UniMate: One Unified Model to Animate Diverse Skeletons
把文字生成动作的模型改造成“读懂骨架连接图”的模型：同一套权重看到鸟、蛇或台灯的骨架，就根据文字生成对应关节的运动，再用已有蒙皮驱动网格。关键是把骨架结构同时放进注意力、位置编码和每层的全局条件。
Turn a text-to-motion model into one that reads a skeleton’s connectivity: the same weights take a bird, snake, or lamp rig, generate joint motion from text, and animate the mesh through its existing skinning. The key is to supply skeletal structure through attention, positional encoding, and a global condition in every block.
本文重点解读论文 v1 的方法和实验；末节单独标注截至研究日期的仓库状态。所有实验数值均为作者报告，本研究未重新训练或复现实验。
This study explains the methods and experiments in paper v1; the final section separately records the repository state at the study date. All experimental numbers are author-reported; this study does not retrain the model or reproduce the experiments.
研究动机
Motivation
自动生成网格、自动绑定骨架越来越容易，但让角色按文字动起来仍需手工关键帧、动作捕捉或逐资产优化。人的两条腿、鸟的翅膀与蛇的长链没有统一的关节列表；固定人体模板的动作模型无法直接共用。
Mesh generation and automatic rigging are becoming easier, but animating a character from text still often requires manual keyframes, motion capture, or per-asset optimization. Human legs, bird wings, and a snake’s chain do not share a joint list, so a fixed human-template motion model cannot serve all three directly.
| 路线 Approach | 可用能力 Capability | 主要门槛 Main obstacle |
|---|---|---|
| 固定模板动作模型 Fixed-template motion models | 在人或特定动物模板上学习动作 Learn motion on a human or species template | 结构改变后难以直接使用 Do not directly accommodate a different structure |
| GANimator / SinMDM | 可处理不同结构 Can handle different structures | 需要为每个骨架单独训练 Require separate training per skeleton |
| AnyTop | 统一建模多种动物骨架 Model multiple animal skeletons jointly | 原版无文字条件；目标骨架需要参考动作估计归一化统计 Original model lacks text conditioning and needs target reference motion for normalization |
| 无骨架网格动画 Skeleton-free mesh animation | 直接预测形变，适合更自由的表面运动 Predict deformations and allow more flexible surface motion | 输出维度高，与标准骨架编辑工具衔接较弱 High-dimensional output and weaker integration with rig editing |
| UniMate | 骨架与文字共同控制同一个模型 Condition one model on skeleton and text | 依赖有效绑定以及训练数据覆盖 Depends on a valid rig and training coverage |
核心想法是把拓扑当作输入，而不是把一种拓扑固定在网络里。但不同骨架不仅连接方式不同，坐标系、关节命名和动作尺度也不同，因此统一数据表示与网络设计同样重要。（论文 §1–4）
The central idea is to make topology an input instead of fixing one topology inside the network. Different rigs also disagree in coordinate systems, joint names, and motion scale, making data standardization as important as the network design. (Paper §§1–4.)
核心方法
Method
1. 把每个关节、每一帧变成 token
1. Tokenize each joint at each frame
骨架表示为有根运动学树，用 BFS 排序，保留静止姿态下的关节坐标、父节点、深度、标准化关节名以及两两关系和图距离。按骨架的拓扑直径缩放，并统一朝向与初始根位置；推理使用训练集统计，避免向新骨架索取参考动作。
Represent the skeleton as a rooted kinematic tree, serialized with BFS, retaining rest-pose positions, parents, depths, standardized joint names, pairwise relations, and graph distances. Scale by topology diameter and standardize facing and initial root position. Inference uses training-set statistics, avoiding the need for reference motion on a new rig.
每个关节每帧有 12 维：3 维位置、6 维连续旋转表示、3 维世界坐标速度。非根关节的水平位置以根为中心并按朝向对齐，竖直位置仍在世界系；根位置保持全局信息。静态骨架 token 来自关节及其父节点的静止坐标，作为额外时间槽放在运动 token 之前，张量大小为 (T+1) × J × d。
Each joint has 12 features per frame: 3 position coordinates, a 6D continuous rotation representation, and 3 world-space velocity coordinates. Non-root horizontal positions are root-relative and facing-aligned, while vertical positions remain in world coordinates; root positions retain global information. Static skeleton tokens encode each joint’s and parent’s rest positions and precede motion tokens as an extra time slot, producing a (T+1) × J × d tensor.
| 术语 Term | 在本文中的具体含义 Concrete meaning here |
|---|---|
| Topology／拓扑 Topology | 哪些关节相连、谁是谁的父节点；不是网格三角面连接 Joint connectivity and parenthood, rather than mesh triangle connectivity |
| Rest pose／静止姿态 Rest pose | 绑定时的中性骨架坐标，提供骨长与结构 Neutral rig coordinates defining bone lengths and structure |
| Graph distance／图距离 Graph distance | 沿骨架树从一个关节到另一个关节的路径关系 Path separation between two joints along the skeleton tree |
| Spectral coordinates／谱坐标 Spectral coordinates | 图 Laplacian 特征向量在某个关节处的数值 Values of graph Laplacian eigenvectors at a joint |
| Flow velocity／流速度 Flow velocity | 把噪声运动特征变成真实运动特征的生成方向；区别于关节的物理速度 Direction transforming noisy motion features into clean features; distinct from physical joint velocity |
| Zero-shot／零样本 Zero-shot | 用已训练权重处理新骨架或新采样约束，不为它重新训练；仍需预处理 Use trained weights on a new rig or sampling constraint without retraining; preprocessing remains necessary |
2. 三种互补的结构信号
2. Three complementary structural signals
| 组件 Component | 实现方式 Mechanism | 直观作用 Intuition |
|---|---|---|
| 图注意力偏置 Graph-aware attention bias | 把父子、兄弟、祖先关系及图距离嵌入为每个 attention head 的标量偏置，跨帧共享 Embed parent/child, sibling, ancestor relations and graph distance as per-head scalar biases shared across frames | 显式告诉关节注意力哪些关节在骨架上相关；仍允许全局协调 Expose skeletal relationships to joint attention while retaining global coordination |
| Spec-RoPE | 取 8 个非平凡 Laplacian 特征向量，用 SignNet 映射为旋转角，旋转 query/key Map 8 nontrivial Laplacian eigenvectors through SignNet to angles rotating queries and keys | 用图结构位置替代任意关节编号 Replace arbitrary joint indices with structural position |
| 全局拓扑条件 Global topological conditioner | 4 个可学习 query 对骨架 token 做注意力池化，均值汇总后与文字、时间相加，通过 AdaLN-Zero 注入各层 Use 4 learned queries to attention-pool skeleton tokens; mean-pool, add text and time conditions, and inject through AdaLN-Zero | 给各层一个固定维度的整副骨架摘要 Give each layer a fixed-size summary of the whole skeleton |
每个 block 先在同一帧内跨关节注意，再在同一关节上跨时间注意，最后经过前馈层。结构偏置与 Spec-RoPE 用在关节分支；时间分支仍用普通帧索引 RoPE。忽略常数后，这种分解把全时空注意力的平方代价，从约 O(T²J²) 降到 O(TJ² + JT²)；这是由其布局得到的复杂度推导，而非论文计时实验。
Each block attends across joints within a frame, then across time at each joint, followed by a feed-forward layer. Structural bias and Spec-RoPE affect the joint branch; the temporal branch uses ordinary frame-index RoPE. Ignoring constants, this factorization reduces attention cost from roughly O(T²J²) to O(TJ² + JT²). This is a complexity derivation from the layout, not a timing experiment reported by the paper.
\[ a_{ij}^{(h)} = \frac{(R(f(s_i))q_i)^{\top}(R(f(s_j))k_j)}{\sqrt{d_h}} + B_{ij}^{(h)},\qquad L=D-A. \]
式中 s 是 Laplacian 的谱坐标，f 产生旋转角，B 是图关系偏置。SignNet 对每个谱分量同时计算正负输入的特征再相加，使特征向量整体翻转符号时结果不变。（论文 §3.2、附录 B–C）
Here s contains Laplacian spectral coordinates, f produces rotation angles, and B is the graph-relation bias. SignNet adds features evaluated at both signs of each spectral component, making the result unchanged when an eigenvector flips sign. (Paper §3.2, Appendices B–C.)
3. 用 flow matching 学动作分布
3. Learn the motion distribution with flow matching
\[ x_\tau=(1-\tau)x_0+\tau x_1,\quad v^*=x_1-x_0,\quad \hat{x}_1=x_\tau+(1-\tau)v_\theta. \]
训练在高斯噪声 x₀ 与真实动作 x₁ 之间插值，用带 padding mask 的 MSE 学目标流速度；再对去噪动作添加 SO(3) 旋转测地误差和速度平滑项，权重分别为 0.5 与 0.1。物理速度通道是生成对象的一部分，不能与流速度混淆。
Training interpolates between Gaussian noise x₀ and real motion x₁, using masked MSE to learn the target flow velocity. It adds an SO(3) geodesic rotation error and a velocity-smoothness term on denoised motion, weighted 0.5 and 0.1. Physical velocity channels are part of the generated object and should not be confused with flow velocity.
推理用 50 步 Euler 积分，文字 classifier-free guidance 默认强度为 3.0；条件与无条件分支合批计算。随后恢复旋转与根轨迹，通过 forward kinematics 和蒙皮驱动网格。无需逐骨架优化不等于单次网络前向：这是多步生成过程。（论文 §3.3、附录 B.6–B.7）
Inference uses 50 Euler steps with text classifier-free guidance at a default scale of 3.0, batching conditional and unconditional branches. Rotations and the root trajectory are recovered, then forward kinematics and skinning animate the mesh. No per-skeleton optimization does not mean one network forward pass: generation is iterative. (Paper §3.3, Appendices B.6–B.7.)
数据与训练
Data and training
| 论文 v1 数据来源 Paper v1 data source | 动作序列数 Motion sequences | 补充说明 Notes |
|---|---|---|
| Truebones | 1,094 | 74 种骨架 74 distinct skeletons |
| Mixamo | 2,425 | 人类动作 Human motion |
| Objaverse-XL | 9,487 | 来自 6,965 个经过筛选和去重的绑定资产 From 6,965 filtered, deduplicated rigged assets |
| 总计 Total | 13,006 | 2,140,232 帧；约 20 小时；3,584 个独特文字提示 2,140,232 frames; roughly 20 hours; 3,584 unique prompts |
数据流程依次为：保留主要骨架树、修正根节点、去掉无作用辅助骨、移除静态及不合理动作；用语言模型规范关节名并选择朝向参考关节，再从四视图渲染生成动作描述；最后进行 BFS 排序、尺度/朝向统一、相对静止姿态旋转重表达，以及训练集级特征归一化。
The pipeline retains the main skeleton tree, corrects roots, removes nonfunctional helper bones, and rejects static or implausible clips. Language models standardize joint names and select facing-reference joints; four-view renders support motion captioning. BFS ordering, scale/facing alignment, rest-pose-relative rotations, and training-set feature normalization complete the process.
覆盖面广但分布不均：双足占 84.1%，海洋类占 0.8%，蛇形占 0.3%。每个骨架类型含 n 个样本时，单个样本权重按 n⁻⁰·⁵ 设置；在线增强包括删叶节点、插入中性节点、合并单子节点链、轻微改变骨长，并重新计算拓扑描述和运动特征。（论文 §4、附录 A）
Coverage is broad but imbalanced: bipeds account for 84.1%, marine rigs 0.8%, and serpentine rigs 0.3%. For a skeletal type with n samples, each sample receives weight n⁻⁰·⁵. Online augmentation removes leaves, inserts neutral joints, collapses single-child chains, and slightly changes bone lengths, recomputing structural descriptors and motion features. (Paper §4, Appendix A.)
展开论文训练配置
Expand the paper’s training configuration
论文配置：8 层、隐藏维度 512、8 个 attention heads；冻结 FLAN-T5-Base。AdamW 学习率 10⁻⁴，全局 batch 256，100k steps，8 张 H100 约一天；文字条件 dropout 0.1，使用 EMA 参数推理。通常生成 60 帧；180 帧示例来自以 180 帧训练的变体，不是标准窗口天然无限外推。
Paper configuration: 8 blocks, hidden size 512, and 8 attention heads; frozen FLAN-T5-Base. AdamW uses a learning rate of 10⁻⁴, global batch 256, and 100k steps, taking about one day on 8 H100 GPUs. Text-condition dropout is 0.1 and inference uses EMA weights. Typical generation uses 60 frames; the 180-frame examples come from a variant trained on 180 frames, rather than unlimited extrapolation of the standard window.
主要实验结果
Results
1. 新骨架上的动作生成
1. Motion generation on held-out skeletons
作者把 AnyTop 加上文字 cross-attention，并在相同 Truebones 数据设置下训练比较模型。测试留出 7 种完整骨架，训练看不到这些骨架的任何动作：Raptor、Leapord、Raindeer、Parrot2、Jaws、Crab、KingCobra，共 58 条序列、5,861 帧。FID 在运动学特征空间计算，越低越好；Diversity 是 5 个生成样本之间的平均关节距离，越高表示差异更大，但不必然更合理。（表 1、附录 D）
The authors add text cross-attention to AnyTop and train the compared models under the same Truebones data setting. Seven complete skeleton types are held out, with none of their motions seen during training: Raptor, Leapord, Raindeer, Parrot2, Jaws, Crab, and KingCobra, totaling 58 sequences and 5,861 frames. FID is computed in kinematic-feature space, where lower is better. Diversity is average joint distance among 5 generated samples; higher means more variation, not necessarily more plausible motion. (Table 1, Appendix D.)
| 方法 Method | FID ↓ | Diversity ↑ |
|---|---|---|
| AnyTop + text | 2.711 | 8.139 |
| UniMate | 0.757 | 9.200 |
FID 相对下降约 72.1%，由表中数值计算。这个结果支持跨所测骨架的泛化，但不能证明所有形态都可靠：测试只涵盖 7 种骨架，附录给出量化生成 seed 10，主表未报告跨训练随机种子的置信区间。
FID falls by approximately 72.1%, calculated from the table. This supports generalization across the tested rigs, but not reliability on every morphology: the test includes only 7 skeleton types, the appendix specifies generation seed 10, and the main table provides no confidence intervals across training seeds.
2. 网格动画与人类评分
2. Mesh animation and human ratings
网格评测使用 18 个不同形态的 mesh，将结果渲染为固定视角的 512 × 512 视频后计算 VBench。V2M4 使用 Wan2.2 从文字生成驱动视频；UniMate 使用已绑定网格。这是不同输入处理路径下的整体系统比较。OC 为整体一致性，MS 为运动平滑度，DD 为动态程度，AQ 为美学质量，均越高越好。（表 2）
The mesh benchmark uses 18 meshes of different morphologies and evaluates fixed-view 512 × 512 rendered videos with VBench. V2M4 uses Wan2.2 videos generated from text, whereas UniMate uses rigged meshes. This compares complete systems with different input-processing paths. OC measures overall consistency, MS motion smoothness, DD dynamic degree, and AQ aesthetic quality; higher is better for all four. (Table 2.)
| 方法 Method | OC ↑ | MS ↑ | DD ↑ | AQ ↑ | 每段生成耗时 ↓ Generation time per clip ↓ |
|---|---|---|---|---|---|
| V2M4 | 0.167 | 0.991 | 0.667 | 0.506 | 1.641 h |
| AnimateAnyMesh | 0.151 | 0.995 | 0.352 | 0.498 | 15.542 s |
| UniMate | 0.186 | 0.993 | 0.833 | 0.544 | 1.214 s |
UniMate 没有在所有自动指标上获胜：AnimateAnyMesh 的 MS 更高，但其 DD 更低，作者的可视化显示它常接近静止。UniMate 的约 50 FPS 是单张 H100 上约 1.2 秒生成 60 帧的吞吐量，包含文字编码与运动学恢复；不能当作逐帧低延迟或整套资产处理/渲染的帧率。
UniMate does not win every automatic metric: AnimateAnyMesh has higher MS but lower DD, and the authors’ visual examples often show near-static output. UniMate’s roughly 50 FPS means generating 60 frames in about 1.2 seconds on one H100, including text encoding and kinematic recovery. It is not a claim of low-latency per-frame streaming or a frame rate for the entire asset-processing/rendering pipeline.
用户研究共 32 人，每人评价 12 个案例，对四项指标打 1–5 分，方法匿名且展示顺序随机化。下面是文字匹配 TA、动作合理性 MP、表现力 ME、形状保持 SP 与平均分。（表 3）
The user study includes 32 participants, each rating 12 cases on four criteria from 1 to 5, with anonymous methods and randomized presentation. The table reports text agreement (TA), motion plausibility (MP), motion expressiveness (ME), shape preservation (SP), and their average. (Table 3.)
| 方法 Method | TA | MP | ME | SP | 平均 Average |
|---|---|---|---|---|---|
| V2M4 | 2.378 | 2.341 | 2.596 | 2.336 | 2.413 |
| AnimateAnyMesh | 2.193 | 2.930 | 2.362 | 3.747 | 2.808 |
| UniMate | 4.617 | 4.568 | 4.646 | 4.630 | 4.615 |
3. 消融：结构与数据各贡献多少
3. Ablations: contributions of structure and data
| 变体 Variant | FID ↓ | Diversity ↑ |
|---|---|---|
| 去掉图注意力偏置 Without graph-aware bias | 0.773 | 8.799 |
| 去掉 Spec-RoPE Without Spec-RoPE | 0.798 | 8.424 |
| 去掉全局拓扑条件 Without global topology condition | 0.825 | 9.475 |
| 完整模型 Full model | 0.757 | 9.200 |
三项移除均使 FID 变差，但增量为 0.016–0.068，缺少误差区间时不宜过度解读大小排名。去掉全局条件后 Diversity 升到 9.475，伴随不稳定和抖动，说明“多样”不等于“好”。数据消融也很关键：不做静止姿态旋转重表达，FID 升到 1.179；不做在线增强，Diversity 降到 8.265。这里数据规范化的损失大于单项架构消融，不能把全部优势归给 Spec-RoPE。（表 4、附录 E.3–E.4）
Removing any component worsens FID, but the increases are 0.016–0.068; ranking their importance too strongly without uncertainty estimates would be unwarranted. Without global conditioning, Diversity rises to 9.475 alongside instability and jitter, illustrating that variety is not quality. Data ablations also matter: omitting rest-pose rotation rebasing raises FID to 1.179; removing online augmentation reduces Diversity to 8.265. The representation change hurts more than any individual architectural removal here, so the overall advantage should not be attributed entirely to Spec-RoPE. (Table 4, Appendices E.3–E.4.)
应用与控制
Applications and control
| 用途 Application | 采样时怎么做 Sampling procedure | 含义边界 What it establishes |
|---|---|---|
| 跨拓扑动作迁移 Cross-topology motion transfer | 把源动作概括为文字，再给目标骨架生成动作 Describe the source behavior in text and generate on the target rig | 传递动作语义；不保证保留精确时序、轨迹或源动作风格 Transfers semantics, without guaranteeing exact timing, trajectories, or source style |
| 动作补间 Motion in-betweening | 固定起止帧或其他关键帧的姿态 token，补全其余部分 Fix pose tokens at endpoints or other keyframes and fill the remainder | 利用已有姿态约束；与任意手脚空间轨迹控制不同 Uses given poses; distinct from arbitrary hand/foot spatial trajectory control |
| 动作延长 Motion expansion | 以前一段末尾姿态作为下一段边界，逐段更换文字提示 Use the preceding final pose as the next segment’s boundary and change prompts | 可串接片段，但没有无限时长无漂移保证 Enables chaining, without proving indefinite drift-free generation |
| 局部文字编辑 Local text-guided editing | 固定不改的关节 token，只重采样选择的关节 Fix unchanged joint tokens and resample selected joints | 沿用同一模型；复杂接触与动力学仍未强制满足 Reuses the same model; complex contacts and dynamics remain unenforced |
这些应用复用预训练模型，无需为任务微调。论文展示的控制核心是采样约束与文字条件；“动作迁移”尤其应理解为文字介导的重新生成，不是建立逐关节对应后的精确 retargeting。（论文 §5.5、附录 F）
These applications reuse the pretrained model without task-specific fine-tuning. Their control comes from sampling constraints and text. In particular, “motion transfer” is text-mediated regeneration, rather than exact retargeting based on joint correspondences. (Paper §5.5, Appendix F.)
局限与批判性解读
Limitations and critical reading
接触与形态覆盖仍是实用瓶颈
Contact and morphology coverage remain practical bottlenecks
骨架结构限制了可表达的形变，却不自动保证接触物理正确。论文明确展示滑步、漂浮、穿地等问题。附录 E.5 在 20 个留出的有腿骨架（10 双足、10 四足）上评测：IK 足部锁定把近地帧中的滑步比例从 0.107 降到 0.023，滑动距离从 0.542 降到 0.191，距离单位为静止根高度的 10⁻²。该后处理不适用于所有蛇、鱼、飞鸟或机械对象，也不是生成器本身已经学会接触的证据。
Skeletal structure limits possible deformations but does not automatically enforce physical contact. The paper shows foot sliding, hovering, and ground penetration. Appendix E.5 evaluates 20 held-out legged rigs, 10 bipeds and 10 quadrupeds: IK foot locking reduces the skating ratio among near-ground frames from 0.107 to 0.023 and sliding distance from 0.542 to 0.191, in units of 10⁻² times rest-pose root height. This post-process does not generalize to every snake, fish, flying bird, or mechanical object, and does not demonstrate that the generator itself has learned contact.
少见结构与动作容易出现语义错误。台灯案例要求向前倒下，实际只低头再抬起，底座没有翻倒。需要有效 rig 的前提，也意味着高度非刚性表面、流体或超出骨架表达能力的形变不在其优势范围。论文未提供面向任意环境接触与细粒度空间轨迹的完整控制接口。
Rare structures and actions can fail semantically. When asked to fall forward, the lamp merely lowers and raises its head without tipping its base. The requirement for a valid rig also limits suitability for highly nonrigid surfaces, fluids, or deformations beyond the skeleton’s expressive capacity. The paper does not provide a complete interface for arbitrary environmental contact or fine-grained spatial trajectories.
理论性质需按实际参数化理解
Read the theoretical properties with their parameterization
Spec-RoPE 的三个限定条件
Three qualifications for Spec-RoPE
第一，谱坐标平移不变性的严格推导要求旋转角是谱坐标的线性函数。默认非线性 SignNet 保证特征向量符号不变性，却不保留相同的严格相对坐标性质；附录 C.1–C.2 明确作了区分。这里的“平移”发生在谱坐标中，不是角色在世界坐标里的移动。
First, strict translation invariance in spectral coordinates is derived for rotation angles linear in those coordinates. The default nonlinear SignNet provides eigenvector-sign invariance but does not preserve the same strict relative-coordinate property; Appendices C.1–C.2 explicitly distinguish them. “Translation” here refers to spectral coordinates, not moving the character in world space.
第二，SignNet 不能消除重复特征值子空间内任意正交换基的歧义；近重复特征值也可能数值不稳定。如果截断恰好切开重复特征空间，问题仍在。因此不能把“对关节排列等变”解读成所有退化情况的无条件保证。
Second, SignNet does not remove arbitrary orthogonal basis ambiguity within a repeated eigenspace, and near-repeated eigenvalues can cause numerical instability. Truncating inside a repeated eigenspace retains this problem. Permutation equivariance should therefore not be read as an unconditional guarantee in every degenerate case.
第三，有效电阻距离与 attention 衰减的推导用了完整缩放谱坐标和高斯随机频率平均。部署模型使用截断谱与学习到的非线性角映射；论文自己将前者称作解释性视角，而非部署网络的字面定理。（附录 C.3–C.6）
Third, the effective-resistance interpretation and attention decay derivation use full scaled spectral coordinates and averaging over Gaussian random frequencies. The deployed model uses a truncated spectrum and learned nonlinear angles. The paper itself describes the former as an interpretive guide, not a literal theorem about the deployed network. (Appendices C.3–C.6.)
真正新增了什么，证据还缺什么
What is new, and what evidence is missing
批判性定位：从组件看，这是图结构条件化的动作 flow transformer：图偏置、谱编码、AdaLN 与 flow matching 都有先例。最具体的贡献是把三种结构信号与统一骨架/动作预处理整合，使一个文字模型可以在无目标参考动作、无逐骨架再训练的条件下驱动异构 rig；Spec-RoPE 是其中较明确的架构变化，UniML3D 则提供关键训练基础。
Critical positioning: at the component level, this is a graph-conditioned motion flow transformer: graph bias, spectral encoding, AdaLN, and flow matching all have precedents. The concrete contribution is integrating three structural signals with unified rig/motion preprocessing so one text-conditioned model can animate heterogeneous rigs without target reference motion or per-rig retraining. Spec-RoPE is a specific architectural change, while UniML3D supplies crucial training support.
更强证据应包括多训练 seed、每形态分项结果、更大的留出拓扑集，以及等算力和等数据的对照。附录还承认 How to Move Your Dragon 已研究文字条件与可变拓扑，但因写作时无代码/权重而未量化比较；AnimaX 同样未纳入。因此不应把结果扩大为对所有相关方法的全面优越性。引入视频先验、更多稀有形态和形态相关接触约束是合理后续方向，目前仍是展望。
Stronger evidence would include multiple training seeds, per-morphology results, a larger held-out topology set, and compute/data-matched controls. The appendix also acknowledges How to Move Your Dragon as prior text-conditioned, topology-adaptive work, omitted quantitatively because code/weights were unavailable at writing; AnimaX is likewise excluded. Results should not be generalized to superiority over every relevant method. Video priors, more rare morphologies, and morphology-dependent contact constraints are plausible future directions, not established results.
版本与复现
Versions and reproducibility
本研究以 2026-09-04 的 arXiv v1 为实验事实依据。方法见 §3，数据见 §4 与附录 A，主比较见表 1–4，理论限定见附录 C，接触分析见附录 E.5。网页图像提取自该论文的公开 HTML，保留原图中的专名与标签。
Experimental claims in this study follow arXiv v1, dated 2026-09-04. Methods are in §3, data in §4 and Appendix A, main comparisons in Tables 1–4, theoretical qualifications in Appendix C, and contact analysis in Appendix E.5. Images were extracted from the paper’s public HTML, preserving original labels and proper names.
研究日期读取的 官方仓库 README 已给出训练/推理代码、预览权重 和自有 rig 预处理入口；因此论文附录的“未来发布”说明已过时。README 的数据数目为 13,769 clips、7,430 skeletons，与论文的 13,006 不同；推荐配置的层数、训练步数等也有变化。本页不将这些仓库设置冒充论文表格使用的实验版本。
The official repository README read on the study date provides training/inference code, preview checkpoints, and a custom-rig preprocessing entry point, so the appendix’s future-release statement is outdated. The README reports 13,769 clips and 7,430 skeletons rather than the paper’s 13,006 clips; recommended layer counts, training steps, and other settings also differ. This page does not present those repository settings as the experimental version underlying the paper’s tables.
README 说明 Truebones 原始动物动作属于商业资产，公开数据不包含其动作和网格；完整复现该部分需要从原来源取得资产并运行处理脚本。对想使用模型的人，先检查自有骨架的命名、朝向、关节规模和接触效果，比只看“任意拓扑”描述更有用。可在 作者的交互演示 中查看动画。
The README states that original Truebones animal motions are commercial assets, and public data excludes their motions and meshes; reproducing that portion requires obtaining the assets from the original source and running preprocessing. For practical use, check joint naming, facing, rig size, and contact behavior on your own assets rather than relying solely on the “arbitrary topology” description. Animations can be explored in the authors’ interactive demo.