A preprint posted as arXiv version 1 on 29 September 2026 describes a neural character-animation system that adjusts the length of the movement used to connect two input clips. It first groups motions into clusters, searches for a route through those clusters, and feeds that route into a Transformer encoder-based generator, a neural model that produces the connecting movement. The route supplies both guidance for generation and information about transition length.
The reported foot-phase comparison showed different total sequence lengths. The full sequence was 64 frames when the second input began with the left foot in contact, and 80 frames when it began with the right foot. The longer case inserted an additional step.
A route through motion clusters
The framework is built in three stages: motion clustering, cluster pathfinding and motion generation. The path found through the cluster graph acts as a guide for the generator and helps determine how long the transition should be.
The main experiments trained the framework on LAFAN1. The authors excluded sequences with obstacles, mirrored the remaining sequences for augmentation, used Subject 5 for validation and used the other subjects for training.
For testing, the paper reports 66,780 within-LAFAN1 input pairs and 10,000 cross-category LAFAN1 pairs. A Mixamo + 100Style set was used to examine transitions between unseen motion sequences.
What the comparisons found
In a component-removal test, the full model recorded L2P 0.436 and L2Q 0.909. Without the guide, those values were 0.567 and 0.972. Without destination loss, they were 0.442 and 0.918. The full model therefore had the lower reported value on both measures than either version with one component removed.
For NMS, the proposed method, the benchmark reported FID 0.20, detail distance 5.18, jitter 382.89, foot sliding 0.23 and ground penetration 0.09 for LAFAN1-within pairs. For LAFAN1-random pairs, the corresponding values were 7.60, 15.86, 385.00, 0.38 and 0.10. On Mixamo + 100Style, the reported FID and detail distance were 14.26 and 26.54.
The authors report the strongest overall comparison performance on LAFAN1-within. On the two more challenging datasets, they report lower jitter for NMS while saying its FID and detail distance remained comparable to the alternatives.
In a qualitative test of generalization, the method produced plausible transitions between two unseen motions.
The authors also ran a user study. It included 20 participants aged 23 to 36, and the average familiarity score was 4.9 on a 7-point scale. Participants rated naturalness in 10 test cases, and the authors report that NMS received the highest score across the methods.
On 1,000 LAFAN1-within test inputs, the average time for estimating length was 6.3 milliseconds and the average motion-generation time was 8.2 milliseconds, for a reported total of 14.5 milliseconds.
The trade-offs at the edges
Scaling up the graph increased the reported resources needed for pathfinding. Across progressively larger dataset configurations, codebook size was 2,048, 4,096 and 8,192; pathfinding time was 0.037, 0.074 and 0.207 seconds; and graph memory was 25.213, 82.273 and 290.952 megabytes.
The paper reports a trade-off around transition length. Output length generally increased as the two inputs became more dissimilar, while motion quality declined for longer outputs. For outputs in the longest reported bin, 128 to 176 frames, the reported values were cosine distance 0.80, pose difference 1.83, FID 12.09, detail distance 17.89, jitter 374.95, foot sliding 0.75 and ground penetration 0.14.
The system also showed a specific failure at the first boundary between clips. The authors report popping when inputs depart from the training distribution, especially for rare poses without two-foot support or for motions from unseen datasets. Because the framework has no physics constraints, they also report physically implausible smoothing, including hips lifting without appropriate support.
The work is an arXiv version 1 preprint dated 29 September 2026. Its reported evidence comes from computational comparisons, a qualitative unseen-motion test and a 20-person naturalness study.
The acknowledgments say the research was supported by a Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency and funded by the Ministry of Culture, Sports and Tourism in 2024 under grant RS-2024-00440434.
Paper data and sources
Original title: Length-varying Neural Motion Stitching via Cluster Transition Graph
Authors: Haemin Kim, Junghyun Nam, Seokhyeon Hong et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-29
DOI: Not available
Original paper · Full text