{"id":"194bc45f-526f-418e-96de-5def90aff4ea","arxiv_id":"2507.19292","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"PINO sequentially composes pairwise interactions and optimizes diffusion noise with overlap and control penalties to generate customizable, artifact-reduced group motion for arbitrary group sizes.","lead":"This paper introduces a training-free method that generates motion for groups of any size by composing pairwise interactions from a pretrained two-person diffusion model, using inference-time noise optimization with physics penalties to prevent collisions and enable user controls. It is relevant to animation, gaming, and robotics, where controllable multi-character motion generation is needed without retraining.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pairwise-decomposition assumption is unvalidated at the group level: the only semantic evaluation decomposes every multi-person sequence into pairs with the pivot, so failures of coordinated group dynamics are invisible to the metrics.","rationale":"The reader identified the same weakest assumption (pairwise decomposition fails for coordinated behaviors), and the paper itself acknowledges it in Sec. 6. I agree, but would sharpen it: the problem is not just theoretical; the evaluation actively prevents detection of the failure. The multi-person semantic numbers in Table 4 are per-pair FIDs with person 1 as pivot, so they are consistent by construction with the pairwise-generation pipeline. No experiment measures group-level semantic fidelity. Since the method's novelty is precisely the pairwise-decomposition strategy, validating that strategy on full-group tasks is essential before the \"arbitrary group\" claim can be accepted. The other concerns (missing error bars, foot-skating degradation) are secondary. The verdict remains CONDITIONAL: the paper is promising and the physics-based penalties work as designed, but the central generalization claim needs group-level evidence.","tokens_in":15209,"tokens_out":4550,"duration_ms":47076,"concrete_test":"Run PINO on a test set of group actions with genuine three-body coordination (e.g., \"three people hug\", \"person A pushes B into C\", \"two people carry a third\") and evaluate with (1) a full-group semantic metric, such as FID or CLIP-style retrieval on the complete multi-person sequence against a multi-person dataset, and (2) human raters asked whether the intended group action is recognizable. Also record whether the pivot character's motion changes when the new character makes physical contact; if the reference stays fixed under contact, the group is non-reactive. If full-group scores drop sharply while pairwise FID stays competitive, the pairwise-decomposition assumption is the bottleneck.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that arbitrary-sized group interactions can be generated by sequential pairwise composition (Abstract, Sec. 4.2). For this to hold, any target group behavior must be representable as a chain of pairwise interactions with a fixed pivot. The paper's only semantic evaluation of multi-person output (Table 4) decomposes each sequence into pairs (1,2), (1,3), ... and computes FID against two-person ground truth; it never measures whether the whole group performs a coherent group-level action. This metric cannot detect failures of high-order coordination. The limitation in Sec. 6 explicitly concedes that \"a person pushing another person into a third person will not be modeled by the underlying two-person model.\" That is a direct consequence of Eqs. 6-7: when adding person p, the reference character's motion is fixed as a condition (xhat^{k_p}_0) and is never updated in response to the new character. Any interaction that requires the reference to react physically to the newcomer (pushes, lifts, coordinated falls, group embraces) is therefore outside what the method can express, regardless of how well the overlap penalty works. The paper's claim of \"arbitrary-sized group\" interactions is thus unsupported for a large class of common group behaviors, and the evaluation design hides this blind spot.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PINO, a training-free framework that extends a pretrained two-person motion-diffusion model (InterGen) to multi-person group generation by sequentially composing pairwise interactions. Each new character is generated while conditioning on a fixed reference character, and inference-time noise optimization applies penalty terms for overlap avoidance and user-controllable spatial/temporal constraints (root position, region, orientation, relative distance). Motion inpainting is added to extend sequences in time, and prompt switching between pairs enables changing roles during long sequences. Experiments report lower overlap and penetration volume on two-person, multi-person, and long-duration tasks, an ablation of the penalty terms, a small user study, and comparisons with InterGen, FreeMotion, and in2IN.","tokens_in":15457,"tokens_out":5775,"duration_ms":60234,"significance":"If the claims hold, PINO is a practically useful, training-free way to obtain customizable multi-person motion without dedicated multi-person datasets, which is valuable for animation and content generation. The paper's strengths include using a publicly available pretrained model, providing fairly detailed supplementary implementation details, reporting a user study, and demonstrating that the proposed framework also works with another two-person base model (in2IN). The significance is tempered, however, by two issues: the headline overlap metric is a strict relaxation of the optimized penalty, so the main quantitative success is partly by construction, and the semantic evaluation of multi-person output decomposes everything into pairs sharing a single pivot, which cannot validate the paper's central claim about coherent group interactions of arbitrary size.","major_comments":[{"comment":"The central claim of generating group interactions of arbitrary size is not validated at the group level. Equations (6)-(7) generate each new character conditioned on the fixed reference motion xhat^{k_p}_0, and that reference is never updated when the new character is added. As Section 6 concedes, a person pushing another person into a third person is therefore outside the model. The semantic evaluation in Table 4 only decomposes each multi-person sequence into the pairs (1,2), ..., (1,5), all sharing person 1 as pivot, and computes FID against two-person ground truth; this protocol cannot detect whether the generated characters perform a coherent group-level action, and for FreeMotion it likely understates that baseline's group-level conditioning. Please add a group-level semantic evaluation (for example, recognition of multi-person actions, evaluation on non-pivot pairings, or explicit tests of synchronized group behaviors such as group lifts or cascading falls) and temper the abstract and conclusion claims accordingly.","section":"Sec. 4.2, Eqs. (6)-(7); Sec. 5.2, Table 4; Sec. 6"},{"comment":"The overlap numbers that form the main quantitative evidence are by construction determined by the optimization objective. Loverlap uses a root-distance threshold delta_overlap=30cm (Supplementary B.1), while the Overlap metric declares a failure when root positions are closer than 25cm (Supplementary B.2). The penalty is thus a stricter version of the metric, so Overlap values near zero in Tables 2-4 certify that the optimizer minimized its own loss, not that the framework produces physically plausible interactions in an independent sense. The remaining non-semantic metrics are more informative; note that PINO's Foot Skate and Max Acc in Table 4 are consistently worse at (1,4) and (1,5) than the InterGen baseline, a degradation that receives no discussion. Reporting a metric that is a strict relaxation of the objective, and discussing the metrics that do move unfavorably, would put the physical-plausibility claim on firmer ground.","section":"Sec. 4.3, Eq. (8); Supplementary B.1-B.2; Tables 2-4"},{"comment":"The comparison with InterGen and FreeMotion is not matched for inference-time optimization: PINO performs up to 100 gradient steps of noise optimization per added character, while the baselines are run without the additional optimization loop, with PINO taking about one minute per person for the overlap penalty alone (Supplementary C.4). The large overlap reduction may therefore reflect the optimization budget rather than the pairwise-composition design. A fairer assessment would include a compute-matched variant (for example, optimizing the baselines' latent initialization with the same loss, or an ablation of PINO without the overlap penalty) so the reader can separate the contribution of sequential composition from that of noise optimization.","section":"Sec. 5.2, Tables 3-4; Supplementary B.1 and C.4"}],"minor_comments":[{"comment":"The notation for the number of individuals is inconsistent: Section 4.2 writes x3,...,xN while Algorithm 1 uses M, and N is already used for the number of frames in the motion representation. Please unify these symbols.","section":"Sec. 4.2 and Algorithm 1"},{"comment":"Line 12 of Algorithm 1 calls Gθ(...) whereas Eq. (7) uses Gmask_theta, and the argument order in the pseudocode appears different from the convention used in Eqs. (3)-(7). Please make the pseudocode consistent with the equations.","section":"Algorithm 1, line 12 and Eq. (7)"},{"comment":"The bar labels in the user-study figure run together as \"InterGenFreeMotionPINO\", and the reported scores are not accompanied by error bars or significance tests across the 35 participants. Please report confidence intervals or a statistical comparison.","section":"Figure B (Supplementary C.3)"},{"comment":"The overlap metric threshold of 25 cm and the penalty threshold of 30 cm appear only in the supplementary material; since these values directly determine the headline results, they should be stated in the main text.","section":"Supplementary B.1-B.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid engineering contribution with a clearly identified limitation that conflicts with the strength of the abstract's \"arbitrary-sized group interactions\" claim. The revision path is feasible: add a group-level semantic evaluation, reframe the claims to match the pairwise-composition scope, and address the overlap-metric circularity. I see no concerns about citation patterns or misconduct."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a read. PINO composes an arbitrary-size group by generating pairwise interactions sequentially, using a pre-trained two-person diffusion model and optimizing the initial noise with physics penalties (overlap, position, orientation, region). The combination of per-pair prompts with noise optimization is new, as far as I know, and it ships some nice engineering: fixed reference plus masked denoising, inpainting for long sequences, and an ablation that shows each penalty does what it claims (Table 5).\n\nThe paper is honest about the main limitation: Section 6 says a person pushing another into a third won't be modeled because the underlying two-person model can't express it. That is not a small caveat. It follows directly from Eqs. 6-7: the reference character is fixed during optimization, so nothing the new person does can cause the reference to react. That means 'arbitrary-sized group interactions' overclaims for any behavior requiring physical reaction chains or synchronized multi-person coordination. The abstract and conclusion say 'arbitrary-sized' without this qualification, and the semantic evaluation (Table 4) decomposes every sequence into pairs with the pivot, so group-level coherence is never measured. The stress-test note is right on both counts.\n\nWhat is not in the stress-test note but worth saying: the ablation is the strongest part. The progressive penalty study shows real control, not just overlap reduction. The overlap improvement is to some extent circular because that is the optimized objective, but the paper also reports FID, foot-skate, and max acceleration, which don't improve but also don't collapse, which is reassuring. However, Foot Skate and Max Acc degrade monotonically as group size grows (Table 4) without discussion. Minor point: no error bars anywhere, and the multi-person evaluation is small (96 sequences for three people).\n\nOverall: the writing is clear, the method is coherent, and the limitations are honestly acknowledged in one place even if the headline overclaims. The gaps, group-level evaluation, reactive interactions, and discussion of degradation, are addressable. I would send this to peer review; it deserves a serious referee, but I would ask for a softened claim or a group-level semantic metric (or both) before accepting.","headline":"A solid training-free composition method whose headline claim ('arbitrary-sized groups') outruns its evaluation, but the core idea and ablation are worth referee time.","tokens_in":16015,"tokens_out":2344,"would_cite":false,"duration_ms":23931,"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":"This paper proposes PINO, a training-free method that generates arbitrary-sized group motion by sequentially composing pairwise interactions from a pretrained two-person diffusion model, using noise-optimization penalties to keep bodies…","keywords":["human motion generation","multi-person interaction","text-conditioned diffusion","noise optimization","motion inpainting","physics-based penalties","training-free inference"],"falsifier":"Run the method on a set of genuinely three-person-coordinated actions, such as a group hug, a two-person lift of a third, or a chain-reaction push, with prompts describing the full group behavior, and measure whether the resulting motion contains the required simultaneous three-body contact; if the composed pairwise motions systematically lose the third-person contact or produce physically impossible configurations, the pairwise-decomposition claim is falsified for those behaviors.","tokens_in":15004,"feed_emoji":"🕺","tokens_out":8029,"duration_ms":69972,"temperature":0.7,"pith_summary":"PINO tackles the problem of generating realistic motion for groups of arbitrarily many people without collecting multi-person training data. The central proposal is to break any group interaction into a sequence of pairwise interactions, each described by its own text prompt, and to build the group by adding one person at a time to a pretrained two-person interaction diffusion model. Because sequentially composed characters tend to overlap or penetrate each other, the method optimizes the diffusion model's initial noise against differentiable physics-based penalties that enforce minimum distances and user-specified spatial, orientational, and positional constraints. The same inpainting mechanism extends sequences in time, so long-duration motions with changing roles can be produced. Evaluations on up to five-person scenes show that the approach nearly eliminates overlap while keeping semantic fidelity comparable to the base model.","feed_headline":"A two-person diffusion model can drive arbitrary-sized group motion","feed_subtitle":"Pairwise prompts plus overlap penalties enable spatial control with no retraining.","key_machinery":"The engine of the method is Person-Interaction Noise Optimization (PINO), defined as the minimization of differentiable penalty losses over the initial latent noise of a masked two-person diffusion model while one character's motion is fixed as the reference. The masked model $G^{\\text{mask}}_\\theta$ denoises only the new target character's noise, interleaving the reference character's already-generated frames at every step, so the new motion is anchored to the existing interaction. The loss $L = L_{\\text{overlap}} + L_{\\text{control}}$ penalizes frames in which two root positions come closer than a threshold $\\delta$, plus user-specified control terms for root position, movement region, facing orientation, and relative distance, each written with thresholded max functions so that small deviations are free. For long-duration generation, motion inpainting blends a fixed reference window $m \\odot \\hat{x}$ with denoised continuation frames and adds a joint-acceleration boundary penalty for smooth transitions.","core_discovery":"On its own terms, the paper claims that group interactions are not monolithic: they are composites of smaller pairwise interactions linked by shared pivot characters, and therefore a model trained only for two-person interactions can generate coherent motion for groups of arbitrary size. The discovery is the demonstration that composing pairwise generations sequentially, with each new character conditioned on one existing character's frozen motion and guided by its own prompt, yields group motion whose quality does not collapse as people are added, provided the initial noise of the diffusion process is optimized with penalties for root-position overlap and, when needed, for target positions, movement regions, orientations, and relative distances. This noise optimization, not additional training, is what removes body penetration and overlap and what gives users fine-grained control. The paper further shows that the same framework extends motion in time via masked inpainting with a boundary acceleration penalty, allowing long sequences and prompt switching between pairs.","pith_inferences":["Because the penalties act only on joint and root positions and use a single distance threshold, the method implicitly treats bodies as point-like at the root; a natural extension would couple body-thickness buffers or hand-aware losses to catch the residual near-hand penetrations the paper reports.","The same training-free recipe could turn any two-person motion prior into a data-generation tool: run PINO over diverse prompts and pairings, then use the artifact-free outputs to train a feed-forward multi-person model, which would remove the per-person optimization cost at deployment.","The pairwise-decomposition assumption means group behaviors whose dynamics are inherently third-order, such as a lift, a chain-reaction fall, or a three-person synchronized jump, will be under-expressed; testing exactly where the dyadic graph fails would map the boundary of the method.","The control penalties are generic differentiable costs, so in principle the same optimizer could steer motions to satisfy non-human constraints such as camera framing, obstacle maps, or collision with scene geometry, not just person-person spacing."],"forward_implications":["Group size is no longer bounded by training data: the same frozen two-person base model can in principle be extended to any number of participants by repeating the pivot-and-optimize step.","Each pairwise interaction keeps its own text prompt, so mixed-role scenes such as a photographer directing posers or dancers with onlookers can be specified directly instead of being forced into one shared group description.","New control signals, including where someone stands, where they face, how fast they move, and how far apart they are, become inference-time penalty choices, so no per-control retraining is needed.","Long sequences with changing roles over time are reachable by extending motions through inpainting while switching prompts between pairs, as demonstrated by the alternating handshake among three people.","The overlap and penetration artifacts that grow with group size in sequential baselines are largely suppressed while FID, diversity, and R-precision stay close to the underlying two-person model."],"supporting_citations":[{"why":"InterGen: supplies the pretrained shared-weight two-person interaction diffusion model and the InterHuman dataset used for evaluation.","marker":"[19]"},{"why":"DDIM: the denoising sampler used for all generation and for backpropagating penalties through the diffusion process.","marker":"[29]"},{"why":"DNO: establishes that optimizing the initial noise of a motion diffusion model can serve as a universal prior for goal-directed generation.","marker":"[16]"},{"why":"ProgMoGen: provides the differentiable noise-optimization loop that PINO adapts to the multi-person setting, as noted in the supplementary implementation details.","marker":"[20]"},{"why":"FreeMotion: the sequential multi-person ControlNet-based baseline PINO compares against and improves on in overlap and interaction fidelity.","marker":"[6]"},{"why":"Shan et al.: the alternative supervised multi-person diffusion approach that requires a dedicated multi-person dataset and fixed group sizes.","marker":"[28]"},{"why":"InterControl: the joint-control baseline that also generates characters sequentially from a shared prompt but lacks per-pair prompts and post-hoc constraints.","marker":"[33]"},{"why":"in2IN: an additional two-person base model used in the supplementary study to show PINO is agnostic to the choice of base model.","marker":"[26]"},{"why":"InterHandGen: defines the penetration-volume metric used to quantify physical plausibility.","marker":"[17]"},{"why":"SMPL: the body mesh model used to compute interpenetration volumes in evaluation.","marker":"[21]"}],"fun_headline_variants":["Two-person model scales to any crowd via pairwise noise tuning","Pairwise noise optimization lets a two-person model drive any group","Crowd motion from a two-person model: pairwise prompts plus physics penalties","Two-person diffusion, arbitrary group size: no training, just noise penalties","Pairwise prompts, physics penalties: two-person model scales to any group"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every group interaction can be decomposed into a chain of independent pairwise interactions, each describable by its own text prompt; the paper's own limitations section concedes that coordinated three-body behaviors such as one person pushing another into a third are not captured by the two-person base model.","fun_headline_variants_meta":{"raw":{"variants":["Two-person model scales to any crowd via pairwise noise tuning","Pairwise noise optimization lets a two-person model drive any group","Crowd motion from a two-person model: pairwise prompts plus physics penalties","Two-person diffusion, arbitrary group size: no training, just noise penalties","Pairwise prompts, physics penalties: two-person model scales to any group"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001013,"raw_usage":{"total_tokens":4246,"prompt_tokens":882,"completion_tokens":3364,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":3272}},"tokens_in":498,"tokens_out":3364,"duration_ms":22584,"temperature":1.0,"reasoning_tokens":3272,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:55:48.136895+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on a set of genuinely three-person-coordinated actions, such as a group hug, a two-person lift of a third, or a chain-reaction push, with prompts describing the full group behavior, and measure whether the resulting motion contains the required simultaneous three-body contact; if the composed pairwise motions systematically lose the third-person contact or produce physically impossible configurations, the pairwise-decomposition claim is falsified for those behaviors.","supporting_citations":[{"cited_title":"Intergen: Diffusion-based multi-human motion genera- tion under complex interactions","cited_arxiv_id":null,"evidence_quote":"InterGen: supplies the pretrained shared-weight two-person interaction diffusion model and the InterHuman dataset used for evaluation."},{"cited_title":"Op- timizing diffusion noise can serve as universal motion priors","cited_arxiv_id":null,"evidence_quote":"DNO: establishes that optimizing the initial noise of a motion diffusion model can serve as a universal prior for goal-directed generation."},{"cited_title":"Programmable motion generation for open- set motion control tasks","cited_arxiv_id":null,"evidence_quote":"ProgMoGen: provides the differentiable noise-optimization loop that PINO adapts to the multi-person setting, as noted in the supplementary implementation details."},{"cited_title":"Freemotion: A unified framework for number- free text-to-motion synthesis","cited_arxiv_id":null,"evidence_quote":"FreeMotion: the sequential multi-person ControlNet-based baseline PINO compares against and improves on in overlap and interaction fidelity."},{"cited_title":"Towards open domain text-driven synthesis of multi-person motions","cited_arxiv_id":null,"evidence_quote":"Shan et al.: the alternative supervised multi-person diffusion approach that requires a dedicated multi-person dataset and fixed group sizes."},{"cited_title":"Intercontrol: Zero-shot human interaction generation by controlling every joint","cited_arxiv_id":null,"evidence_quote":"InterControl: the joint-control baseline that also generates characters sequentially from a shared prompt but lacks per-pair prompts and post-hoc constraints."},{"cited_title":"in2in: Leveraging individual information to generate human interactions","cited_arxiv_id":null,"evidence_quote":"in2IN: an additional two-person base model used in the supplementary study to show PINO is agnostic to the choice of base model."},{"cited_title":"Interhandgen: Two-hand interaction generation via cascaded reverse diffusion","cited_arxiv_id":null,"evidence_quote":"InterHandGen: defines the penetration-volume metric used to quantify physical plausibility."},{"cited_title":"Smpl: A skinned multi- person linear model","cited_arxiv_id":null,"evidence_quote":"SMPL: the body mesh model used to compute interpenetration volumes in evaluation."}],"review_version":1}