{"id":"c4360f48-1c1c-412d-a26e-a4c67a2050e7","arxiv_id":"2508.01878","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A mixed-initiative system combining video motion capture with editable skinning-weight transfer lets stylized 3D characters mimic human videos.","lead":"VidAnimator lets people upload a video of a human moving and a 3D cartoon character, then automatically transfers the motion onto the character. Users can fix mistakes by editing poses and by repainting the character's skinning weights, which the paper says improves both similarity and smoothness.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline claim that user-guided refinement improves quality is not established because §7.1.4 compares author-curated refinements against an automatic baseline, not refinements produced by target users.","rationale":"The reader's weakest assumption identifies the same gap: the evaluation does not show that target users can effectively perform the refinements. My reading strengthens that concern by noting that §7.1.4's paired comparison is between automated output and refinements whose authorship is undisclosed and likely expert-driven, so it cannot support the claim that the interactive system helps non-experts. I do not see an internal inconsistency in the pipeline itself, and the case studies demonstrate feasibility, so the paper is not a rejection. A conditional verdict is appropriate: the paper should be accepted only if the authors provide a proper user study, full statistics, and transparency about who performed the refinements. Since the reader already set CONDITIONAL, my verdict is UNCHANGED. I agree with the reader's assessment rather than partially or fully disagreeing.","tokens_in":19351,"tokens_out":2861,"duration_ms":36477,"concrete_test":"Recruit at least 20 participants matching the target-user profile (no professional 3D animation experience). Give each participant a brief tutorial, then have each independently use VidAnimator on one assigned video/character pair and perform refinement until satisfied. Blind raters (not authors) score the resulting animations and the automatic baseline on the same 7-point Likert items, with presentation order counterbalanced. Report per-participant improvement, the proportion of participants who achieve a meaningful gain (e.g., ≥0.5 on similarity), time spent, and a paired t-test with full statistics. Also disclose in the paper who performed the refinements in the current study, and share the refined outputs and code so the existing result can be inspected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that user guidance adds value—rests on §7.1.4, where refined outputs score 4.30 vs 4.03 on similarity and 4.48 vs 4.00 on smoothness, with a paired t-test asserted at α=0.05. The paper never states who performed the refinements. §7.3 describes a single 'user' making weight-label edits but gives no recruitment, expertise, or independence information; the questionnaire participants (§7.1.3) only rated outputs and never used the system. The comparison is therefore between the automatic pipeline and an expert- or author-curated result. If the designers made those edits knowing the failure modes, the reported gains do not show that the intended users can achieve them. §7.3.3 even reports the Pose Editor was 'challenging to navigate and might require excessive manual intervention.' Thus the only quantitative evidence for the mixed-initiative claim lacks construct validity: the independent variable is not 'user guidance' but 'researcher-selected refinement.' The paired t-test itself is also reported without test statistics, degrees of freedom, or p-values, making the significance claim unverifiable from the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents VidAnimator, a mixed-initiative system that animates stylized 3D characters from single-view human videos. The pipeline combines DeepMotion's monocular motion capture for a human character with the skeleton-free pose transfer model SFPT to drive the stylized target character. Two interactive tools are introduced: a pose editor for frame-level joint correction and a skinning-weight editor that maps vertex color edits to weight changes through a color-weight converter based on inverse distance weighting and kernel density estimation. The authors evaluate the system via a questionnaire study (N=30) comparing automatic (MoCap+MoTrans) output against a refined version that includes interactive adjustments, reporting higher similarity and smoothness ratings for the refined output. Three case studies with humanoid and non-humanoid characters illustrate the use of the weight editor, and the paper distills two refinement guidelines from this experience.","tokens_in":19606,"tokens_out":6418,"duration_ms":76774,"significance":"If the central claim is supported, the work makes a useful contribution to accessible stylized character animation by integrating established MoCap and pose-transfer models with a lightweight, intuitive refinement interface. The color-based weight editor is an appealing idea and the case studies give qualitative evidence that manual weight-label adjustments can repair visible artifacts. The paper is clearly written, and the design considerations (C1, C2) are sensible. However, the quantitative evidence for the headline claim that user guidance improves quality is currently not convincing: the evaluation lacks full inferential statistics, and the identity/independence of the person who performed the refinements is not disclosed. As a result, the significance of the measured improvements is uncertain.","major_comments":[{"comment":"The paper claims that the refinement method 'achieved significantly higher scores' at α=0.05 with a paired t-test, but it reports only means and standard deviations, with no t-statistic, degrees of freedom, p-value, effect size, or confidence interval. The reader cannot verify the significance claim from the manuscript. Please report the full test results or soften the claim accordingly. In addition, the paper never states who performed the interactive refinements presented to questionnaire participants; §7.3 describes 'the user' without any recruitment or expertise information. If the refinements were made by the system's designers, the comparison conflates system capability with author expertise, and the results do not demonstrate that target users can achieve these gains. The study should either include target users performing the refinement or be explicitly framed as a demonstration of expert-curated refinement.","section":"7.1.4 (Effectiveness)"},{"comment":"The skinning weights are defined in Eq. (1) to satisfy ||W||_1 = 1 with entries in [0,1]. However, Eq. (3) normalizes the edited weight vector by its L2 norm. For non-negative vectors, L2 normalization does not preserve the sum-to-one property (e.g., a vector (0.5, 0.5) becomes (0.707, 0.707) after L2 normalization). Thus the color-weight converter as written can output weights that violate Eq. (1), which should affect the subsequent linear blending skinning deformation. Please correct the normalization (e.g., use L1 normalization or a softmax) and clarify how the predicted skinning weights maintain the stated constraint.","section":"4.5.2, Eq. (3)"},{"comment":"Even apart from the missing test statistics, the reported mean differences are small (0.27 for similarity and 0.48 for smoothness on a 0–6 scale), and the sample is a convenience sample of 30 students recruited through WeChat. The paper does not report confidence intervals or effect sizes, making it impossible to judge whether the differences are practically meaningful, and the two paired t-tests are conducted without correction for multiplicity or any check of the normality assumption. Please provide a more complete statistical account, including effect sizes and confidence intervals, or temper the claim that interactive adjustments 'yield superior performance.'","section":"7.1.4 / 7.1.5"}],"minor_comments":[{"comment":"Section 6 is also titled 'Method Analysis' but contains only a duplicate of the Implementation paragraph from §5.5; Section 7 then begins with the actual evaluation. Please renumber or remove the duplicate section.","section":"Sections 6 and 7"},{"comment":"The Gaussian kernel in Eq. (7) is written with h^2 in the normalization, which is not the standard form and is inconsistent with the definition of K_h in Eq. (6). Please check the kernel formula and the bandwidth normalization.","section":"4.5.2, Eqs. (6)–(7)"},{"comment":"The sentence about assigning the k'th value of W_v to 0 for a non-covered deformation body part is repeated verbatim later in the same subsection; please remove the duplicate.","section":"4.5.2"},{"comment":"There is a typo: 'constrants' should be 'constraints'.","section":"4.1.2, Eq. (1)"},{"comment":"The text uses 'Phd students'; this should be 'PhD students'.","section":"7.1.3"},{"comment":"The description of the Likert scale is slightly confusing: the scale is described as ranging from 0 to 6, and a score above 4 is said to align with 'somewhat agree.' Clarify the scale anchors, especially whether 3 is the neutral midpoint.","section":"7.1.4"},{"comment":"The bar charts in Fig. 4(B) and (C) do not show error bars or confidence intervals, making it hard to gauge the variability; consider adding them to support the statistical claims.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The central idea and system integration are plausible, and the paper could become acceptable after a revision that (1) clarifies who performed the refinements and adds proper statistical reporting, (2) fixes the normalization inconsistency in Eq. (3), and (3) rebalances any over-strong claims about the mixed-initiative benefit. The case studies and guidelines are useful but do not by themselves establish that novice users can achieve the reported improvements. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Truth first. This is a mixed-initiative system paper, and the genuinely new piece is the weight editor. Users recolor vertices on the target character to adjust skinning weights, and the color-weight converter (Eqs. 2–7) maps those edits to new weights with inverse distance weighting plus a KDE density term to downweight nearby mislabeled vertices. That is a real interaction technique for skeleton-free pose transfer, and it is described with enough detail to reimplement. The rest of the pipeline is a sensible integration of existing components: DeepMotion for MoCap, SFPT for pose transfer, a pose editor for frame-level joint correction. The case studies (Pikachu, Nailong, Charizard, a dinosaur, a flower) visually support the claim that the tools fix common artifacts, and the two guidelines in Section 7.2 are concrete and practical. The authors also acknowledge real limitations: penetration on chubby characters, difficulty with abstract characters, and the pose editor's tediousness.\n\nThe soft spots are in the evaluation. The questionnaire (N=30, convenience-recruited students) only shows that refined outputs are rated higher than the automatic baseline on similarity and smoothness. It never states who did the refinements. Section 7.3 describes \"the user\" in the case studies but gives no recruitment, expertise, or independence; Section 7.3.3 notes the Pose Editor was \"challenging to navigate and might require excessive manual intervention.\" So the comparison is at risk of being author-curated outputs versus the automatic pipeline. That is a construct validity problem for the mixed-initiative claim. The significance claim is also unverifiable as written: no test statistics, degrees of freedom, or p-values are reported. No code or data are provided, and the paper contains an obvious editorial slip—Section 6 repeats the implementation text from Section 5.5, with the evaluation then mislabeled as Section 7 \"Method Analysis.\" These are fixable but should be fixed before publication.\n\nThe central idea holds up conceptually: the weight editor is a meaningful step toward making skeleton-free pose transfer editable for non-experts. But the evidence that target users can achieve the reported gains is missing. This deserves a serious referee, not a desk reject, with the expectation that the user evaluation gets redone or substantially reframed.\n\nWho is this for: researchers in HCI and animation tool design. I'd cite the weight-editor equations in related work, but I wouldn't rely on the user-study numbers.","headline":"A genuinely novel color-based weight editor for skeleton-free pose transfer, wrapped in an under-evaluated mixed-initiative system that deserves peer review but not acceptance as-is.","tokens_in":20095,"tokens_out":3806,"would_cite":true,"duration_ms":40726,"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 mixed-initiative pipeline lets stylized 3D characters mimic human videos, and user-guided edits measurably improve the automatic result.","keywords":["motion capture","motion transfer","stylized 3D characters","mixed-initiative design","skinning weights","character animation","human-AI collaboration","video-driven animation"],"falsifier":"Recruit novice participants, give them only the VidAnimator interface and a short tutorial, and ask them to refine the same video-character pairs used in the questionnaire; if their refined outputs are not rated higher than the automatic baseline on similarity and smoothness, the paper's central claim about user-guided improvement is contradicted.","tokens_in":19190,"feed_emoji":"🎬","tokens_out":6079,"duration_ms":63285,"temperature":0.7,"pith_summary":"VidAnimator proposes a mixed-initiative pipeline for making stylized 3D characters—both humanoid and non-humanoid—mimic the motion in an ordinary single-view human video, without requiring the user to rig a skeleton. The paper's central claim is that adding two interactive editing tools, a pose editor for frame-level joint corrections and a weight editor for adjusting how character body parts correspond to human body parts, measurably improves the result over running motion capture and motion transfer automatically. A questionnaire study with 30 participants found that refined outputs scored higher than the automatic baseline on both motion similarity (4.30 vs. 4.03) and smoothness and naturalness (4.48 vs. 4.00), with the differences reported as statistically significant. The work matters because stylized character animation is normally a specialist skill, and the authors argue their framework lowers the barrier for general users while keeping creative control.","feed_headline":"User-guided edits beat automatic animation transfer for stylized 3D","feed_subtitle":"Interactive pose and weight tweaks measurably improve stylized 3D animation from a single human video.","key_machinery":"The load-bearing mechanism is the skinning-weight correspondence between the source human character and the target character: per-vertex skinning weights encode which deformation primitive moves each part of the mesh, and SFPT transfers poses using these weights rather than a skeleton. On top of this, the paper contributes a color-weight converter that turns a user's vertex recoloring into a new skinning weight vector, using inverse-distance weighting combined with a kernel density estimate so that nearby mispredicted vertices do not dominate. The pose editor supplies the second mechanism, letting users adjust joint rotations and translations on chosen keyframes of the captured human animation. Together these make the otherwise opaque learned correspondence editable by a non-expert.","core_discovery":"On its own terms, the paper establishes that automatic motion capture with DeepMotion plus skeleton-free pose transfer with SFPT can already drive stylized 3D characters from human videos, but the result is imperfect and the implicit correspondence between the human character and the target character is often not what users want. The discovery is that letting users edit either the captured human pose or the skinning-weight correspondence—through a color-based weight editor that translates vertex colors into modified skinning weights—fixes the largest visible failures, such as deformed hands, mismatched legs, and lifeless abstract characters. Across seven target characters, the refined animations were rated more similar to the source video and smoother and more natural than the fully automatic baseline; humanoid characters benefited more than non-humanoid ones, and three case studies show specific weight-label corrections that remove deformations.","pith_inferences":["If the color-weight converter pattern generalizes, any learned correspondence field—for example in 2D image animation or avatar control—could expose a directable color or weight overlay to end users.","The reported gains are modest (about 0.3 points on a 7-point scale), so a reasonable next test is a between-subjects study where novices and experts each refine the same clips to see whether the improvement survives without researcher involvement.","The Pose Editor's frame-by-frame joint editing is a bottleneck for longer videos; pairing it with recent motion-synthesis models could keep the interactive benefit while reducing manual keyframing.","A larger validation with more target characters and source videos, plus measurable metrics such as foot sliding or mesh penetration, would test whether the perceived improvements correspond to objective motion quality."],"forward_implications":["Non-professional users can produce usable stylized 3D character animations from a single human video, skipping rigging and weight painting.","Interactive refinement is worth doing: the questionnaire results indicate that the mixed-initiative result beats the automatic MoCap-plus-MoTrans baseline on perceived similarity and smoothness.","Humanoid characters are the near-term sweet spot; non-humanoid characters like flowers can be animated but require more correspondence tuning and remain more abstract.","The weight-label guidelines (mirror good correspondences across symmetric sides, map unusual labels to human anatomy) give a repeatable recipe for fixing common deformations."],"supporting_citations":[{"why":"DeepMotion MoCap model that captures human motion from the single-view input video.","marker":"[5]"},{"why":"Skeleton-Free Pose Transfer model that transfers the captured human poses to the target stylized character via predicted skinning weights.","marker":"[33]"},{"why":"AMASS motion-capture dataset used in training the pre-trained SFPT model.","marker":"[37]"},{"why":"Mixamo character dataset used in training the pre-trained SFPT model.","marker":"[3]"},{"why":"RigNet dataset used in training the pre-trained SFPT model.","marker":"[57]"}],"fun_headline_variants":["User-guided edits improve stylized 3D character animation","Interactive tweaks beat automatic 3D motion transfer","Editing poses and weights fixes stylized character animation","User oversight enhances 3D character mimicry from video","Manual adjustments refine stylized 3D animation from clips"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole benefit of interactive refinement rests on the assumption that the intended users, not just the researchers, can actually operate the pose editor and the weight editor; the paper does not report who performed the refinements, and one user already found the Pose Editor hard to navigate.","fun_headline_variants_meta":{"raw":{"variants":["User-guided edits improve stylized 3D character animation","Interactive tweaks beat automatic 3D motion transfer","Editing poses and weights fixes stylized character animation","User oversight enhances 3D character mimicry from video","Manual adjustments refine stylized 3D animation from clips"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1396,"prompt_tokens":891,"completion_tokens":505,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":427}},"tokens_in":507,"tokens_out":505,"duration_ms":5812,"temperature":1.0,"reasoning_tokens":427,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:18:49.886612+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recruit novice participants, give them only the VidAnimator interface and a short tutorial, and ask them to refine the same video-character pairs used in the questionnaire; if their refined outputs are not rated higher than the automatic baseline on similarity and smoothness, the paper's central claim about user-guided improvement is contradicted.","supporting_citations":[{"cited_title":"DeepMotion","cited_arxiv_id":null,"evidence_quote":"DeepMotion MoCap model that captures human motion from the single-view input video."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Skeleton-Free Pose Transfer model that transfers the captured human poses to the target stylized character via predicted skinning weights."},{"cited_title":"Troje, Gerard Pons-Moll, and Michael J","cited_arxiv_id":null,"evidence_quote":"AMASS motion-capture dataset used in training the pre-trained SFPT model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Mixamo character dataset used in training the pre-trained SFPT model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RigNet dataset used in training the pre-trained SFPT model."}],"review_version":1}