EVENT DOSSIER
September 7, 2026: UniMate unified model and UniML3D datasets were released on arXiv and Hugging Face.
The research team proposed UniMate on arXiv preprint site on September 4, 2026, and officially released it on the Hugging Face Papers page on September 7. This model is a unified diffusion transformer that does not require re-training or optimization for specific skeletons during testing, and can generate articulated motions of any topology based on rigged 3D assets and text prompts. UniMate integrates skeleton topologies into the attention layer through three mechanisms: graph-aware attention bias based on joint relationships and geodesic distances, spectral rotation position embeddings (RoPE) based on graph-Laplacian generalization, and a global topological conditionalizer aggregated from the residual posture skeletons. To train this model, the team constructed the UniML3D dataset, which includes 13,006 standardized text-paired motion sequences for bipeds, quadrupeds, birds, marine organisms, insects, snakes, and rigid objects. Experiments show that UniMate…
UniMate UniMate 发布统一骨架动画生成模型
September 7, 2026: UniMate unified model and UniML3D datasets were released on arXiv and Hugging Face.
The research team released UniMate unified base model and UniML3D datasets on arXiv and Hugging Face. This model is based on topological-aware diffusion Transformer and can generate models based on rigged 3D assets and text prompts without optimization during testing or single-bone retraining.
2 reportsUniMate 是一款无需测试时优化或单骨骼重训的统一基础模型，能根据 rigged 3D 资产和文本提示合成任意骨骼的 articulated motion。该模型引入拓扑感知扩散 Transformer，通过三种机制将骨骼拓扑集成至注意力：基于成对关节关系与测地距离的图感知注意力偏置、基于图拉普拉斯矩阵推广 RoPE 到任意运动树的谱旋转位置嵌入，以及从余下姿态骨架聚合的全局拓扑条件器注意力。研究团队还整理了 UniML3D 数据集，包含跨越双足、四足、鸟类、海洋生物、昆虫状、蛇形及刚性物体等 13,006 条带统一规范化和文本配对的 motion sequences。在 UniML3D 上训练后，UniMate 在质量、泛化能力和效率方面均优于最先进基线，并支持零样本跨拓扑迁移、补间、扩展及文本引导编辑…
UniMate 是一款无需每骨架重新训练或测试时优化的统一扩散 Transformer，能根据 rigged 3D 资产和文本提示生成任意骨架动画。该模型引入拓扑感知扩散 Transformer，通过图感知注意力偏置、基于图拉普拉斯的谱 RoPE 以及从静止姿态池化的全局拓扑条件器将骨骼拓扑整合到注意力中。作者同时发布了 UniML3D 数据集，包含跨双足动物、四足动物、鸟类、海洋生物、昆虫、蛇及可动刚性物体的 13,006 个标准化文本配对运动序列。
UniMate 是一款无需测试时优化或单骨骼重训练的统一基础模型，可基于 rigged 3D 资产和文本提示合成任意骨骼的 articulated motion。该模型引入拓扑感知扩散 Transformer，通过三种机制将骨骼拓扑集成至注意力：(1) 来自成对关节关系与测地距离的图感知注意力偏差；(2) 基于图拉普拉斯泛化 RoPE 到任意运动树的谱旋转位置嵌入；(3) 从余部姿态骨架聚合的全局拓扑条件器注意力。研究团队整理了 UniML3D 数据集，包含跨越双足、四足、鸟类、海洋、昆虫、蛇形及刚性物体的 13,006 条统一规范化和文本配对的 motion sequences。在 UniML3D 上训练的 UniMate 在质量、泛化能力和效率方面超越现有最先进基线，并支持零样本跨拓扑迁移、补帧、扩展及文…