{"id":"955fe7f6-078d-4d5d-a755-8db5efd37b2e","arxiv_id":"2508.09383","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A self-supervised framework encodes whole-body motion into four identity-agnostic latent tokens and uses them to animate reference images, outperforming skeleton-based baselines.","lead":"X-UniMotion learns a compact four-part motion code (face, body, left hand, right hand) from a single video frame, then transfers that motion to a photo of a different person to generate an animated video. The code is trained to ignore the identity of the person, which makes cross-identity animation more robust than explicit skeleton-based methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Identity-agnosticism is asserted but never directly measured; §3.2.2 admits residual leakage, and the only ablation (Table 2) is on synthetic renderings. A linear probe on the latent codes would settle the central claim.","rationale":"The paper's central claim is that four compact latents are expressive, identity-agnostic, and outperform SOTA cross-identity. For that to be true, the self-supervised training of Sec. 3.2.2 must actually strip identity cues from the latents. The paper provides no direct test of this; in fact it discloses residual leakage in Sec. 3.2.2 and tests the synthetic remedy only on synthetic data (Sec. 4.3/Table 2). I do not dispute that the method produces strong qualitative/quantitative results on the in-the-wild benchmark; those results are consistent and the user-study margins are large. But they measure end-to-end transfer, not the identity-agnostic property that is the stated contribution. A method that retains identity in the latent could still win on a benchmark where reference identities and driving identities are drawn from a limited distribution, while failing when body proportions are genuinely novel. This is the single most load-bearing gap. The proposed linear-probe check on the synthetic 3D pairs with shared poses is a direct, low-cost test: if the probe is at chance, the concern is resolved; if above chance, the paper must either add an explicit identity-removal loss or soften the identity-agnostic claim. I agree with the reader's identified weakest assumption, and the conditional verdict stands.","tokens_in":13473,"tokens_out":5856,"duration_ms":61041,"concrete_test":"Train a linear (or single-hidden-layer MLP) classifier to predict subject identity from the extracted latent codes z_full, using the authors' synthetic 3D dataset: 500 rigged characters rendered under identical shared poses. Split by identity and by pose so training and test use disjoint identities and disjoint poses. If held-out identity classification accuracy is significantly above chance, identity information persists in the latent and the Sec. 3.2.2 claim is falsified; chance-level accuracy would support the identity-agnostic premise. This check is feasible with the described synthetic pipeline and directly targets the missing measurement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is Sec. 3.2.2's claim that identity-agnosticism can be enforced by 2D augmentations plus synthetic 3D cross-identity pairs. Training uses the same person as source and target (Sec. 3.2), so the only pressure pushing identity out of z is these augmentations and 3 hours of synthetic data against 200 hours of real video. The paper itself concedes 'identity leakage remains observable' after 2D augmentations for atypical body ratios, and the only quantitative ablation of the synthetic fix (Table 2, row c vs ours) is measured on synthetic 3D test pairs whose renderings are 'not photorealistic' (Sec. 4.3). No experiment measures identity content in z or z_full: no classification probe, no nearest-neighbor analysis, no intervention that removes identity cues and shows the latent is invariant. Without that measurement, the link from 'self-supervised training' to 'identity-agnostic motion latent' to '55.7%/70.3% cross-identity user-study advantage' is not established. If z retains face shape or body proportions, cross-identity results on the curated benchmark may oversell generalization to truly novel body structures.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes X-UniMotion, a self-supervised framework that learns an implicit, compact latent representation for whole-body human motion from a single driving image, using four disentangled tokens for global body, face, left hand, and right hand. These latents are retargeted to a reference subject through a ViT decoder and injected into a pretrained Seaweed-7B DiT video generator. Motion-identity disentanglement is pursued via 2D augmentations and synthetic 3D cross-identity pairs, with auxiliary dual decoders providing pose and hand-normal supervision. The authors report state-of-the-art results on self- and cross-identity reenactment benchmarks, including a user study with 55.7% full-body identity similarity and 70.3% motion accuracy, and ablate components on synthetic keypoint metrics.","tokens_in":13791,"tokens_out":4321,"duration_ms":45311,"significance":"If the central claims hold, the paper would make a solid advance: it replaces explicit pose skeletons with a learned, compact, identity-agnostic motion latent that captures facial expressions, body pose, and hand gestures jointly, enabling a simpler inference pipeline and better handling of occlusions and depth ambiguities. The paper is well structured, and the experiments include multiple baselines, a user study, and ablation studies. However, the most load-bearing claim—identity-agnosticism of the latent codes—is asserted but never directly measured. The absence of code, data, or checkpoints, the bespoke in-house benchmark, and the large training-data/backbone advantage over baselines leave the quantitative claims only partially verified. The paper deserves major revision rather than rejection, because the open concerns are addressable with additional experiments and analysis.","major_comments":[{"comment":"The central claim that the motion latents are identity-agnostic is not directly tested. The paper itself admits (Sec. 3.2.2) that 2D augmentations leave identity leakage for atypical body ratios, and the only quantitative evidence for the synthetic-pair fix (Table 2, row c vs ours) is measured on synthetic 3D test pairs, which Sec. 4.3 says are not photorealistic. No experiment measures identity content in z or z_full: no linear classification probe, no nearest-neighbor analysis, no intervention study that removes identity cues. A linear probe on the latent codes is standard for representation disentanglement and would directly settle whether the 55.7%/70.3% cross-identity user-study gains reflect true identity-agnosticism or the benchmark's composition. This is load-bearing because the title and abstract rest on identity-agnosticism.","section":"Sec. 3.2.2, Table 2"},{"comment":"The cross-identity user study is reported as aggregate percentages with no statistical detail. With 60 participants and 10 randomly selected pairs each, it is unclear how many comparisons involved each baseline, whether the reported numbers are head-to-head win rates or average preference, and no confidence intervals or significance tests are given. Given the very large gaps (55.7% vs 15% FullID-Sim; 70.3% vs 11.7% Mot-Acc), a paired-preference design with per-pair breakdowns and significance testing would strengthen the claim. Also, the text says 'we adopt ID-SIM↓' while the table header is 'FaceID-Sim↑', which is confusing and should be corrected.","section":"Table 1, Sec. 4.2"},{"comment":"There is a training/evaluation confound: the proposed method is trained on ~200h of in-house video plus 3h of synthetic data and fine-tunes Seaweed-7B, while most baselines are released models trained on smaller public datasets. The bespoke benchmark of 100 Midjourney/Pexels references and 100 driving clips may also favor implicit methods because explicit keypoint baselines are more sensitive to viewpoint and composition shifts. The large margins in Table 1 could therefore reflect data scale and backbone capacity rather than the proposed representation. The only controlled same-backbone comparison is ablation (d) in Table 2, which uses keypoint L1 on synthetic renderings, not the user-study metrics. Please provide same-backbone comparisons with explicit pose conditioning in the user study, or report baseline performance after fine-tuning on the same training data.","section":"Sec. 4.1, Sec. 4.2, Table 1"},{"comment":"No code, data, or checkpoints are released. The training data is in-house and the evaluation benchmark is not public. This is a significant verification gap for a counter-intuitive central claim—that a 512-d latent from a single image can outperform explicit pose guidance across identities. At minimum, releasing the evaluation benchmark, the motion encoder/decoder code, and a demo checkpoint would allow independent assessment. Without these, the reported quantitative results cannot be reproduced or compared by the community.","section":"Reproducibility (General)"}],"minor_comments":[{"comment":"There are several typos: 'identity-agonistic' in Sec. 3.2.1 and 3.2.2 (should be 'agnostic'); 'diverse diverse identities' in the Figure 1 caption; 'Seawead et al.' should be 'Seaweed et al.' These should be fixed.","section":"Throughout"},{"comment":"The table formatting for 'X-UniMotion (Ours)' differs from the other rows; also, the metric naming inconsistency (ID-SIM↓ vs FaceID-Sim↑) should be resolved in both text and table.","section":"Table 1"},{"comment":"The text states that all 2D skeleton-based methods are evaluated with per-skeleton alignment, but Table 1 only marks OmniHuman with an asterisk. Please clarify which baselines received alignment and what the alignment procedure was.","section":"Sec. 4.2"},{"comment":"The quantitative ablation in Table 2 is evaluated only on synthetic 3D pairs; the authors note the renderings are not photorealistic. It would be helpful to also report the ablation on the real cross-identity benchmark used in Table 1, even if only for the user-study subset.","section":"Sec. 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is from an industrial lab and uses in-house data; the lack of release may be a policy constraint, but for a journal submission the evaluation benchmark should at least be made public. The most urgent issue is the unmeasured identity-agnosticism: a linear probe on the latent codes is necessary to support the paper's central claim. I also note the paper is closely related to the authors' earlier X-NeMo and X-Portrait work; the novelty over those should be stated more crisply."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a credible product-quality paper from a strong lab, and it deserves serious refereeing. The genuinely new thing is the four-token decomposition—global body, face, left/right hand—learned end-to-end with a DiT video generator, with dual decoder heads for joint heatmaps and hand normals. That is a real step beyond X-NeMo's facial latents and the skeleton-based baselines, and the reported numbers are consistently better across FID/FVD, self reenactment, and cross-identity user study. I believe the method works.\n\nWhat it does well: the local tokens let the model carry fine hand and face detail without inflating the global bottleneck; the dual decoding gives the latent a structured supervisory signal rather than hoping diffusion loss does all the work; and the synthetic 3D pairs are a honest attempt to attack the hardest part of cross-identity transfer. The qualitative results show cases where skeleton-based methods clearly break.\n\nSoft spots, in order of seriousness. First, identity-agnosticism is the load-bearing property and is never directly measured. The paper says 2D augmentations leave observable leakage for atypical body proportions and adds 3 hours of synthetic renderings against 200 hours of real video, but there is no probe on the latent code itself—no linear classifier, no nearest-neighbor check, no intervention showing z is invariant to identity. The only quantitative ablation for the synthetic fix (Table 2) uses synthetic renderings the authors themselves call not photorealistic. A linear probe on z (or on z_full) would settle this and is cheap. Second, the cross-identity benchmark is purpose-built and the numbers are not accompanied by any significance testing, so the 55.7%/70.3% user-study edge over the next best baseline is a point estimate with unknown variance. Third, no code/data/checkpoints are released, and the training set is proprietary—which is normal for industrial systems but makes independent verification harder. I don't see circularity or fabricated results; this is a trained system, not a derivation.\n\nOverall: important for the human-animation subfield, but the paper overstates the confidence in its central mechanism. I'd send it to referees with the explicit request that they check whether identity is actually gone from the latents, and maybe ask for a probe in the revision.","headline":"Solid, well-engineered whole-body motion transfer paper with a genuinely new four-token implicit latent, but the central identity-agnosticism claim is asserted rather than directly measured.","tokens_in":14283,"tokens_out":1991,"would_cite":true,"duration_ms":20592,"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 compact set of four learned latent tokens—body, face, left hand, right hand—can carry identity-agnostic whole-body motion for cross-identity human image animation, outperforming explicit skeleton-based methods.","keywords":["human image animation","motion transfer","latent motion representation","identity disentanglement","diffusion transformer","whole-body reenactment","self-supervised learning","video generation"],"falsifier":"Train a probe classifier on the 512-dim global motion latent $z$ extracted from held-out driving videos to predict the identity of the person in the source frame; if the probe exceeds chance accuracy substantially, identity has not been fully removed. A second observation: run cross-identity animation on an extreme body-proportion mismatch (e.g., an adult driving a child reference, or a cartoon with an oversized head) and check whether motion fidelity collapses when the synthetic 3D pair component is ablated.","tokens_in":13384,"feed_emoji":"🎭","tokens_out":8985,"duration_ms":76493,"temperature":0.7,"pith_summary":"X-UniMotion claims that whole-body human motion—facial expression, body pose, and hand gestures—can be compressed into just four latent tokens (one global body vector plus face, left hand, and right hand vectors) that are expressive enough for fine detail and free enough of identity cues to transfer across different subjects. The tokens are learned end-to-end with a pretrained video diffusion transformer, with identity disentanglement pushed by motion-invariant 2D augmentations and synthetic 3D pairs of different characters sharing poses. The paper reports that this implicit representation outperforms state-of-the-art skeleton- and parametric-model-based methods on both self and cross reenactment, including user-study preferences of 55.7% for full-body identity similarity and 70.3% for motion accuracy over the nearest baseline. If the claim holds, it removes the need for pose detectors at inference and resolves depth ambiguities that plague 2D skeletons.","feed_headline":"Four motion tokens beat skeletons for cross-identity animation","feed_subtitle":"Compact face-body-hand latents transfer expressive motion between subjects, preserving identity without pose detectors","key_machinery":"The four-token motion latent $z_{\\mathrm{full}} = (z, z_f, z_{\\mathrm{lh}}, z_{\\mathrm{rh}})$, extracted by ViT encoders from the driving image, is retargeted by an identity-aware ViT decoder into spatial guidance for a pretrained DiT video generator; the face token is fed via cross-attention. Three supervision signals jointly enforce the identity-agnostic property: motion-invariant 2D augmentations (color jitter, scaling up to 30%, piecewise affine), synthetic 3D cross-identity pairs with shared poses and randomized limb/head/hand scales, and dual decoders predicting per-joint heatmaps and hand normal maps from the intermediate motion features.","core_discovery":"The central discovery is that a low-dimensional, carefully factorized motion latent—global body $z$, face $z_f$, left hand $z_{\\mathrm{lh}}$, right hand $z_{\\mathrm{rh}}$, each a 512-dim vector—can carry identity-agnostic whole-body motion when learned jointly with a DiT-based video generator. A ViT encoder extracts the global token from the whole driving frame, and two lighter ViT encoders extract face and hand tokens from crops; an identity-aware ViT decoder fuses these tokens with reference-image patches to produce spatial motion guidance aligned to the reference subject, and the facial token is injected through cross-attention. Disentanglement from appearance is not imposed by explicit i","pith_inferences":["The paper does not directly probe how much identity information survives in the latents; a linear classifier or nearest-neighbor retrieval on $z$ over held-out identities would give a more direct test of the disentanglement claim than the user-study numbers.","The four-token factorization suggests a modular control interface—editing or replacing the face token, or one hand token, should affect only that body part—a capability the paper does not demonstrate but that would follow naturally if the tokens are truly disentangled.","The reliance on synthetic 3D pairs implies a testable scaling prediction: as the diversity of body proportions in the 2D real data grows, the contribution of synthetic data should shrink; measuring the ablation gap on increasingly diverse real videos would expose how much the 3D renderings are load-bearing.","Because the non-global tokens are local crops, their viewpoints are limited to what the crop contains; cropping a hand that is occluded or out of frame may produce a degenerate token, and a test with heavy occlusion could reveal whether the global token compensates or the hand token degrades output."],"forward_implications":["Cross-identity animation no longer requires per-skeleton alignment or bone-length rescaling; motion is extracted directly from raw pixels, so occlusion, unusual lighting, and self-overlap are handled by the learned latent rather than by detector heuristics.","Depth ambiguities of 2D skeletons (e.g., finger crossings, limb interleaving) are reduced because hand normal map supervision and the implicit bottleneck force the latent to preserve depth ordering.","The same identity-agnostic latent can be used for motion outpainting: a small motion diffusion model predicts future latent tokens from a short context, and those tokens are retargeted to any subject, yielding longer generated videos with coherent identity.","Stylized and non-anthropomorphic characters with atypical body proportions become animatable because synthetic 3D pairs explicitly train the encoder to ignore head/hand/limb scale differences."],"supporting_citations":[{"why":"Supplies the pretrained Seaweed-7B DiT video generator that the motion encoder and retargeting module are trained against.","marker":"[Seawead et al. 2025]"},{"why":"Provides the reference-image appearance conditioning design (concatenating reference latents with noised video latents) adopted for identity preservation, and serves as the DiT-based baseline.","marker":"[Lin et al. 2025]"},{"why":"Sapiens supplies the per-joint heatmaps and hand normal maps used as auxiliary supervision in the dual decoders.","marker":"[Khirodkar et al. 2024]"},{"why":"X-NeMo supplies the identity-agnostic facial motion latent idea and cross-attention facial conditioning that X-UniMotion extends to body and hands.","marker":"[Zhao et al. 2025]"},{"why":"Defines the DiT diffusion-transformer architecture underlying the video generator.","marker":"[Peebles and Xie 2023]"},{"why":"ArcFace provides the face identity similarity metric used to quantify identity preservation in cross-identity evaluation.","marker":"[Deng et al. 2019]"},{"why":"Supplies the GAN-based face renderer used for auxiliary facial expression supervision.","marker":"[Wang et al. 2023]"},{"why":"MimicMotion is a key skeleton-based baseline whose confidence-aware pose guidance and reported performance anchor the comparison in the user study.","marker":"[Zhang et al. 2024]"}],"fun_headline_variants":["Four tokens, no skeletons, full-body animation","Compact motion latents transfer face, body, hands across subjects","Motion tokens, not skeletons, for identity-free animation","Four disentangled latents animate any subject expressively","No skeletons needed: four tokens carry whole-body motion"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"Identity-agnosticism rests entirely on self-supervised training signals—2D augmentations and synthetic 3D cross-identity pairs—without any explicit identity label or removal loss, so if those cues are insufficient the latents will carry residual identity and cross-identity transfer will degrade.","fun_headline_variants_meta":{"raw":{"variants":["Four tokens, no skeletons, full-body animation","Compact motion latents transfer face, body, hands across subjects","Motion tokens, not skeletons, for identity-free animation","Four disentangled latents animate any subject expressively","No skeletons needed: four tokens carry whole-body motion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1394,"prompt_tokens":767,"completion_tokens":627,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":550}},"tokens_in":511,"tokens_out":627,"duration_ms":6995,"temperature":1.0,"reasoning_tokens":550,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:05:15.316965+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a probe classifier on the 512-dim global motion latent $z$ extracted from held-out driving videos to predict the identity of the person in the source frame; if the probe exceeds chance accuracy substantially, identity has not been fully removed. A second observation: run cross-identity animation on an extreme body-proportion mismatch (e.g., an adult driving a child reference, or a cartoon with an oversized head) and check whether motion fidelity collapses when the synthetic 3D pair component is ablated.","supporting_citations":[],"review_version":1}