NECromancer: Breathing Life into Skeletons via BVH Animation
Mingxi Xu, Qi Wang, Zhengyu Wen, Phong Dao Thien, Zhengyu Li, Ning Zhang, Xiaoyu He, Wei Zhao, Kehong Gong, Mingyuan Zhang
TL;DR
NECromancer introduces a universal motion tokenizer that encodes, compresses, and synthesizes motion data across diverse skeletons, enabling cross-species transfer and advanced motion analysis.
Contribution
It presents NECromancer, a novel framework with skeletal encoding, topology-invariant tokenization, and a large BVH dataset, facilitating cross-morphology motion processing.
Findings
High-fidelity motion reconstruction across skeletons
Effective disentanglement of motion and skeletal structure
Supports cross-species transfer and generation tasks
Abstract
Motion tokenization is a key component of generalizable motion models, yet most existing approaches are restricted to species-specific skeletons, limiting their applicability across diverse morphologies. We propose NECromancer (NEC), a universal motion tokenizer that operates directly on arbitrary BVH skeletons. NEC consists of three components: (1) an Ontology-aware Skeletal Graph Encoder (OwO) that encodes structural priors from BVH files, including joint semantics, rest-pose offsets, and skeletal topology, into skeletal embeddings; (2) a Topology-Agnostic Tokenizer (TAT) that compresses motion sequences into a universal, topology-invariant discrete representation; and (3) the Unified BVH Universe (UvU), a large-scale dataset aggregating BVH motions across heterogeneous skeletons. Experiments show that NEC achieves high-fidelity reconstruction under substantial compression and…
Peer Reviews
Decision·Submitted to ICLR 2026
* This work tackles an interesting and important problem. Unified skeleton representation has the potential to greatly expand the scope of motion generation models and break free of limitations imposed by scarcity of 4-D data for a variety of skeleton types. * The proposed method of learning a unified token representation of motions is a reasonable and potentially useful direction to solve this problem. * Experimental results show improved reconstruction over a naive padding-based approach for s
* The motion transfer and generation directions are not explored very thoroughly, and only a few examples are provided. * The details of tokenization in the paper are somewhat difficult to follow, especially the relation between the graph embedder and the tokenization model. Perhaps it would help to move Figure 3 and 4 into the supplementary material and provide more mathematical details of training these models in the main text. * The training of the graph embedder is based on heuristic objecti
* The goal to unify the motion tokenization is ambitious. * The combined dataset with the curation strategy could help the community
* Hard to interpret the provided videos * What do the "transfer" examples suppose to mean? All of them have the prefix "gt" (ground truth?). Which ones are the source motions, which ones are the transferred motions? * I see only quadrupeds in the transfer folder. Any non-quadruped transfer examples? Humanoid-to-quadruped or quadruped-to-humanoid? * Questionable generalizability to different skeletal morphologies * For example, it looks like the joint semantic loss is taken over the same jo
1. **Addresses a significant problem**: This paper directly confronts a core limitation in the field of motion generation—the model's dependency on specific skeleton topologies. The proposed universal tokenizer, capable of handling arbitrary BVH skeletons, greatly expands the applicability of motion models and holds significant research and practical value. 2. **Systematic contribution**: The contribution is comprehensive and solid. The authors not only propose a new model (NEC) but also buil
1. **Strong dependency on data quality**: The `OwO` encoder relies on extracting semantic features from joint names. This means the model's performance is likely highly dependent on the standardization and consistency of joint naming within the BVH dataset. For data from the wild with messy or non-semantic names, the model's generalization capability might be compromised.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsHuman Motion and Animation · Human Pose and Action Recognition · 3D Shape Modeling and Analysis