MotionInsight: Diagnosing Object Motion Deficiencies in Generated Videos

Jiahao Zhan; Yongrui Ma; Qunliang Xing; Xuanyu Zhang; Jingqi Tong; Junlin Li; Li zhang; Shijie Zhao

cs.CVSep 29, 2026

MotionInsight: Diagnosing Object Motion Deficiencies in Generated Videos

Authors: Jiahao Zhan, Yongrui Ma, Qunliang Xing, Xuanyu Zhang, Jingqi Tong, Junlin Li, Li zhang, Shijie Zhao, +1 more

Abstract

Despite rapid progress in video generation models, they still exhibit obvious motion deficiencies, often manifested as incorrect object motion. However, most existing video quality evaluations focus on aesthetic quality or text-video alignment. To address this gap, we study object-centric motion fidelity assessment, evaluating target objects along object consistency, motion continuity, and physical plausibility. To achieve this, we first introduce VidMotion, a diagnostic dataset of 6,879 videos with designated moving objects and fine-grained annotations including dimension-wise scores and failure causes. We further propose MotionInsight, a diagnostic evaluator that shifts assessment from implicit RGB-frame observation to explicit motion-space diagnosis. By constructing motion-aware representations, MotionInsight makes subtle motion deficiencies more observable. We also introduce motion-specific rewards during GRPO to transform observed motion into a diagnostic assessment. Experiments demonstrate that MotionInsight provides an effective basis for diagnosing object motion deficiencies, producing human-aligned scores along three dimensions and grounded explanations.

Figures & tables

Dim.	Description
OC	Stable identity, appearance, and structure during motion.
MC	Smooth and continuous trajectory without abrupt jumps or jitter.
PP	Motion consistent with forces and basic physical dynamics.

Table 1: Definitions of the three dimensions in object-centric motion fidelity assessment: object consistency (OC), motion continuity (MC), and physical plausibility (PP).

Open-source models
Model	Object Consistency $\uparrow$	Motion Continuity $\uparrow$	Physical Plausibility $\uparrow$
Wan 2.2 Wan et al. (2025)	$3.420\pm 0.441$	$3.634\pm 0.530$	$2.355\pm 0.261$
LTX 2.1 HaCohen et al. (2024)	$1.978\pm 0.690$	$3.512\pm 0.546$	$2.930\pm 0.402$
LongCat-Video Team et al. (2025)	$1.335\pm 0.214$	$2.242\pm 0.699$	$1.821\pm 0.582$
Cosmos-Predict 2.5 Ali et al. (2025)	$2.088\pm 0.649$	$3.109\pm 0.679$	$2.156\pm 0.472$
HunyuanVideo-1.5 Kong et al. (2024)	$3.223\pm 0.504$	$3.939\pm 0.672$	$2.370\pm 0.274$

Table 2: Comparison of video generation models and real videos on object consistency, motion continuity, and physical plausibility, evaluated on the 166 challenging prompts in VidMotion-Test. Values are reported as mean

\pm

standard deviation.

Figure 2: Overview of MotionInsight. Given an input video and a target prompt, we uniformly sample RGB frames as the standard visual input. Then, we extract motion features from the frames using ViPE and CoTracker3, and aggregate them into a motion-aware representation, which is fed into the VLM.

Figure 3: Qualitative result of MotionInsight. The reasoning texts are presented below the video frames, and the scores for the three dimensions are visualized in the radar chart at the top right.

	Object Consistency			Motion Continuity			Physical Plausibility
Method	SRCC $\uparrow$	PLCC $\uparrow$	KRCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	KRCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	KRCC $\uparrow$
Independent Human	0.732	0.774	0.671	0.614	0.684	0.580	0.766	0.804	0.692
GPT-5.4	0.370	0.400	0.322	0.276	0.303	0.243	0.246	0.257	0.203
Gemini 3.1 Pro	0.333	0.318	0.281	0.218	0.217	0.193	0.166	0.162	0.144
Qwen-3-VL-8B	0.074	0.092	0.065	0.055	0.062	0.051	-0.009	0.020	-0.008
Qwen-3-VL-8B-FT	0.286	0.301	0.242	0.213	0.238	0.190	0.181	0.205	0.154

Table 3: Correlation with human annotations across three dimensions. Baselines include GPT-5.4 Hurst et al. (2024) , Gemini 3.1 Pro Team et al. (2024) , Qwen-3-VL-8B Bai et al. (2025) , Qwen-3-VL-8B-FT (supervised fine-tuning on VidMotion-Train), VideoPhy2 Bansal et al. (2025) , VMBench Ling et al. (2025) , and WorldModelBench Li et al. (2025) . Independent Human denotes a single volunteer evaluator who did not participate in dataset annotation or calibration, while the benchmark labels are aggregated from three calibrated annotators using MOS.

Method	Jaccard $\uparrow$	Prec. $\uparrow$	Rec. $\uparrow$
GPT-5.4	0.15	0.29	0.19
Gemini 3.1 Pro	0.12	0.24	0.16
Qwen-3-VL-8B	0.09	0.20	0.12
Ours	0.58	0.74	0.63

Table 4: Failure-cause grounding of diagnostic reasoning. We compare GPT-5.4 Hurst et al. (2024) , Gemini 3.1 Pro Team et al. (2024) , Qwen-3-VL-8B Bai et al. (2025) , and our MotionInsight. We parse each model’s reasoning into 12 predefined failure causes and compare the parsed causes with human annotations using Jaccard similarity, precision, and recall.

Figure 4: Sensitivity to localized motion failures. We divide each video into $K$ temporal clips and compute the gap between the minimum clip-level score and the human full-video score. MotionInsight keeps a consistently small gap across $K$ , indicating better sensitivity.

Method	#Fr.	OC $\uparrow$	MC $\uparrow$	PP $\uparrow$
Uniform Sampling	16	0.510	0.455	0.445
2 $\times$ Uniform Sampling	32	0.435	0.382	0.421
4 $\times$ Uniform Sampling	64	0.406	0.334	0.397
Trajectory Overlay	16	0.475	0.438	0.430
Motion-Aware Rep.	16	0.761	0.633	0.727

Table 5: Ablation on video perception strategies under the same GRPO training, reported with SRCC. Rep. denotes representations. #Fr. refers to the number of sampled RGB frames, OC: Object Consistency, MC: Motion Continuity, and PP: Physical Plausibility.

Setting	OC $\uparrow$	MC $\uparrow$	PP $\uparrow$
Semantic Alignment Only	0.253	0.183	0.224
+ Scoring Reward	0.703	0.588	0.670
+ Failure-cause Reward	0.761	0.633	0.727

Table 6: Ablation on the GRPO design of MotionInsight, reported with SRCC. OC: Object Consistency, MC: Motion Continuity, and PP: Physical Plausibility.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Method	OC $\uparrow$	MC $\uparrow$	PP $\uparrow$
Motion-Aware Rep.	0.761	0.633	0.727
w/ Temporal Shuffle	0.161	0.109	0.113
w/o Camera Motion	0.697	0.580	0.683

Appendix

Table 7: Ablation on the design of motion-aware representations for GRPO, reported with SRCC. Rep. denotes representations. w/ Temporal Shuffle randomly permutes the temporal order of the motion representations, while w/o Camera Motion removes camera-motion tokens and keeps only object-motion features. OC: Object Consistency, MC: Motion Continuity, and PP: Physical Plausibility.

Figure 5: Distribution of target objects in VidMotion. The objects are grouped into six coarse categories: Human, Animal, Vehicle, Toy and Ball, Tool, and Other. The distribution highlights the diversity of object-centric motion scenarios covered by the dataset.

Dimension	Krippendorff’s $\alpha$
Object Consistency	0.7981
Motion Continuity	0.6846
Physical Plausibility	0.7432

Appendix

Table 8: Inter-annotator reliability for the three scalar motion-quality dimensions in VidMotion. Krippendorff’s

\alpha

is used to measure agreement among annotators, with higher values indicating stronger consistency.

Method	Motion Fidelity	Overall Quality
over Wan 2.1	92.86%	86.67%
over VideoPhy2-DPO	87.56%	87.50%

Appendix

Table 9: 2AFC human study results for DPO-based video generation. We compare the model trained with MotionInsight-derived preference pairs against the original Wan2.1 and the VideoPhy2-DPO Bansal et al. (2025) baseline.

Figure 6: Qualitative comparison of DPO-based video generation.

Figure 7: Interface of the annotation website.

Figure 8: Distribution of annotated issue categories. The 12 cause candidates are grouped into three dimensions: Object Consistency, Motion Continuity, and Physical Plausibility.

Figure 9: Additional qualitative results of MotionInsight. For each example, we show the sampled video frames, the model’s reasoning for the designated object, and the predicted scores across the three motion-quality dimensions.

Figure 10: Overview of the VidMotion construction and annotation pipeline. We collect real videos with prominent object motion, extract captions and target objects, generate corresponding videos using nine open-source and proprietary models, and collect expert annotations for both real and generated videos, including three-dimensional scores, failure causes, and confidence levels.

Explore similar work

HumanScore: Benchmarking Human Motions in Generated Videos
Apr 22, 2026Yusu Fang, Tiange Xiang, Tian Tan +4Generative Video Models Human Motion Generation
Stanford University · Peking University
Generative Action Tell-Tales: Assessing Human Motion in Synthesized Videos
Dec 1, 2025Xavier Thomas, Youngsun Lim, Ananya Srinivasan +2Human Motion Generation Human Motion
Boston University · Belmont High School · Canyon Crest Academy +1
Quantitative Video World Model Evaluation for Geometric-Consistency
May 14, 2026Jiaxin Wu, Yihao Pi, Yinling Zhang +2Generative Video Models Video World Models
1Tsinghua University - IEI Lab · 2UW-Madison · 3Adobe Research

Apr 22, 2026, cs.CVJ/K move · Enter open · S save

HumanScore: Benchmarking Human Motions in Generated Videos

Yusu Fang, Tiange Xiang, Tian Tan, Narayan Schuetz+3

Stanford University · Peking University

Generative Video Models Human Motion Generation Video Generation Human Motion Physical Plausibility Biomechanics

Open paper AskPDF arXiv

Recent advances in model architectures, compute, and data scale have driven rapid progress in video generation, producing increasingly realistic content. Yet, no prior method systematically measures how faithfully these systems render human bodies and motion dynamics. In this paper, we present HumanScore, a systematic framework to evaluate the quality of human motions in AI-generated videos. HumanScore defines six interpretable metrics spanning kinematic plausibility, temporal stability, and biomechanical consistency, enabling fine-grained diagnosis beyond visual realism alone. Through carefully designed prompts, we elicit a diverse set of movements at varying intensities and evaluate videos generated by thirteen state-of-the-art models. Our analysis reveals consistent gaps between perceptual plausibility and motion biomechanical fidelity, identifies recurrent failure modes (e.g., temporal jitter, anatomically implausible poses, and motion drift), and produces robust model rankings from quantitative and physically meaningful criteria.

Abstract

Explore similar work

HumanScore: Benchmarking Human Motions in Generated Videos

Generative Action Tell-Tales: Assessing Human Motion in Synthesized Videos

Quantitative Video World Model Evaluation for Geometric-Consistency