Can Text-to-Video Models Generate Responsible Realistic Human Motion?

Zhengqing Yuan, Xiaoyu Ma, Yunhong He, Zixuan Weng, Rong Zhou, Weixiang Sun, Lifang He, Lichao Sun, Yanfang Ye

Keywords: Benchmarking, Human Motion Generation, Video Evaluation

Abstract: Recent advances in text-to-video (T2V) generation have yielded impressive progress in resolution, duration, and prompt fidelity, with models such as Pika, Gen-3, and Sora producing clips that appear compelling at first glance. Yet, in everyday use and public demos, generated people often “look right but move wrong,” exhibiting artifacts like foot sliding, joint hyperextension, and desynchronized limbs. Such failures are not cosmetic: 1) unsafe motions can be copied by viewers, especially juveniles, raising injury risks; 2) in clinical and sports contexts, implausible kinematics corrupt analytics for angle, cadence, and phase, causing misdiagnosis and unsafe return-to-play; and 3) in simulation pipelines, non-physical motion distributions contaminate training and evaluation, degrading sim-to-real transfer. However, existing benchmarks remain inadequate: 1) they lack kinematics awareness, rewarding visual resemblance while joint trajectories violate physiological ranges; 2) they lack rhythm- and body-level temporal metrics, overlooking gait-cycle timing, symmetry, and inter-limb coordination; and 3) they fail to disentangle camera from body motion, letting pans and zooms mask biomechanical errors. To address these gaps, we present \textbf{Movo}, the first kinematics-centric benchmark for T2V motion realism. Movo unifies three components: 1) a posture-focused dataset with camera-aware prompts that isolate representative upper- and lower-body actions; 2) skeletal-space metrics, Joint Angle Change (JAC), Dynamic Time Warping (DTW), and Motion Consistency Metric (MCM), that operationalize biomechanical plausibility across joints, rhythms, and constraints; and 3) human validation studies that calibrate thresholds and show strong correlation between skeletal scores and perceived realism. Evaluating 14 leading T2V models reveals persistent gaps: some excel in specific motions but struggle with cross-action consistency, and performance varies widely between open-source and proprietary systems. Movo provides a rigorous, interpretable foundation for improving human motion generation and for integrating biomechanical realism checks into model development, selection, and release workflows. The code and scripts are available at Supplementary Material.

Submission Number: 94