One video can animate many animal and human characters—about 20 times faster

A new system predicts motion for different digital skeletons while reducing errors in the characters’ joint rotations.

Editorial illustration — not from the study.

The system predicts 3D joint positions from video and then converts them into rotations for a target character. Unlike earlier pipelines that use a separate mathematical solver for this conversion, both stages are learned together. A rest pose and one example pairing of a pose with its rotations provide information about the target skeleton’s coordinate system.

Tests on the Truebones Zoo and Objaverse collections included skeletons with different structures and conventions. The researchers report lower rotation errors, including 6.54 degrees on unseen skeletons, and qualitative results on internet videos. The system does not use a mesh as an intermediate step, which the researchers say helps it run about 20 times faster than mesh-based pipelines.

Lower errors, faster motion capture

On Truebones Zoo and Objaverse, the system reduced average rotation-angle error from about 17 degrees for the earlier factorized pipeline to about 10 degrees. On unseen skeletons, the reported error reached 6.54 degrees. The system also achieved about 20 times faster inference than mesh-based pipelines. Results shown for Objaverse assets, Truebones Zoo, and in-the-wild internet videos produced motion that the researchers describe as accurate and temporally consistent across varied subjects and capture conditions.

Putting motion on more characters

Different digital characters can have different numbers of joints, joint layouts and local rotation conventions. A joint-position estimate alone cannot fully specify how those joints should rotate. By giving the system a reference pose and its rotations for the target character, the researchers make that conversion learnable and allow the video-to-pose and pose-to-rotation stages to be optimized together. This could make it more practical to transfer motion from a single video to a wider range of animated characters without the extra mesh-based processing used by earlier pipelines.

Evidence and open limits

The evidence comes from experiments on the Truebones Zoo and Objaverse datasets, along with results shown for internet videos. The reported comparisons include rotation errors and inference speed relative to a factorized learned video-to-pose system with an analytical inverse-kinematics stage, and to mesh-based pipelines. The system assumes one foreground subject, little camera movement and does not explicitly handle heavy occlusion. Its rotation decoder depends on motion patterns represented in training: unusual combinations of skeletons and movements can produce implausible rotations even when the predicted joint positions are reasonable. Training coverage also varies by species, with the researchers noting that rarely represented species may have lower rotation quality.

An AI system lets users change a video’s lighting as it plays

Researchers introduced LiveLight, an AI system that relights streaming video in real time while allowing users to adjust a three-dimensional light source. Tests on real-world and synthetic benchmarks found stronger temporal stability, lighting control and user preference than offline alternatives.

10 hours ago·7 views

AI & Computing

AI models shared human values—but not how strongly to rank them

Most of 75 large language models ranked six AI priorities in the same order as 376 human respondents. But some models overstated the gaps between those priorities, and closer agreement with people did not consistently come with greater size, recency or capability.

16 hours ago·8 views

AI & Computing

A client-side AI system blocked 1.2 million malicious email logins

A lightweight language model running on client devices detected brute-force and credential-stuffing attacks against email accounts. Across 2.5 million authentication sessions, it reported 98.4% attack-detection accuracy and blocked more than 1.2 million malicious login attempts during a six-month deployment.

23 hours ago·6 views