{"id":"09fa9600-d1ef-4504-b42f-b7db0f13ec51","arxiv_id":"2506.19851","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"AnimaX is a feed-forward system that animates arbitrary articulated 3D meshes by jointly generating multi-view videos and 2D pose maps with a shared-positional-encoding diffusion model, then reconstructing 3D joint motion via triangulation and inverse kinematics.","lead":"AnimaX turns a static 3D character into a moving one from a text description, using a diffusion model that generates both the video and the skeleton's joint positions at once. It aims to make category-agnostic 3D animation fast and automatic, without per-character optimization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Sec. 3.3 pose-map-to-3D reconstruction bridge is never validated: no joint-angle or bone-length error is reported, so the 'arbitrary skeletons' claim rests on an untested assumption that VBench appearance metrics cannot certify.","rationale":"The reader identified the reconstruction pipeline in Sec. 3.3 as the weakest assumption, and my reading converges on the same point: it is the only bridge between the diffusion output and the claimed 3D animation, yet it receives no direct quantitative validation. This is more load-bearing than the narrow-baseline or dataset-count issues, because even a perfect joint video-pose diffusion model would fail the headline claim if the k-means/triangulation/IK stage cannot recover correct joint angles for arbitrary skeletons. The paper's strengths should be acknowledged: the joint video-pose diffusion design with shared positional encodings is a credible mechanism for transferring video priors, and the ablation showing pose-only generation degrades supports the core architectural choice. However, the quantitative evidence for the central claim is indirect. VBench is designed for video appearance and motion statistics, not for verifying that a recovered 3D skeleton matches ground truth; the user study is small and preference-based; and the paper's own note that Dynamic Deg. can be inflated by failed generations weakens the motion-fidelity comparison, where MotionDreamer scores higher than AnimaX. The dataset inconsistency (Table 5 sums to about 202k, not 161k) and the lack of error bars are real but secondary; they affect the scale/evaluation claims, not the core mechanism. A direct pose-space evaluation, comparing recovered joint angles and bone lengths against ground truth on held-out clips across categories, would settle whether the 'arbitrary skeletons' claim is supported. Since the reader already set CONDITIONAL and this concern supports that verdict rather than moving it, I recommend no change.","tokens_in":19350,"tokens_out":5816,"duration_ms":63203,"concrete_test":"Use the held-out evaluation assets (or a new 50-clip split spanning humanoid, quadruped, bird, and articulated furniture) with ground-truth animation parameters. Run the full AnimaX pipeline, then compute: (i) 2D joint detection error between k-means centers in generated pose maps and ground-truth projected joints; (ii) mean per-joint 3D position error and joint-angle error after triangulation/IK; (iii) bone-length drift relative to the template skeleton. Also add one topology absent from training (e.g., a hexapod rig) and report the same metrics. If mean joint-angle error exceeds ~15° or more than 10% of clips show >5% bone-length drift, the Sec. 3.3 reconstruction does not support the 'arbitrary skeletons' claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—feed-forward, category-agnostic animation of arbitrary skeletons—depends on a two-part bridge: the joint video-pose diffusion model must generate pose maps that encode the right motion, and Sec. 3.3 must recover correct 3D joint angles from those maps via k-means color clustering, multi-view triangulation with an unspecified bone-length consistency weight, and inverse kinematics. The paper validates only the outside of this bridge. Tables 2 and 4 report VBench scores on rendered videos; Table 3 reports a 30-participant preference study. No experiment measures 2D joint localization accuracy, triangulation reprojection error, bone-length preservation, or joint-angle error against ground truth. The 'arbitrary skeletons' claim therefore holds only if k-means on generated color-coded pose maps yields accurate 2D joints for unseen topologies and if the triangulation/IK stage reliably maps them to valid 3D poses. The paper itself notes that Dynamic Deg. can be misleadingly high for failed generations, and MotionDreamer scores higher (0.827) than AnimaX (0.661) on this metric, so VBench cannot be read as direct evidence of motion fidelity. Because the RGB stream can look plausible while the pose stream is wrong, a failure in Sec. 3.3 would invalidate the headline result even if the generated videos are attractive. This is a missing-evidence correctness risk rather than an internal contradiction; the fix is direct pose-space evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents AnimaX, a feed-forward framework for animating articulated 3D meshes with arbitrary skeletons from a text prompt. Motion is represented as multi-view multi-frame 2D pose maps, and a diffusion transformer jointly generates RGB videos and pose-map sequences conditioned on template renderings and camera information. Shared positional encodings and modality embeddings are used to align the two modalities. The generated multi-view pose sequences are converted to 3D joint positions by color clustering and multi-view triangulation, then to joint angles by inverse kinematics. The model is fine-tuned from Wan2.1 and trained on a newly curated dataset of 161,023 rigged animation clips. Experiments report VBench scores, a 30-participant user study, and ablations on the joint video-pose architecture.","tokens_in":19605,"tokens_out":6680,"duration_ms":64898,"significance":"If validated, AnimaX would be a practically useful and conceptually clean contribution: it offers a category-agnostic, text-driven, feed-forward alternative to optimization-based 3D-to-4D methods, and the 161k-clip dataset would be a valuable resource for future work. The joint video-pose diffusion design with shared positional encoding is a sensible and clearly motivated way to ground video priors into skeleton-based motion, and the qualitative results in Figs. 1, 2, and 5 are notably more coherent than the shown optimization baselines. The paper is also honest about some limitations (fixed camera viewpoints, bounded video length) in Appendix A.3. However, the quantitative evidence in the current version is too thin to support the headline claims of state-of-the-art motion fidelity and arbitrary-skeleton generalization, because the pose-to-3D reconstruction stage is never directly evaluated.","major_comments":[{"comment":"The reconstruction bridge in Section 3.3 is never validated, and it is load-bearing for the paper's central claim of transferring video motion to arbitrary skeletons. The paper reports no 2D joint localization accuracy, no triangulation reprojection error, no bone-length preservation metric, and no joint-angle error against the rigged ground truth. Because the RGB stream can look plausible while the pose stream is inaccurate, the current Tables 2 and 4, which evaluate rendered videos with VBench, cannot certify that the final 3D animations are correct. Please add direct pose-space evaluation (e.g., PCK/OKS on the extracted 2D joints, triangulation reprojection error, and joint-angle/bone-length error on held-out clips), and specify the bone-length consistency weight in the triangulation objective, which is currently left undefined.","section":"3.3 and 4.1"},{"comment":"The quantitative comparison does not support the abstract's claim of state-of-the-art motion fidelity. On Dynamic Degree, AnimaX scores 0.661 while MotionDreamer scores 0.827; the paper's caveat that Dynamic Degree can be inflated by failed generations is reasonable, but it also means this metric cannot be used as evidence of motion fidelity. No error bars, significance tests, or per-category breakdowns are provided for the 35-case evaluation, and VBench is a video-level proxy whose validity for judging 3D animation correctness is not established. The authors should either add a direct motion-quality metric with confidence intervals or soften the SOTA claim accordingly.","section":"4.1, Table 2"},{"comment":"The evaluation protocol does not yet substantiate the generalization claim. The 35 test pairs are sampled from the same sources used for training (Objaverse, Mixamo, VRoid), and the training distribution is heavily humanoid (Table 5 lists 140,000 'character' clips and 22,881 'anthropomorphic' clips), so 'category-agnostic' and 'arbitrary skeletons' have not been tested out-of-distribution. The user study is based on only 30 participants and does not report the number of judged clips, the instructions, or the statistical significance of the preference differences. Please evaluate on held-out skeleton topologies and categories absent from the training data, and report standard errors or significance.","section":"4.1, Table 3 and A.1"}],"minor_comments":[{"comment":"There is a typo in the Table 4 caption: 'diffusin model' should be 'diffusion model'.","section":"Table 4 caption"},{"comment":"The text prompt in Fig. 6 contains a typo: 'Abat flaps r h ythmically as if it is in flight' should be cleaned to 'A bat flaps rhythmically as if it is in flight'.","section":"Fig. 6"},{"comment":"The comparison includes only Animate3D and MotionDreamer; since Table 1 lists AKD as a recent skeleton-based alternative, a brief discussion of why AKD is excluded from quantitative comparison (e.g., runtime or different output representation) would help the reader calibrate the claims.","section":"4.1"},{"comment":"The efficiency claim in Table 1 would be easier to verify if the GPU hardware and per-method variance were reported; '5 minutes' and '6 minutes' are given without specifying the hardware.","section":"A.2"},{"comment":"The 35 evaluation sequences are described as 'picked out' without a selection procedure; specify whether they are random or curated to avoid the appearance of cherry-picking.","section":"4.1"}],"recommendation":"major_revision","confidential_remarks":"This is a missing-evidence problem rather than a fundamental flaw: the method is trained on ground-truth animation data, and I do not see a circularity concern. The reconstruction stage and the generalization claims can in principle be validated with additional experiments. I recommend major revision; if the authors add direct pose-space evaluation and tighten the claims, the paper could become a strong contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"AnimaX deserves a serious referee, but the referee should make the authors close one specific gap before publication: the pose-map-to-3D reconstruction stage is never validated.\n\nWhat's actually new: representing 3D motion as multi-view, multi-frame 2D pose maps and training a joint video-pose diffusion model with shared positional encodings. That combination is not in the cited literature, and the ablation suggests it matters—pose-only fine-tuning degrades, and shared PE improves alignment. The 160k-sequence dataset is a real contribution. The feed-forward 6-minute runtime is a genuine improvement over optimization-based approaches like AKD.\n\nWhat's solid: the architecture description is clear, the training details are complete, the qualitative results look coherent, and the user study is a reasonable subjective check. The paper does not oversell its runtime claim—the appendix gives the breakdown.\n\nWhere it's soft: the evaluation rests on VBench scores over 35 hand-picked in-distribution sequences, two baselines, no error bars. That alone would be borderline but acceptable for a systems paper. The bigger issue is that the central claim—arbitrary-skeleton, category-agnostic animation—depends entirely on the reconstruction stage in Sec. 3.3. The diffusion model's pose-map outputs are clustered with k-means, triangulated with an unspecified bone-length weight, and converted to joint angles via IK, but the paper reports no 2D joint localization error, no reprojection error, no bone-length preservation, no joint-angle error against ground truth. VBench cannot certify this; the paper itself notes Dynamic Degree can be misleadingly high, and MotionDreamer scores higher on that one metric. The RGB stream can look plausible while the pose stream is off, so the headline claim is unvalidated at exactly the point where it is most brittle.\n\nMinor issue: dataset counts are inconsistent (abstract says ~160k, appendix sums to more; Table 5 categories don't reconcile cleanly). The paper also uses VBench as if it were a motion-fidelity benchmark when it is not.\n\nFor a revision, I'd ask for direct pose-space evaluation on held-out skeletons: joint-angle error, bone-length consistency, and success/failure rates, plus a third baseline (e.g., AnyTop or AKD) and a broader test set. This is fixable and the method is worth fixing.","headline":"New joint video-pose diffusion architecture that could be useful for feed-forward 3D animation, but the paper's key reconstruction stage is unvalidated and the evaluation is too thin to support the arbitrary-skeleton claim.","tokens_in":20190,"tokens_out":3277,"would_cite":true,"duration_ms":29805,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"AnimaX claims that a single feed-forward joint video-pose diffusion model can animate arbitrary rigged 3D meshes from a text prompt, recovering motion by triangulating multi-view pose maps and applying inverse kinematics.","keywords":["3D animation generation","video diffusion models","pose maps","multi-view generation","skeleton-based animation","inverse kinematics","text-to-motion","articulated meshes"],"falsifier":"Render a set of rigged meshes with ground-truth animation, run AnimaX's pose-map generation and reconstruction on them, and compare the recovered joint angles with the ground-truth angles using mean per-joint rotation error; the central claim collapses if the generated videos look plausible but the recovered 3D joints are wrong on typical non-humanoid skeletons.","tokens_in":19103,"feed_emoji":"🎬","tokens_out":10322,"duration_ms":90284,"temperature":0.7,"pith_summary":"AnimaX is a feed-forward system that takes an articulated 3D mesh and a text prompt and produces an animation of that mesh. Motion is represented not as a fixed-topology pose sequence but as multi-view, multi-frame 2D pose maps, so the same model can handle humanoids, animals, and furniture alike. A joint video-pose diffusion model generates the RGB video and the pose maps together, conditioned on template renderings and the prompt, and a shared positional encoding keeps spatially corresponding video and pose tokens aligned. The pose maps are then lifted to 3D joint positions by multi-view triangulation and turned into joint angles by inverse kinematics. The authors argue that this makes category-agnostic text-driven 3D animation a matter of minutes rather than hours of per-asset optimization.","feed_headline":"Text prompt animates arbitrary 3D meshes in six minutes","feed_subtitle":"Joint video-pose diffusion transfers video motion priors to skeletons, replacing hours of optimization.","key_machinery":"The load-bearing object is the joint multi-view video-pose diffusion model. It concatenates RGB video latents and pose-map latents into a single token sequence $x_{\\mathrm{total}}\\in \\mathbb{R}^{(2f+4)\\times h\\times w\\times c}$, with the RGB stream in the first half and the pose stream in the second half. The key identity is the shared positional encoding $\\mathrm{PE}_{i,j,k}=\\mathrm{PE}_{i+(f+2),j,k}=\\boldsymbol{R}(i,j,k)$, which gives spatially matching RGB and pose tokens the same rotary position embedding, while modality-specific embeddings tell the attention layers which stream each token belongs to. Plucker ray maps, which encode each view's camera geometry per pixel, and multi-view attention over all views enforce cross-view consistency. On the output side, per-frame 2D joint positions are recovered by k-means color clustering of the pose maps, lifted to 3D by multi-view triangulation with bone-length consistency, and converted to joint angles by FABRIK inverse kinematics.","core_discovery":"The central claim is that the motion priors of a pretrained video diffusion model can be transferred to skeleton-based 3D animation by generating pose maps jointly with video, rather than by optimizing a deformation field. AnimaX represents 3D motion as multi-view, multi-frame 2D pose maps and trains a diffusion model to predict both RGB videos and pose sequences from template views, template pose maps, camera rays, and a text prompt. The authors report that this joint generation, with shared positional encodings across the two modalities, avoids the distorted or near-static pose outputs produced when pose maps are generated alone, and that the reconstructed animations score higher than the compared baselines on VBench while taking about six minutes per clip. The result, as stated, is a scalable, category-agnostic animation pipeline for diverse articulated meshes with arbitrary skeletons.","pith_inferences":["A direct test of the pipeline would separate diffusion error from reconstruction error by training a learned joint detector on clean pose maps and comparing its recovered joints against k-means clustering.","The fixed-viewpoint limitation described in the paper is likely training-data-specific rather than architectural, since Plucker ray conditioning already supports arbitrary camera trajectories.","The joint video-pose interface could be extended to partially articulated or non-rigid objects by adding a depth stream, because the RGB stream already carries appearance information about the deforming surface.","If reconstruction accuracy matches the qualitative appearance, the pose-map-plus-video representation may serve as a general exchange format for language-driven animation across skeleton definitions."],"forward_implications":["A rigged mesh and a text prompt yield a full mesh animation in about six minutes, replacing the roughly 25-hour optimization used by articulated kinematics distillation.","The same trained model covers humanoids, quadrupeds, birds, furniture, and fictional objects because the pose-map representation and conditioning are not tied to one skeleton topology.","The output includes both a viewable RGB video and an editable, rig-based motion, since the video and pose streams are generated together and the final representation is joint angles on the input skeleton.","Sharing positional encodings between video and pose tokens is reported to be the difference between coherent motion and degraded or static pose output, as shown in the ablation.","Because the final output is a skeleton parameterization, the recovered animation can be re-targeted or edited with standard rigging tools."],"supporting_citations":[{"why":"Supplies the pretrained video diffusion backbone whose motion priors are transferred to pose generation.","marker":"[Wang et al. 2025]"},{"why":"Defines the VBench metrics used to compare AnimaX against baselines in the quantitative experiments.","marker":"[Huang et al. 2024b]"},{"why":"Is the articulated kinematics distillation baseline that AnimaX's feed-forward pipeline is designed to beat on time and generality.","marker":"[Li et al. 2025]"},{"why":"Is the multi-view video diffusion baseline whose deformation-field optimization AnimaX compares against qualitatively and quantitatively.","marker":"[Jiang et al. 2024a]"},{"why":"Is the zero-shot optimization baseline compared in the experiments.","marker":"[Uzolas et al. 2024]"},{"why":"Supplies the non-linear least-squares triangulation used to lift multi-view 2D joints to 3D.","marker":"[Hartley and Zisserman 2003]"},{"why":"Provides the FABRIK inverse kinematics solver that converts triangulated joint positions into joint angles.","marker":"[Aristidou and Lasenby 2011]"},{"why":"Is the raw asset source for the curated 161,023-clip training dataset.","marker":"[Deitke et al. 2023a,b]"},{"why":"Supplies the LoRA technique used in stage-one fine-tuning of the diffusion backbone.","marker":"[Hu et al. 2021]"},{"why":"Defines the rotary position embedding reused as the shared positional encoding across video and pose tokens.","marker":"[Su et al. 2024]"}],"fun_headline_variants":["AnimaX: text-driven 3D animation for arbitrary skeletons in six minutes","No optimization: text prompts drive fast 3D animation on any mesh","Joint video-pose diffusion: from text to 3D animation in six minutes","Animate any mesh from a text prompt in six minutes","Feed-forward text-to-3D animation for any skeleton in six minutes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The final animation is only as accurate as the 2D joint positions that k-means color-clustering recovers from the generated pose maps, and the paper does not report a direct measurement of that recovery accuracy against known ground truth.","fun_headline_variants_meta":{"raw":{"variants":["AnimaX: text-driven 3D animation for arbitrary skeletons in six minutes","No optimization: text prompts drive fast 3D animation on any mesh","Joint video-pose diffusion: from text to 3D animation in six minutes","Animate any mesh from a text prompt in six minutes","Feed-forward text-to-3D animation for any skeleton in six minutes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001059,"raw_usage":{"total_tokens":4439,"prompt_tokens":939,"completion_tokens":3500,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":3403}},"tokens_in":555,"tokens_out":3500,"duration_ms":24579,"temperature":1.0,"reasoning_tokens":3403,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:24:04.597315+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a set of rigged meshes with ground-truth animation, run AnimaX's pose-map generation and reconstruction on them, and compare the recovered joint angles with the ground-truth angles using mean per-joint rotation error; the central claim collapses if the generated videos look plausible but the recovered 3D joints are wrong on typical non-humanoid skeletons.","supporting_citations":[],"review_version":2}