers. A sympathetic reader would take the paper as establishing that diffusion-based generation, rather than example-based matching or physics-based control, is a viable route to interactive soccer animation.","feed_headline":"Two-stage diffusion model puts soccer animation under user control","feed_subtitle":"A two-stage diffusion pipeline with contact guidance generates dribbling, tricks, and shooting from coarse controls in real time","key_machinery":"The carrying object is the two-stage pipeline itself. Stage one is a single-step diffusion transformer: conditioned on skill label, target trajectory point, past trajectory, and injected Gaussian noise, it predicts a future ground-plane trajectory, and a temporal blending strategy prevents abrupt turns. Stage two is a transformer-based autoregressive diffusion model that takes as condition the skill label, the refined future trajectory, and past soccer motion, and predicts the future motion state $x_i = \\{h, b, c\\}$ with human pose $h$, ball state $b$, and binary contact labels $c$; it is trained with the denoising diffusion probabilistic model (DDPM) simple objective plus auxiliary position, velocity, and foot-contact losses. The contact-guidance module then detects a ball-foot contact when ball acceleration exceeds $\\tau_a = 2\\,\\mathrm{m/s^2}$ (Eq. 10), picks the lifted foot nearest the ball, and uses a distance-based contact loss with the DSG spherical-Gaussian guidance rule to steer the last two denoising steps. The Soccer-X dataset supplies the learned distribution: 1.08 million frames, six skill classes, 30 players.","core_discovery":"The discovery the paper argues for is a generation recipe: coarse user commands are converted into a global trajectory by a single-step diffusion transformer, and that trajectory conditions an autoregressive diffusion model that generates future soccer motion jointly for the SMPL body, the ball, and binary foot-ground and foot-ball contacts. The ball-foot contact detail is then improved at inference by a contact-loss guidance that detects contact from a threshold on ball acceleration and pulls the nearest lifted foot toward the ball. The paper reports that this recipe yields FID 0.1813, foot-sliding distance 0.8543 m, and skill accuracy 93.3% on its evaluation split, against best baseline values of 0.2494, 1.0678 m, and 73.3%, and that the full pipeline runs at interactive speed with 8 denoising steps.","pith_inferences":["An extension the paper leaves open is to replace the acceleration-threshold contact detector with a learned classifier trained on the same binary contact labels the dataset contains, and to compare FID and skill accuracy under both detectors.","The trajectory-then-motion decomposition is generic enough that the same recipe could transfer to other fast object-interaction animation tasks such as basketball or tennis, where user control is a coarse target rather than a full pose sequence.","Because the contact guidance is enabled only for the final two denoising steps, varying both the number of guided steps and their position in the schedule would isolate whether the reported FID gain comes from contact correctness itself or from the extra gradient signal at low noise.","The reported real-time figure covers model inference and network communication; an end-to-end measurement that includes rendering would quantify the latency a player actually perceives."],"forward_implications":["On the paper's reported evaluation, SMGDiff reaches FID 0.1813, foot sliding 0.8543 m, and skill accuracy 93.3%, while the best compared baseline reaches 0.2494, 1.0678 m, and 73.3%.","The system runs at interactive speeds with 8 denoising steps and contact guidance confined to the last two steps, with a reported inference time of 12 ms for a generated segment.","The Soccer-X dataset, with 1.08 million frames across six skill categories and 30 players, provides a new training and evaluation resource for data-driven soccer motion generation.","Contact guidance specifically reduces missed ball-foot contacts when the ball changes direction, a failure mode visible in the compared controllers."],"supporting_citations":[{"why":"It supplies the autoregressive diffusion architecture and the trajectory-blending strategy that the soccer motion stage adapts for real-time control.","marker":"[5]"},{"why":"It supplies the simple diffusion objective and the auxiliary position, velocity, and foot-contact losses used to train the motion generator.","marker":"[61]"},{"why":"It defines the DDPM forward and reverse diffusion processes and the training objective that both stages build on.","marker":"[15]"},{"why":"It provides the spherical-Gaussian-constrained loss guidance rule used to steer the last denoising steps with the contact loss.","marker":"[67]"},{"why":"It provides the ball-control-weight representation that converts the global ball position into a body-relative ball state.","marker":"[56]"},{"why":"It is one of the real-time character controllers used as a baseline that the method must beat on motion quality.","marker":"[57]"},{"why":"It is the categorical codebook matching controller that sets the strongest baseline FID among the methods compared.","marker":"[60]"},{"why":"It is part of the MANN-DP baseline, combining mode-adaptive networks with deep phase features for the comparison.","marker":"[72]"},{"why":"It defines the SMPL parametric body format used for the human state in the motion representation.","marker":"[39]"}],"fun_headline_variants":["Diffusion model delivers real-time soccer motion control","Two-stage diffusion animates soccer with real-time ball control","Soccer motion from user commands via diffusion and contact feedback","Diffusion-based system turns coarse controls into soccer motion in real time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a ball-foot contact is happening whenever the generated ball's acceleration exceeds 2 m/s², and that using that inferred contact as a target for the guidance loss makes motions more realistic; the paper does not validate this detector against the ground-truth contact labels it collects.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model delivers real-time soccer motion control","Two-stage diffusion animates soccer with real-time ball control","Soccer motion from user commands via diffusion and contact feedback","Diffusion-based system turns coarse controls into soccer motion in real time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001191,"raw_usage":{"total_tokens":4883,"prompt_tokens":882,"completion_tokens":4001,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":3934}},"tokens_in":498,"tokens_out":4001,"duration_ms":28899,"temperature":1.0,"reasoning_tokens":3934,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:21:34.776682+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, on the held-out split of Soccer-X, precision and recall of the Eq. (10) contact detector ($\\|b_a\\| > 2\\,\\mathrm{m/s^2}$) against the recorded binary ball-foot contact labels; if many true contacts have acceleration below threshold or many non-contacts exceed it, the contact-guidance signal is driven by an unvalidated proxy, and the FID improvement attributed to the module needs a different explanation.","supporting_citations":[{"cited_title":"Taming diffusion probabilistic models for character control","cited_arxiv_id":null,"evidence_quote":"It supplies the autoregressive diffusion architecture and the trajectory-blending strategy that the soccer motion stage adapts for real-time control."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the simple diffusion objective and the auxiliary position, velocity, and foot-contact losses used to train the motion generator."},{"cited_title":"Denoising diffu- sion probabilistic models","cited_arxiv_id":null,"evidence_quote":"It defines the DDPM forward and reverse diffusion processes and the training objective that both stages build on."},{"cited_title":"Guidance with spherical gaussian con- straint for conditional diffusion","cited_arxiv_id":null,"evidence_quote":"It provides the spherical-Gaussian-constrained loss guidance rule used to steer the last denoising steps with the contact loss."},{"cited_title":"Local motion phases for learning multi-contact character movements","cited_arxiv_id":null,"evidence_quote":"It provides the ball-control-weight representation that converts the global ball position into a body-relative ball state."},{"cited_title":"Local motion phases for learning multi-contact character movements","cited_arxiv_id":null,"evidence_quote":"It is one of the real-time character controllers used as a baseline that the method must beat on motion quality."},{"cited_title":"Categorical codebook matching for embodied character controllers","cited_arxiv_id":null,"evidence_quote":"It is the categorical codebook matching controller that sets the strongest baseline FID among the methods compared."},{"cited_title":"Mode-adaptive neural networks for quadruped motion control","cited_arxiv_id":null,"evidence_quote":"It is part of the MANN-DP baseline, combining mode-adaptive networks with deep phase features for the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It defines the SMPL parametric body format used for the human state in the motion representation."}],"review_version":1}