{"id":"3e88c06e-135d-4e0d-8354-55fd358a42df","arxiv_id":"2601.02716","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MorphGS retargets motion from a monocular video onto a rigged 3D character by optimizing target morphology and pose with image-space losses, without 3D source reconstruction or parametric templates.","lead":"This paper presents MorphGS, a system that transfers motion from a single video to a rigged 3D character by directly tuning the character's shape and pose until its rendered views match the video, skipping explicit 3D reconstruction of the source. It reports better motion-fidelity scores than reconstruct-then-retarget baselines on human and animal benchmarks, without using category-specific parametric templates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sec. 4.2's source-video generation may render the target mesh itself, collapsing the synthetic benchmark into same-rig pose estimation and undermining the cross-morphology transfer claim.","rationale":"The reader's weakest_assumption focused on the compatible-topology limitation, which is honestly acknowledged in the conclusion and is a genuine scope restriction. However, the most load-bearing concern is the evaluation protocol ambiguity in Sec. 4.2. The sentence 'we animate the target with the ground-truth pose sequence' can be read as rendering the source video from the target mesh itself, which would make the synthetic benchmark a self-reconstruction task rather than cross-morphology retargeting. If that reading is correct, the headline improvements (Table 1) are not evidence for the central claim of transferring motion between morphologically different characters. Even if the sentence is a typo, the absence of code/data and the lack of any further clarification in the appendix (e.g., no per-pair mesh identities) prevent verification. This is not an accusation of misconduct; it is a correctness risk that must be resolved before the empirical claims can be trusted. The concrete test—rendering source videos from a genuinely different source mesh—would settle the issue. The reader flagged this ambiguity in their rationale as point (1), but did not make it the weakest assumption; hence partial agreement. Since the appropriate disposition remains conditional acceptance pending this clarification, the verdict is unchanged.","tokens_in":21173,"tokens_out":5766,"duration_ms":54428,"concrete_test":"Obtain the dataset generation code/scripts from the authors or, if unavailable, run a minimal control: animate a source mesh A (e.g., a humanoid with different limb proportions) with a motion sequence, render it to video from a fixed camera, then run MorphGS with a different target mesh B; if PMD/FVMD are comparable to the reported values, the benchmark is valid; if performance collapses or matches the no-morphology baseline, the reported results reflect same-rig pose estimation, not retargeting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The synthetic benchmark's central quantitative claim depends on how source videos are generated. Sec. 4.2 states: 'To generate source videos, we animate the target with the ground-truth pose sequence and render a monocular video from a fixed viewpoint.' If 'the target' is the same mesh used as the retargeting target, then source and target are identical in shape and rig, so the task reduces to pose estimation of a known character; the reported gains over baselines (e.g., PMD 1.91 vs SPT+ 2.96) would not demonstrate cross-morphology transfer. Alternatively, if this is a typo for 'source,' the manuscript never clarifies, and the released data/code are absent, so the ambiguity cannot be resolved post hoc. Since Tables 1–2 and the central claim of 'morphology-adaptive' retargeting rest entirely on this protocol, this is the most load-bearing weak point. The compatible-topology assumption (Conclusion) is a real but secondary limitation; the evaluation protocol is more immediately threatening.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MorphGS, a target-driven analysis-by-synthesis framework for transferring articulated motion from a monocular video to a rigged 3D character. The method optimizes a small set of morphology parameters (global scale, bone lengths, local offsets) together with a time-conditioned pose network so that differentiable Gaussian splatting renders the animated target to match the source video. Dense semantic 2D-3D correspondences and diffusion-based synthesized views provide additional supervision. The authors report quantitative improvements over reconstruct-then-retarget baselines on Mixamo, DT4D, and real-world video, with ablations showing the contribution of each component.","tokens_in":21381,"tokens_out":8476,"duration_ms":84539,"significance":"If the result holds, the paper would make a useful contribution: it replaces the common reconstruct-then-retarget pipeline with a direct, template-free optimization in the target's parameter space, and the idea of coupling morphology parameters to the rig to reduce shape-pose ambiguity is well motivated. The manuscript includes a substantial ablation study, per-scene tables, and an identifiability proof sketch, which are strengths. However, the current text contains several formal ambiguities and at least one equation-level issue that block assessment of the central claim. The source-video generation protocol in Sec. 4.2 is ambiguous in a way that could invalidate the cross-morphology claim, and the keypoint loss in Eq. (10)-(12) is written as a constant with respect to the optimized variables. These are load-bearing and must be corrected before the empirical contributions can be evaluated.","major_comments":[{"comment":"The evaluation protocol is ambiguous at the most load-bearing point. The sentence 'To generate source videos, we animate the target with the ground-truth pose sequence and render a monocular video from a fixed viewpoint' can be read as rendering the same target mesh that is later used as the retargeting target. If so, source and target are identical, and the synthetic benchmark reduces to pose estimation of a known character; Tables 1-2 would not test morphology-adaptive cross-character transfer. If this is a typo for 'source' or 'source mesh', the text must say so explicitly, and the construction of the ground-truth target mesh sequence used in PMD must be specified. Without code/data release, this ambiguity cannot be resolved post hoc. This must be fixed before the empirical claims can be assessed.","section":"Sec. 4.2 (Datasets)"},{"comment":"There is a mismatch between optimization and deployment. The optimization updates bone lengths ℓ_b and local offsets o_k, which change the rest-pose skeleton and canonical vertex positions used in forward kinematics and skinning (Eqs. 4-8). At the end, the paper states that only the pose trajectory {P_t(ϕ)} is applied to the original target rig Mtgt, 'preserving its original geometry.' But local rotations estimated against the adapted skeleton do not produce the same global joint trajectories when applied to a rig with different bone lengths; end-effector and body-part trajectories change. The manuscript must specify exactly which object is used for evaluation (the adapted morphology or the original rig) and how the final animation is generated from (ϕ, m). This is load-bearing because the ablation in Table 3 shows bone lengths and offsets account for most of the improvement.","section":"Sec. 3.1, Eq. (1)-(8); final paragraph of Sec. 3.1"},{"comment":"As written, the dense keypoint loss is independent of the optimization variables. Eq. (10) defines the matching score between a source pixel u and a vertex x ∈ V(M_tgt) of the static target mesh, using rendered images I^tgt_v of that same static mesh. The matched point x̃_{u,t} in Eq. (12) is therefore a constant; with π a fixed camera projection, L_keyp is constant with respect to ϕ and m and cannot provide any gradient. The ablation in Table 3 attributes a large improvement to L_keyp (PMD 4.52→2.96), which is only possible if the implemented loss uses the currently posed/deformed vertices or an equivalent differentiable dependence on pose. The formal description must be corrected to match the implemented objective; as it stands, the method section does not support the reported ablation.","section":"Sec. 3.4, Eq. (10)-(12)"}],"minor_comments":[{"comment":"The text 'These results indicate that our target-driven, morphology-adaptive optimization is effective on unconstrained monocular videos [28, 72]' cites SMPL and SMAL references rather than video datasets; the citation should be fixed.","section":"Sec. 4.2"},{"comment":"The pooling operator 'pool_{v,x∈V(M_tgt)}' is underspecified: it should state explicitly which dimensions are pooled and how the similarity map over vertices is reduced to a scalar.","section":"Eq. (10)"},{"comment":"The main-text loss omits the regularization term L_reg, which appears later in Eq. (14). For clarity, define all terms in the main objective or refer to the appendix when L is first introduced.","section":"Sec. 3.4 / Eq. (9)"},{"comment":"The loss weights are listed in a sentence; a small table would improve readability and reproducibility.","section":"Appendix A.1 / loss weights"}],"recommendation":"major_revision","confidential_remarks":"The Sec. 4.2 source-video generation sentence is likely a typo ('target' vs 'source'), but because no code or data is released, the ambiguity is not resolvable by the reader. If the synthetic source videos were rendered from the target mesh itself, the central cross-morphology claim collapses; the authors should be asked to state the protocol unambiguously and, if necessary, rerun the affected experiments. The keypoint loss equation also strongly suggests the written method does not match the implementation, so the derivation and ablation need to be reconciled."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is real: instead of reconstructing the source in 3D and then retargeting, MorphGS directly optimizes the target's morphology and pose in an analysis-by-synthesis loop against the source video. That is a genuine step beyond the reconstruct-then-retarget baselines, and it is not present in the cited prior work. The rig-coupled morphology parameterization—bone lengths, global scale, local offsets anchored to the skeleton—is sensible, and the ablations (Table 3) show each component earns its keep: fixed morphology PMD 3.72 drops to 1.91 with the full parameterization; adding keypoint and multi-view losses takes PMD from 4.52 to 1.91. The appendix ablations on auto-rigging, feature extractors, and view count are also useful and internally consistent. Credit where due: the paper ships a coherent method and a careful ablation story.\n\nNow the soft spots, in proportion. The biggest one is the exact sentence in Sec. 4.2: \"To generate source videos, we animate the target with the ground-truth pose sequence.\" If \"the target\" means the target character, then source and target are the same mesh and rig, and the synthetic benchmark collapses into same-rig pose estimation—the reported gains over SPT+ and NPR+ would not demonstrate morphology-adaptive cross-morphology transfer at all. If it is a typo for \"the source,\" the protocol is fine, but the paper must say so. This is not a manufactured flaw; it is the single most load-bearing sentence in the evaluation, and the ambiguity is real.\n\nBeyond that: no code or data are released, so the ambiguity cannot be resolved post hoc; Tables 1 and 2 have no error bars; and the real-video numbers rely on FVMD, a perceptual metric that is itself not validated against any ground truth. These are real but standard deficiencies. The compatible-topology assumption is acknowledged honestly in the conclusion, though it does limit the scope. The theoretical identifiability result is explicitly a proof sketch under weak-perspective, piecewise-rigid assumptions that the authors admit real videos violate—fine as motivation, not a proof of the empirical results.\n\nNone of this is fatal on its own. The central approach is novel and the internal evidence is strong. The paper deserves peer review, but the referee should demand a clarification of the synthetic protocol, and ideally code/data, before trusting the headline numbers.","headline":"The target-driven optimization idea is genuinely new and the ablations are solid, but the synthetic benchmark sentence in Sec. 4.2 is ambiguous at the exact point that decides whether the headline claim tests cross-morphology transfer at all.","tokens_in":21966,"tokens_out":2025,"would_cite":true,"duration_ms":22932,"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":"MorphGS transfers articulated motion from a monocular video to a rigged 3D character by directly optimizing the target's morphology and pose through image-space supervision, without reconstructing the source in 3D or using category-specific","keywords":["motion retargeting","monocular video","analysis-by-synthesis","morphology parameterization","3D Gaussian splatting","differentiable rendering","articulated pose estimation","rigged characters"],"falsifier":"A cross-topology transfer test: take a source video of a human and retarget to a target whose rig lacks the corresponding degrees of freedom (e.g., a whale or spider), or to a target with extra joints; if the method silently fails or produces geometrically implausible motion there, the compatible-topology assumption is confirmed as the limiting factor. More directly, a synthetic experiment with known ground truth and deliberately degenerate motion (e.g., planar motion only) should break the claimed identifiability of bone-length ratios.","tokens_in":20974,"feed_emoji":"🎬","tokens_out":4409,"duration_ms":41035,"temperature":0.7,"pith_summary":"The paper tries to establish that motion retargeting from monocular video can be solved as a target-driven analysis-by-synthesis problem: instead of reconstructing the source subject and then transferring its motion, one can directly optimize the target character's skeletal proportions and per-frame joint rotations so that a differentiable renderer reproduces the observed video. The central move is a rig-coupled morphology parameterization that ties all shape changes to bone lengths, global scale, and skeleton-anchored offsets, factorizing time-invariant character identity from time-varying pose and reducing shape-pose ambiguity. Dense 2D-3D semantic correspondences and diffusion-synthesized views supply structural and multi-view constraints, and 3D Gaussian splatting provides the differentiable rendering. If this is right, a single video is enough to animate a morphologically different rigged character without parametric body templates or explicit 3D supervision.","feed_headline":"One video drives a new 3D character, no template needed","feed_subtitle":"The system optimizes target bone lengths and pose to match the video, beating reconstruct-then-retarget baselines.","key_machinery":"The rig-coupled morphology parameterization: each canonical vertex is written as a global scale times a skeleton-anchored anchor plus a local offset, where the anchor is a skinning-weighted average of rest-pose joint positions that depend on learnable bone lengths. This couples surface geometry to the kinematic tree so that changing one bone length coherently moves all attached vertices, preventing shape updates from absorbing pose error. The pose trajectory comes from a time-conditioned MLP that outputs per-joint axis-angle rotations and root translation, applied through forward kinematics and linear blend skinning, then rendered by 3D Gaussian splatting. Structural guidance comes from dens","core_discovery":"MorphGS claims that video-to-3D motion retargeting can be formulated and solved entirely in the target's parameter space. Given a source video and a rigged target mesh, it optimizes time-invariant morphology parameters—global scale, learnable bone lengths, and local rest-pose offsets anchored to skeleton joints—together with a time-conditioned pose network that outputs per-joint rotations and root translation. The posed target is rendered with differentiable Gaussian splatting, and losses compare rendered frames against the source video, including a dense correspondence term and synthesized-view rendering. The paper reports consistently lower motion-transfer error than reconstruct-then-retar","pith_inferences":["If the framework holds, motion transfer could become a per-target optimization service: a user supplies any rigged mesh and any video, and the same optimization code adapts without retraining or category-specific assets—an implication the paper does not develop.","The paper's own stated limitation points to a sharp boundary: transfer across very different kinematic topologies (e.g., human arm to whale flipper) should fail; a concrete test would be to measure where retargeting breaks as joint counts and limb configurations diverge.","The dense correspondence and synthesized-view modules are replaceable components; swapping in a different semantic feature extractor or view generator may shift results, so the core claim's practical robustness may be sensitive to these auxiliary choices even though the paper only ablates a few variants.","The identifiability argument assumes weakly perspective cameras and piecewise rigidity; a natural extension is to enforce physics-based contact and soft-tissue constraints, which the authors flag as future work, to handle real-world perspective effects and non-rigid deformation."],"forward_implications":["Monocular RGB video alone can drive a rigged character with different body proportions, removing the need for intermediate 3D reconstruction of the source.","Retargeting quality no longer depends on the accuracy of a source reconstruction step, which is the main bottleneck in reconstruct-then-retarget pipelines.","Because the optimization is target-driven, the approach can handle categories without parametric templates (e.g., birds or whales) as long as the target has a rig, possibly obtained by auto-rigging.","Under the paper's identifiability argument, weak-perspective monocular observations with sufficiently non-degenerate motion determine bone-length ratios uniquely, giving a theoretical foundation for optimizing skeletal morphology directly from 2D video.","The same framework extends naturally to multi-view supervision: synthesized views simply add image-space objectives over the same target parameters, so additional viewpoints can be plugged in without changing the formulation."],"fun_headline_variants":["Video to 3D motion without templates or reconstruction","MorphGS: optimize bone lengths and pose from one video","Target-driven retargeting beats reconstruct-then-retarget","No intermediate 3D: motion transfer directly from video"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The source subject and target character must share a compatible articulated topology: every motion the source can make must be expressible by the target's fixed kinematic tree, and the paper states that this assumption limits transfer across subjects with significantly different kinematic structures.","fun_headline_variants_meta":{"raw":{"variants":["Video to 3D motion without templates or reconstruction","MorphGS: optimize bone lengths and pose from one video","Target-driven retargeting beats reconstruct-then-retarget","No intermediate 3D: motion transfer directly from video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000439,"raw_usage":{"total_tokens":2025,"prompt_tokens":660,"completion_tokens":1365,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":404,"completion_tokens_details":{"reasoning_tokens":1298}},"tokens_in":404,"tokens_out":1365,"duration_ms":13206,"temperature":1.0,"reasoning_tokens":1298,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T12:30:13.804104+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A cross-topology transfer test: take a source video of a human and retarget to a target whose rig lacks the corresponding degrees of freedom (e.g., a whale or spider), or to a target with extra joints; if the method silently fails or produces geometrically implausible motion there, the compatible-topology assumption is confirmed as the limiting factor. More directly, a synthetic experiment with known ground truth and deliberately degenerate motion (e.g., planar motion only) should break the claimed identifiability of bone-length ratios.","supporting_citations":[],"review_version":1}