Paper deep dive
MotionInsight: Diagnosing Object Motion Deficiencies in Generated Videos
Jiahao Zhan, Yongrui Ma, Qunliang Xing, Xuanyu Zhang, Jingqi Tong, Junlin Li, Li zhang, Shijie Zhao, Tianfan Xue
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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.
Tags
Links
- Source: https://arxiv.org/abs/2609.37030v1
- Canonical: https://arxiv.org/abs/2609.37030v1
Trouble viewing inline? Open PDF directly â
Full Text
57,812 characters extracted from source content.
Expand or collapse full text
MotionInsight: Diagnosing Object Motion Deficiencies in Generated Videos Jiahao Zhan 1,2 , Yongrui Ma 1,2 , Qunliang Xing 2 , Xuanyu Zhang 4 , Jingqi Tong 3 , Junlin Li 2 , Li Zhang 2 , Shijie Zhao 2,â ,B , Tianfan Xue 1,5,B 1 MMLab, CUHK, 2 ByteDance Inc., 3 Fudan University, 4 Peking University, 5 CPII under InnoHK â Project LeadB Correspondence: zhaoshijie.0526@bytedance.com, tfxue@ie.cuhk.edu.hk Wan 2.6 Seedance 2.0 Veo 3.1 The blue car's front/back orientation changes inconsistently. The tennis ball appears to teleport. The volleyball bounces up without being touched. no touch Prompt: A boy repeatedly pushes a volleyball against the wall. Prompt: A person juggles a tennis ball with a tennis racket. Prompt: Two bumper cars collide with each other. Figure 1: Examples of motion deficiencies in generated videos. The left-side sampled frames highlight the object exhibiting the motion failure in each video, while the right-side panels describe the corresponding failures. Abstract Despite rapid progress in video generation models, they still exhibit obvious motion de- ficiencies, often manifested as incorrect ob- ject 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 assess- ment, evaluating target objects along object consistency, motion continuity, and physical plausibility. To achieve this, we first intro- duce 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 evalua- tor that shifts assessment from implicit RGB- frame observation to explicit motion-space di- agnosis. By constructing motion-aware repre- sentations, MotionInsight makes subtle motion deficiencies more observable. We also intro- duce motion-specific rewards during GRPO to transform observed motion into a diagnostic assessment. Experiments demonstrate that Mo- tionInsight provides an effective basis for di- agnosing object motion deficiencies, produc- ing human-aligned scores along three dimen- sions and grounded explanations. The code is publicly available athttps://github.com/ JohnZhan2023/MotionInsight. 1 Introduction Video generation has undergone rapid development in visual quality (Yang et al., 2024), while it may still generate unrealistic motion in dynamic scenar- ios (Bansal et al., 2024). Current video evaluation methods (He et al., 2025) emphasize aesthetic qual- ity and text-video alignment, while paying limited attention to whether objects move in a stable, con- tinuous, and physically reasonable manner. As shown in Figure 1, visually compelling videos may still contain severe motion deficiencies. Such de- ficiencies restrict the applications in film produc- tion (Jiang et al., 2025) and world modeling (Zhan et al., 2026), highlighting the need for a dedicated evaluation task that systematically diagnoses mo- tion deficiencies in generated videos. To this end, we formulate a diagnostic task of object-centric motion fidelity assessment. We fo- cus the evaluation on designated moving objects, including people, animals, vehicles, and everyday physical objects, since they often attract substan- tial visual attention in videos. This object-centric formulation provides a fine-grained basis for diag- nosis by localizing the assessment to a concrete tar- get. We further assess object motion fidelity along three complementary dimensions: object consis- arXiv:2609.37030v1 [cs.CV] 29 Sep 2026 tency, motion continuity, and physical plausibility, whose definitions are shown in Table 1. Beyond predicting dimension-wise scores, a diagnostic as- sessment should provide grounded explanations of concrete motion failures, such as the volleyball in the top row of Figure 1 bouncing upward without any physical contact. To support this study, we build VidMotion, a diagnostic dataset for object-centric motion fidelity assessment. Most prior video evaluation datasets primarily provide holistic quality scores, which of- fer limited interpretability and cannot support fine- grained diagnosis of object motion deficiencies. In contrast, each video in VidMotion designates a moving object to be evaluated and provides diag- nostic annotations that include scores along three dimensions and failure causes when object motion deficiencies occur. In total, VidMotion contains 5,713 generated videos from nine representative models and 1,166 real videos, covering most every- day dynamic scenarios. Based on the human anno- tations, we further reveal a substantial gap between generated and real motion, as shown in Table 2. This gap raises a natural question: Can existing evaluators diagnose object motion deficiencies in generated videos? Unfortunately, as shown in Table 3, existing eval- uators remain limited in identifying subtle motion deficiencies. Although VLMs have strong seman- tic understanding, they typically rely on sampled RGB frames to infer motion. Such RGB-frame in- puts provide only partial temporal observations and contain substantial background redundancy. As a result, subtle motion deficiencies, such as fingers becoming distorted, can be easily diluted by redun- dant visual content and become difficult to capture. This becomes especially problematic for human- aligned evaluation, since human judgments are of- ten driven by the most severe local artifacts (Tu et al., 2020), while VLM-based evaluators may miss them. Such missed evidence further limits their ability to provide grounded explanations. To address these challenges, we propose Motion- Insight, which diagnoses object motion deficiencies in an explicit motion space. Specifically, MotionIn- sight constructs motion-aware representations from object tracking features and global camera motion, rather than solely relying on sampled RGB frames. Such structured motion evidence makes the evalu- ator more sensitive to subtle deficiencies that may be diluted in RGB-frame observations. Moreover, to help the VLM interpret this motion evidence, Dim.Description OCStable identity, appearance, and structure during motion. MCSmooth and continuous trajectory without abrupt jumps or jitter. PPMotion 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 (P). we further perform motion description alignment, which uses automatically generated motion descrip- tions to align the encoded motion embeddings with the VLMâs semantic space. Finally, to turn motion evidence into human-aligned diagnosis, we intro- duce motion-specific rewards for Group Relative Policy Optimization (GRPO) (Guo et al., 2025), derived from multi-dimensional scores and failure- cause supervision. These rewards align the evalua- torâs scoring criteria with human judgments while enabling diagnostic insights into concrete object motion failures. Experiments show that MotionInsight aligns well with human judgments, producing accurate dimension-wise scores, grounded diagnostic rea- soning, and strong sensitivity to localized motion failures. 2 Related Work Video motion evaluation and benchmarks. Ex- isting benchmarks for video generation mainly focus on overall visual quality and text-to- video alignment, such as VBench (Huang et al., 2024), VideoScore2 (He et al., 2025), and T2V- CompBench (Sun et al., 2025). To address the eval- uation of motion, several studies incorporate low- level or structured signals. For example, VBench uses optical-flow-based metrics, but optical flow may fail to maintain reliable temporal correspon- dence when the target object suddenly disappears. VMBench (Ling et al., 2025) evaluates motion con- sistency via rule-based filtering of tracks, while Hu- manScore (Fang et al., 2026) leverages 3D human body models to assess human motion realism. Sim- ilarly, WorldScore (Duan et al., 2025) introduces Structure-from-Motion (Schönberger and Frahm, 2016) to measure whether generated videos follow physical camera constraints. These approaches rely heavily on priors for specific tasks, which makes ModelObject ConsistencyâMotion ContinuityâPhysical Plausibilityâ Open-source models Wan 2.2 (Wan et al., 2025) 3.420± 0.441 3.634± 0.530 2.355± 0.261 LTX 2.1 (HaCohen et al., 2024) 1.978± 0.690 3.512± 0.546 2.930± 0.402 LongCat-Video (Team et al., 2025) 1.335± 0.214 2.242± 0.699 1.821± 0.582 Cosmos-Predict 2.5 (Ali et al., 2025) 2.088± 0.649 3.109± 0.679 2.156± 0.472 HunyuanVideo-1.5 (Kong et al., 2024) 3.223± 0.504 3.939± 0.672 2.370± 0.274 Proprietary models Wan 2.6 (Wan et al., 2025) 3.577± 0.426 3.746± 0.644 2.461± 0.314 Sora 2 (Brooks et al., 2024) 3.792± 0.375 4.113± 0.462 3.507± 0.270 Veo 3.1 (Wiedemer et al., 2025) 3.189± 0.532 3.636± 0.698 2.250± 0.382 Seedance 2.0 (Seedance et al., 2026) 3.810± 0.379 4.143± 0.573 3.081± 0.250 Real Video 4.740± 0.269 4.915± 0.265 4.814± 0.112 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± standard deviation. them inherently constrained to narrow domains and difficult to generalize to broader video generation scenarios. Other studies evaluate physical plausi- bility or world modeling capabilities of video gen- erators (Meng et al., 2024; Hu et al., 2025), often relying on predefined rules or constrained scenarios. As a result, they cannot serve as general evaluators for diagnosing object motion deficiencies. VLM-based evaluators for video generation. Re- cent works have explored VLM-based evaluators for video generation (Qin et al., 2024; Bansal et al., 2024; Zhang et al., 2026b; Zhao et al., 2026; Wang et al., 2025; Wu et al., 2022), leveraging the strong generalization ability of VLMs. Some recent works further improve video perception by strengthening temporal attention (Motamed et al., 2025) or re- ducing redundant visual tokens (Shi et al., 2026). Nevertheless, these methods still rely solely on sam- pled RGB frames to perceive motion. Many object motion deficiencies manifest as subtle pixel-level changes in RGB space and can be easily diluted during visual-token aggregation. In contrast, Mo- tionInsight models the object motion in motion space, enabling greater sensitivity to local artifacts. 3 VidMotion Dataset VidMotion is a diagnostic dataset designed for object-centric motion fidelity assessment. Unlike prior evaluation datasets that mainly provide over- all scores for physical plausibility, artifact severity, or video realism (Bansal et al., 2025; Zhang et al., 2026a), VidMotion evaluates the motion fidelity of a designated object. Each video is associated with a designated moving object and annotated with three-dimensional motion fidelity scores and fine-grained failure causes. Together, these annota- tions provide a strong basis for diagnosing object motion deficiencies by localizing to a concrete tar- get, explaining the failure cause, and measuring its severity along different motion dimensions. Since our focus is perceptual motion fidelity rather than solver-based correctness, the annotations aim to capture whether the target objectâs motion appears realistic from human experience. So real videos are evaluated under the same human annotation proto- col rather than assigned perfect scores by default, enabling direct comparison between generated and natural object motion. In total, VidMotion contains 6,879 annotated videos. Figure 10 summarizes the overall data construction and annotation pipeline. 3.1 Dataset Construction To build VidMotion,we start from real- worldvideoscollectedfromopen-source datasets (Caelles et al., 2019; Wu et al., 2016; Chow et al., 2025; Kay et al., 2017; Huang et al., 2019; Motamed et al., 2025) and additional curated sources, prioritizing samples in which one salient moving object dominates the motion.After screening by five human experts, we obtain a final set of 1,166 real videos. We then derive textual prompts and targets for these videos through captioning with Gemini 3.1 Pro (Team et al., 2024), followed by human verifi- cation. Using the resulting prompts, we generate corresponding videos with a diverse set of video generation models, including open-source models such as Wan 2.2 (Wan et al., 2025), LTX 2.1 (Ha- Cohen et al., 2024), LongCat-Video (Team et al., 2025), Cosmos-Predict 2.5 (Ali et al., 2025), and HunyuanVideo-1.5 (Kong et al., 2024), as well as proprietary models including Wan 2.6 (Wan et al., 2025), Sora 2 (Brooks et al., 2024), Veo 3.1 (Wiede- mer et al., 2025), and Seedance 2.0 (Seedance et al., 2026). For proprietary models, we generate videos only for a challenging subset of 166 prompts manu- ally selected from the 1,166 prompts. Since videos from these proprietary models are not included in VidMotion-Train, this design makes VidMotion- Test more challenging and allows us to evaluate generalization to unseen generators. In this way, each prompt is paired with a corre- sponding real video and multiple generated videos from different models. After filtering out invalid samples with severe degeneration or unusable con- tent, we use the 1,386 videos associated with the 166 challenging prompts as VidMotion-Test, which serves as our benchmark split. The remaining 5,493 videos are used as VidMotion-Train. 3.2 Human Annotation We collect human annotations on the motion fi- delity of the designated object along three dimen- sions: object consistency, motion continuity, and physical plausibility. For videos judged to contain motion deficiencies, annotators further select one or more applicable failure causes from 12 predefined candidates, which provide diagnostic explanations for the underlying motion artifacts. In total, 21 annotators participated in the an- notation process. Before annotation, all annota- tors underwent training, completed a trial annota- tion assessment, and reviewed representative exam- ples with reference labels. Each video is indepen- dently evaluated by three annotators, each of whom provides three-dimensional scores, a confidence level, and applicable failure causes. Krippendorffâs α(Krippendorff, 2011) averaged over the three dimensions reaches 0.7420, indicating reliable an- notations. More annotation details are provided in Appendix A. We aggregate the three annotatorsâ scores using the Mean Opinion Score (MOS) to obtain the final three-dimensional motion fidelity scores, and take the intersection of their selected failure causes as the final multi-label failure-cause annotations to ensure label reliability. Additional analysis of VidMotion is included in Appendix B. 4 MotionInsight Our goal is to assess the motion fidelity of a desig- nated object in a video. Given a videov =f t T t=1 and a target object prompto, our evaluatorE outputs a three-dimensional motion score vector and corresponding diagnostic reasoning,(s, r) = E(v,o). Here,scovers object consistency, motion continuity, and physical plausibility. For VLM- based evaluators, r corresponds to the explicit rea- soning text generated before the final scores to jus- tify the scoring results. In the benchmark setting, the target object promptois provided by the dataset annotation or specified by the user. When no target object is given, such as in reward-model applica- tions for video generation, we first prompt a VLM to identify salient moving objects in the video, ap- ply MotionInsight to each identified object, and aggregate the resulting object-centric scores into a video-level signal, as described in Appendix F. Since RGB-frame observations provide only sparse temporal information and contain substantial background redundancy, subtle motion deficiencies can be difficult to perceive. Our key design princi- ple is therefore to explicitly represent the complete motion in a structured motion space. Based on this motion evidence, MotionInsight performs human- aligned diagnosis through semantic alignment with the VLM and preference alignment with human judgments. As illustrated in Figure 2, MotionIn- sight complements sampled frames with motion- aware representations, making object motion defi- ciencies observable (Section 4.1). Next, we align these motion embeddings with the VLMâs seman- tic space using automatically generated motion de- scriptions (Section 4.2). Finally, we design motion- specific rewards for GRPO to align the resulting diagnosis with human judgments (Section 4.3). 4.1 Motion-Aware Representations Motion-aware representations aim to make object motion deficiencies explicit beyond RGB-frame observations. However, the motion in a video is entangled with both object motion and camera mo- tion. We therefore fuse object motion with camera poses to form motion embeddings, enabling a dis- entangled understanding of object dynamics. Specifically, given a videov = f t T t=1 and a text promptospecifying the target object, we first use SAM3 (Carion et al., 2025) to obtain an initialization mask for the target object: m = SAM3(v,o).(1) Using this mask, we sample query pointsP and track them throughout the video using Co- ... Frame 1 Frame 16 Frame 8 Frame 9 x z Qwen 3 VL ... Motion Feature Aggregation Vision Encoder ... Cam Pose Frame 10Frame 7 Tracking Feature Fusion (Flatten + Concat) Motion Adapter Linear Projection Self- Attention ... Attn Pooling N points K motion tokens ... ... Motion Embeddings ... ... ... ......... ... Interleave ... ... Vision Encoder Frame-1Visual Tokens Vision Encoder Frame-16Visual Tokens 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. Tracker3 (Karaev et al., 2025). This yields frame- wise point-level tracking featuresX = x t T t=1 , where eachx t âR NĂd o andNis the number of sampled tracked points. Since the number of tracked pointsNvaries with the size ofm, we use an attention pooling module withKlearnable queries to aggregate the point- level features into a fixed number of object-motion tokens: H t = AttnPool(x t )âR KĂd m .(2) In parallel, we feed the entire video into ViPE (Huang et al., 2025) to estimate the camera poses for all frames: C =c t T t=1 = ViPE(v),(3) where eachc t = [r t ;Ï t ]consists of a 6D rotation parameterr t âR 6 and a 3D translation parameter Ï t âR 3 for framef t . These camera poses provide a global motion reference that helps separate object motion from viewpoint changes. For each frame, we construct a motion represen- tation by flattening theKobject-motion tokens and concatenating them with the corresponding camera poses: z t = [vec(H t ); c t ]âR Kd m +9 .(4) The resulting sequenceZ =z t T t=1 is processed by a lightweight Motion Adapter, consisting of a linear projection and a self-attention layer, to produce frame-wise motion embeddings. As shown in Figure 2, we uniformly sample frames from the video and interleave them with the corresponding motion embeddings produced by the Motion Adapter. This interleaved sequence enables the VLM to jointly reason over appearance information and structured motion representations. 4.2 Motion Description Alignment Although the motion-aware representations encode object motion and camera poses, they are not di- rectly interpretable by the frozen VLM. We there- fore align the extracted tracking features and cam- era poses with the VLMâs semantic space using automatically generated motion descriptions as su- pervision. To obtain supervision without additional human annotations, we design 16 questions about the targetâs motion. Each video is divided into clips of 16 consecutive frames, which are fed sequen- tially to the VLM (Bai et al., 2025) to generate clip-level descriptions. These descriptions are then summarized into a video-level motion description. Following this procedure, we collect 80,721 question-answer pairs from OpenVid (Nan et al., 2024) for semantic alignment. We then fine-tune the model with this motion-description supervi- sion for semantic alignment. During this stage, we freeze the VLM backbone and optimize only the motion encoding components, including the attention pooling module and the Motion Adapter. 4.3 GRPO for Human Preference Alignment We adopt GRPO (Guo et al., 2025) with two motion-specific rewards: a multi-dimensional re- ward and a failure-cause reward.The multi- dimensional reward aligns MotionInsight with hu- man judgments by encouraging accurate predic- tion of annotated scores along object consistency, motion continuity, and physical plausibility. The failure-cause reward further refines its diagnostic reasoning by encouraging the evaluator to associate observed motion deficiencies with interpretable failure causes. Built on the richer and more ex- plicit motion evidence provided by motion-aware representations, the motion-specific rewards guide MotionInsight to learn human-aligned scoring and grounded reasoning. Together, these two capabili- ties constitute the core of a diagnostic evaluator for object-centric motion fidelity. Multi-dimensional scoring reward. For the scor- ing reward, we supervise the model using the three- dimensional scores in VidMotion-Train. Given a queryq, the model predicts scores forMevalua- tion dimensionsËs j M j=1 . To discourage overcon- fident scoring, we use a normalized asymmetric error penalty that penalizes overestimation more heavily than underestimation: r score = 1â M X j=1 λ j clip(d 2 j , 0, 1),(5) where d j = ïŁ± ïŁŽ ïŁŽ ïŁČ ïŁŽ ïŁŽ ïŁł α |Ës j â s gt j | 4 , if Ës j > s gt j , |Ës j â s gt j | 4 ,otherwise. (6) Here,s gt j denotes the ground-truth MOS for thej- th dimension,λ j is its weight, andα > 1controls the additional penalty for overestimation. Failure-cause reward. To improve MotionIn- sightâs ability to identify motion deficiencies and provide grounded reasoning, we introduce a failure- cause reward using the failure-cause labels from VidMotion-Train. Given four candidate failure de- scriptions, the model selects the one that best ex- plains the observed issue. We assign a reward of 1 if the selected option matches the annotated failure causes and 0 otherwise. 5 Experiments Dataset and metrics. We use 5K videos from OpenVid (Nan et al., 2024) to construct 80,721 QA pairs for semantic alignment. In the GRPO stage, VidMotion-Train is used to align the evaluator with human judgments, while VidMotion-Test is used to assess the scoring and diagnostic reasoning of Mo- tionInsight. We use PLCC, SRCC, and KRCC to measure the correlation between predicted scores and human annotations. For diagnostic reasoning, we parse each modelâs reasoning texts into 12 pre- defined failure causes using GPT-5.1 and evaluate their consistency with annotated failure causes us- ing Jaccard similarity, precision, and recall. Baselines. To evaluate both the scoring and di- agnostic reasoning capabilities of MotionInsight, we compare it with several general-purpose VLM evaluators, including GPT-5.4 (Hurst et al., 2024), Gemini 3.1 Pro (Team et al., 2024), and Qwen-3- VL-8B (Bai et al., 2025). To ensure a fair com- parison, we carefully design and validate a unified prompt for both GRPO training and VLM baselines. The prompt explicitly specifies the designated tar- get object and the three evaluation dimensions, re- quiring each evaluator to reason first and then out- put continuous scores in a consistent format. The full prompt template is provided in Appendix C. To provide a stronger VLM baseline, we further fine-tune Qwen-3-VL-8B for score regression on VidMotion-Train using only RGB-frame inputs. For scoring evaluation, we additionally compare with specialized video evaluators. For evaluators specifically designed for physical assessment, such as VideoPhy2 (Bansal et al., 2025) and WorldMod- elBench (Li et al., 2025), we compare their scoring results on Physical Plausibility. For methods specif- ically designed for motion assessment, we compare the Motion Smoothness Score in VMBench (Ling et al., 2025) on Motion Continuity. As a reference, we also report the scoring performance of a single volunteer evaluator on VidMotion-Test. 5.1 Results Qualitative comparisons to baselines. For the qualitative results shown in Figure 3, MotionIn- sight provides diagnostic reasoning from three per- spectives. In the illustrated example, MotionInsight identifies that the foam box becomes shorter after the collision, indicating a structural collapse. Its scores are also better aligned with human judg- ments, as shown in the accompanying radar chart. More qualitative results are shown in Appendix D. Scoring alignment with human judgments. As shown in Table 3, existing evaluators show limited alignment with human annotations on VidMotion- Test, highlighting the difficulty of motion fidelity evaluation for designated targets. VMBench relies on rule-based filtering over pixel-level motion sig- nals, which limits its generalization across diverse dynamic scenarios. Although fine-tuning Qwen-3- VL-8B on VidMotion-Train improves adaptation, Question: Give the target object " foam box " in this video three continuous scores from 1 to 5 for the following aspects in order: object consistency, motion continuity, and physical plausibility. Object Consistency Motion Continuity Physical Plausibility The foam box becomes noticeably shorter after being struck by the sliding box , indicating an object consistency failure. Its motion remains relatively smooth overall, with no obvious abrupt jumps. The overall interaction is generally consistent with physical behavior apart from the structural issue. Motion Continuity Object Consistency Physical Plausibility GPT-5.4 Human MotionInsight 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 ConsistencyMotion ContinuityPhysical Plausibility MethodSRCCâ PLCCâ KRCCâ SRCCâ PLCCâ KRCCâ SRCCâ PLCCâ KRCCâ Independent Human0.7320.7740.6710.6140.6840.5800.7660.8040.692 GPT-5.40.3700.4000.3220.2760.3030.2430.2460.2570.203 Gemini 3.1 Pro0.3330.3180.2810.2180.2170.1930.1660.1620.144 Qwen-3-VL-8B0.0740.0920.0650.0550.0620.051-0.0090.020-0.008 Qwen-3-VL-8B-FT0.2860.3010.2420.2130.2380.1900.1810.2050.154 VideoPhy2------0.3020.3030.247 VMBench---0.0360.0130.032--- WorldModelBench------0.2960.3100.230 Ours0.7610.7500.6670.6330.6250.5630.7270.7180.622 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. MethodJaccardâ Prec. â Rec. â GPT-5.40.150.290.19 Gemini 3.1 Pro0.120.240.16 Qwen-3-VL-8B0.090.200.12 Ours0.580.740.63 Table 4: Failure-cause grounding of diagnostic reason- ing. 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 modelâs reasoning into 12 predefined failure causes and compare the parsed causes with human annotations using Jaccard similarity, precision, and recall. it still generalizes poorly to VidMotion-Test. Other VLM-based evaluators also perform poorly: RGB- frame inputs provide only partial temporal observa- tions and may miss frames containing key artifacts, while many object motion deficiencies remain sub- tle without explicit target motion modeling. By contrast, MotionInsight achieves performance com- Score Gap Number of Temporal Clips K GPT-5.4 Gemini3.1Pro MotionInsight Figure 4: Sensitivity to localized motion failures. We divide each video intoKtemporal clips and compute the gap between the minimum clip-level score and the human full-video score. MotionInsight keeps a consis- tently small gap across K, indicating better sensitivity. parable to a human evaluator across the three di- mensions. We also explore how MotionInsightâs diagnostic scores along three dimensions can be used to optimize video generation in Appendix F. Grounded diagnostic reasoning. As shown in Table 4, existing VLMs exhibit poor diagnostic rea- Method#Fr.OCâMCâPâ Uniform Sampling160.510 0.455 0.445 2Ă Uniform Sampling320.435 0.382 0.421 4Ă Uniform Sampling640.406 0.334 0.397 Trajectory Overlay160.475 0.438 0.430 Motion-Aware Rep.160.761 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 P: Physical Plausibility. SettingOCâMCâPâ Semantic Alignment Only0.2530.1830.224 + Scoring Reward0.7030.5880.670 + Failure-cause Reward0.7610.6330.727 Table 6: Ablation on the GRPO design of MotionIn- sight, reported with SRCC. OC: Object Consistency, MC: Motion Continuity, and P: Physical Plausibility. soning capability. Their explanations frequently miss the fine-grained motion deficiencies identified by human annotators, resulting in low consistency with the annotated failure causes. In contrast, Mo- tionInsight produces more grounded explanations, achieving higher Jaccard similarity, precision, and recall. This suggests that explicitly modeling target motion and incorporating failure-cause supervision help MotionInsight provide more insight into ob- ject motion deficiencies. Sensitivity to localized motion failures. To eval- uate sensitivity to localized motion failures, we select 327 videos containing object consistency de- ficiencies from VidMotion-Test for scoring evalua- tion. These deficiencies are localized to the object and often appear only in a few frames, making them a suitable testbed for localized failure detec- tion. For each video, we divide the full sequence intoKtemporal clips and compare the minimum clip-level score with the human score assigned to the full video. If an evaluator performs better when the video is divided into more clips, it suggests that the evaluator cannot reliably detect localized mo- tion failures when the entire video is provided as input (K = 1), where local artifacts can be diluted by redundant visual tokens. As shown in Figure 4, GPT-5.4 and Gemini 3.1 Pro exhibit a large score gap when evaluating the full video directly, but the gap decreases as the video is divided into more temporal clips. In con- trast, MotionInsight maintains a consistently small score gap across different values ofK. This sug- gests that MotionInsight is more sensitive to lo- calized motion failures, consistent with the human visual system (Tu et al., 2020). 5.2 Ablation Study Motion-awarerepresentationsoutperform RGB-frame sampling. As shown in Table 5, we compare different video input strategies while keeping the GRPO training procedure unchanged. Specifically,we evaluate standard uniform sampling, denser variants with 2Ăand 4Ămore sampled frames, as well as a trajectory-visualized setting, where the target objectâs trajectory is drawn on the sampled frames as visual guidance. The results show that neither increasing the sampling density nor overlaying trajectories leads to clear performance gains. Moreover, denser sampling introduces a higher computational cost while making it harder for the evaluator to focus on localized artifacts, as the critical motion defi- ciencies can be diluted by redundant visual tokens. In contrast, our motion-aware representations consistently achieve the best performance across all three evaluation dimensions, suggesting that structuring target motion in motion space helps the evaluator perceive artifacts that may remain subtle in RGB-frame observations. Additional ablations on the design of motion-aware representations are provided in Appendix G. Motion-specific rewards strengthen the diagnos- tic ability. In Table 6, we further ablate the reward design used in GRPO in terms of scoring perfor- mance. After semantic alignment, the model ex- hibits limited scoring capability without additional preference alignment. Incorporating the scoring re- ward substantially improves its performance, while further adding the failure-cause reward provides more explicit supervision on motion failure causes and strengthens diagnostic reasoning, leading to further gains in human-aligned scoring. 6 Conclusion We formulate the diagnostic task of object-centric motion fidelity assessment, evaluated along object consistency, motion continuity, and physical plausi- bility. To support this study, we introduce VidMo- tion, a diagnostic dataset with multi-dimensional scores and failure causes. Built on VidMotion, we further propose MotionInsight, a diagnostic evalua- tor that shifts the assessment from implicit RGB- frame observations to explicit motion-space diag- nosis. We demonstrate the superior performance of MotionInsight in scoring and interpretable failure diagnosis. Limitations Although VidMotion is designed for object-centric motion fidelity assessment, its scope is mainly lim- ited to entities with clear spatial boundaries, per- sistent identities, and trackable motion trajectories. Therefore, VidMotion and MotionInsight may be less suitable for dynamic phenomena such as fluids, smoke, fire, splashes, or highly deformable materi- als, where object-centric scoring and failure causes can become inherently ambiguous. In addition, using MotionInsight as a reward model may introduce Goodhartâs-law risks. Since the evaluator remains fixed during generator opti- mization and does not automatically evolve with advancing generators, generators may overfit to its scoring patterns and obtain higher rewards without corresponding improvements in real object motion fidelity. Ethical Considerations This work complies with the ACL Ethics Pol- icy. VidMotion is constructed from open-source datasets and curated video sources. All external datasets, models, and tools used in this work are properly cited and used in accordance with their respective licenses and terms of use. Our use of these artifacts is limited to academic research on video generation evaluation, which is consistent with their intended research use, where specified. The artifacts released by this work are intended solely for research purposes. All collected videos are manually screened before annotation to filter out invalid, sensitive, offensive, or privacy-risk con- tent to the best of our ability, including videos con- taining personal information, inappropriate content, or clearly privacy-sensitive visual content. The re- leased annotations focus only on target objects, mo- tion quality scores, confidence levels, and failure causes, and do not include annotator identities or personally identifying information. Acknowledgments The work is supported by the National Key R&D Program of China (No. 2025YFE0201300). References Arslan Ali, Junjie Bai, Maciej Bala, Yogesh Balaji, Aaron Blakeman, Tiffany Cai, Jiaxin Cao, Tianshi Cao, Elizabeth Cha, Yu-Wei Chao, and 1 others. 2025. World simulation with video foundation models for physical ai. arXiv preprint arXiv:2511.00062. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, and 1 others. 2025. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631. Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. 2024. Videophy: Evaluating physical com- monsense for video generation.arXiv preprint arXiv:2406.03520. Hritik Bansal, Clark Peng, Yonatan Bitton, Roman Gold- enberg, Aditya Grover, and Kai-Wei Chang. 2025. Videophy-2: A challenging action-centric physical commonsense evaluation in video generation. arXiv preprint arXiv:2503.06800. Tim Brooks, Bill Peebles, and 1 others. 2024. Video generation models as world simulators. OpenAI Technical Report. Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. 2019.The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv preprint arXiv:1905.00737. Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoub- hik Debnath, Ronghang Hu, Didac Suris, Chai- tanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, and 1 others. 2025. Sam 3: Segment anything with concepts. arXiv preprint arXiv:2511.16719. Wei Chow, Jiageng Mao, Boyi Li, Daniel Seita, Vi- tor Guizilini, and Yue Wang. 2025. Physbench: Benchmarking and enhancing vision-language mod- els for physical world understanding. arXiv preprint arXiv:2501.16411. Haoyi Duan, Hong-Xing Yu, Sirui Chen, Li Fei-Fei, and Jiajun Wu. 2025. Worldscore: A unified evaluation benchmark for world generation. In ICCV. Yusu Fang, Tiange Xiang, Tian Tan, Narayan Schuetz, Scott Delp, Li Fei-Fei, and Ehsan Adeli. 2026. Hu- manscore: Benchmarking human motions in gener- ated videos. arXiv preprint arXiv:2604.20157. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948. Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. 2024. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103. Xuan He, Dongfu Jiang, Ping Nie, Minghao Liu, Zhengxuan Jiang, Mingyi Su, Wentao Ma, Junru Lin, Chun Ye, Yi Lu, and 1 others. 2025. Videoscore2: Think before you score in generative video evaluation. arXiv preprint arXiv:2509.22799. Lanxiang Hu, Abhilash Shankarampeta, Yixin Huang, Zilin Dai, Haoyang Yu, Yujie Zhao, Haoqiang Kang, Daniel Zhao, Tajana Rosing, and Hao Zhang. 2025. Benchmarking scientific understanding and reason- ing for video generation using videoscience-bench. arXiv preprint arXiv:2512.02942. Jiahui Huang, Qunjie Zhou, Hesam Rabeti, Aleksandr Korovko, Huan Ling, Xuanchi Ren, Tianchang Shen, Jun Gao, Dmitry Slepichev, Chen-Hsuan Lin, and 1 others. 2025. Vipe: Video pose engine for 3d geo- metric perception. arXiv preprint arXiv:2508.10934. Lianghua Huang, Xin Zhao, and Kaiqi Huang. 2019. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE transac- tions on pattern analysis and machine intelligence, 43(5):1562â1577. Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianx- ing Wu, Qingyang Jin, Nattapol Chanpaisit, and 1 others. 2024. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807â21818. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276. Zeyinzi Jiang, Zhen Han, Chaojie Mao, Jingfeng Zhang, Yulin Pan, and Yu Liu. 2025. Vace: All-in-one video creation and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17191â17202. Nikita Karaev, Yuri Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. 2025. Cotracker3: Simpler and better point track- ing by pseudo-labelling real videos. In Proceedings of the IEEE/CVF International Conference on Com- puter Vision, pages 6013â6022. Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, and 1 others. 2017. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950. Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, and 1 others. 2024. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603. Klaus Krippendorff. 2011. Computing krippendorffâs alpha-reliability. Dacheng Li, Yunhao Fang, Yukang Chen, Shuo Yang, Shiyi Cao, Justin Wong, Michael Luo, Xiaolong Wang, Hongxu Yin, Joseph E Gonzalez, and 1 oth- ers. 2025. Worldmodelbench: Judging video gen- eration models as world models. arXiv preprint arXiv:2502.20694. Xinran Ling, Chen Zhu, Meiqi Wu, Hangyu Li, Xiaokun Feng, Cundian Yang, Aiming Hao, Jiashu Zhu, Ji- ahong Wu, and Xiangxiang Chu. 2025. Vmbench: A benchmark for perception-aligned video motion generation. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 13087â 13098. Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. 2024. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv preprint arXiv:2410.05363. Saman Motamed, Minghao Chen, Luc Van Gool, and Iro Laina. 2025. Travl: A recipe for making video- language models better judges of physics implausi- bility. arXiv preprint arXiv:2510.07550. Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhen- heng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. 2024. Openvid-1m: A large-scale high- quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371. Yiran Qin, Zhelun Shi, Jiwen Yu, Xijun Wang, Enshen Zhou, Lijun Li, Zhenfei Yin, Xihui Liu, Lu Sheng, Jing Shao, and 1 others. 2024. Worldsimbench: To- wards video generation models as world simulators. arXiv preprint arXiv:2410.18072. Johannes Lutz Schönberger and Jan-Michael Frahm. 2016. Structure-from-motion revisited. In Confer- ence on Computer Vision and Pattern Recognition (CVPR). Team Seedance, De Chen, Liyang Chen, Xin Chen, Ying Chen, Zhuo Chen, Zhuowei Chen, Feng Cheng, Tianheng Cheng, Yufeng Cheng, and 1 others. 2026. Seedance 2.0: Advancing video generation for world complexity. arXiv preprint arXiv:2604.14148. Baifeng Shi, Stephanie Fu, Long Lian, Hanrong Ye, David Eigen, Aaron Reite, Boyi Li, Jan Kautz, Song Han, David M Chan, and 1 others. 2026. Attend before attention: Efficient and scalable video under- standing via autoregressive gazing. arXiv preprint arXiv:2603.12254. Kaiyue Sun, Kaiyi Huang, Xian Liu, Yue Wu, Zihan Xu, Zhenguo Li, and Xihui Liu. 2025. T2v-compbench: A comprehensive benchmark for compositional text- to-video generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8406â8416. Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, and 1 others. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530. Meituan LongCat Team, Xunliang Cai, Qilong Huang, Zhuoliang Kang, Hongyu Li, Shijun Liang, Liya Ma, Siyu Ren, Xiaoming Wei, Rixu Xie, and 1 others. 2025. Longcat-video technical report. arXiv preprint arXiv:2510.22200. Zhengzhong Tu, Chia-Ju Chen, Li-Heng Chen, Neil Birkbeck, Balu Adsumilli, and Alan C Bovik. 2020. A comparative evaluation of temporal pooling meth- ods for blind video quality assessment. In 2020 IEEE international conference on image processing (ICIP), pages 141â145. IEEE. Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, and 1 others. 2025. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314. Yibin Wang, Zhimin Li, Yuhang Zang, Chunyu Wang, Qinglin Lu, Cheng Jin, and Jiaqi Wang. 2025. Unified multimodal chain-of-thought reward model through reinforcement fine-tuning. arXiv preprint arXiv:2505.03318. ThaddĂ€us Wiedemer, Yuxuan Li, Paul Vicol, Shixi- ang Shane Gu, Nick Matarese, Kevin Swersky, Been Kim, Priyank Jaini, and Robert Geirhos. 2025. Video models are zero-shot learners and reasoners. arXiv preprint arXiv:2509.20328. Haoning Wu, Chaofeng Chen, Jingwen Hou, Liang Liao, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. 2022. Fast-vqa: Efficient end-to-end video quality as- sessment with fragment sampling. In European con- ference on computer vision, pages 538â554. Springer. Jiajun Wu, Joseph J Lim, Hongyi Zhang, Joshua B Tenenbaum, and William T Freeman. 2016. Physics 101: Learning physical object properties from unla- beled videos. In BMVC, volume 2, page 7. Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, and 1 oth- ers. 2024.Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072. Jiahao Zhan, Zizhang Li, Hong-Xing Yu, and Jiajun Wu. 2026. Perpetualwonder: Long-horizon action- conditioned 4d scene generation. arXiv preprint arXiv:2602.04876. Qin Zhang, Peiyu Jing, Hong-Xing Yu, Fangqiang Ding, Fan Nie, Weimin Wang, Yilun Du, James Zou, Jiajun Wu, and Bing Shuai. 2026a. Physion-eval: Evaluat- ing physical realism in generated video via human reasoning. arXiv preprint arXiv:2603.19607. Xuanyu Zhang, Weiqi Li, Shijie Zhao, Junlin Li, Li Zhang, and Jian Zhang. 2026b. Vq-insight: Teach- ing vlms for ai-generated video quality understand- ing via progressive visual reinforcement learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 12870â12878. Shijie Zhao, Xuanyu Zhang, Weiqi Li, Junlin Li, Li Zhang, Tianfan Xue, and Jian Zhang. 2026. Rea- soning as representation: Rethinking visual reinforce- ment learning in image quality assessment. Proceed- ings of the International Conference on Learning Representations (ICLR). A Annotation Details Annotators were recruited from students and re- searchers with experience in video understanding. All 21 annotators had at least a bachelorâs degree. All annotators were fairly compensated for their work. Before participation, annotators were in- formed that their annotations would be used for academic research on motion fidelity assessment, and they provided consent to participate. Figure 7 shows the interface of the annotation website. As shown in Table 8, Krippendorffâsα reaches 0.7981, 0.6846, and 0.7432 for object con- sistency, motion continuity, and physical plausibil- ity, respectively, indicating reliable inter-annotator agreement across all three scalar dimensions. For the multi-label failure-cause annotations, we com- pute pairwise Jaccard similarity among annotators by treating each annotatorâs selected causes as a set. The resulting Jaccard similarity reaches 0.61, sug- gesting reasonable agreement despite the inherent ambiguity of fine-grained motion artifacts. B VidMotion Analysis As shown in Figure 5, VidMotion shows great di- versity in target objects, covering most dynamic ob- jects commonly encountered in daily life. We also analyze the failure cause annotations in Figure 8. For object consistency, we find that deformation of humans or objects caused by fast motion is the most common issue. For motion continuity, generated object motions are still often stiff and unnatural. For physical plausibility, current video generation models still struggle with physical causality, fre- quently producing misaligned motion for observed force. C Prompt Template We present the prompt used for GRPO training and VLM baselines in Table 10. D More Qualitative Results Figure 9 shows two additional examples. In the field hockey case, the ball suddenly flashes back after being hit, which reflects a localized continuity failure. In the skateboard case, the object remains structurally stable but starts moving without a clear applied force, revealing failure in physical plau- sibility. These results suggest that MotionInsight exhibits reasoning ability for motion, enabling it to evaluate object motion progressively from ap- pearance stability to temporal continuity and even physical fidelity. E Implementation Details We use Qwen-3-VL-8B-Instruct (Bai et al., 2025) as the pretrained VLM backbone. For modal- ity alignment, we fine-tune the motion encoding components for 3 epochs on 8 NVIDIA A100 GPUs with a learning rate of1 Ă 10 â6 . For GRPO training, we set the number of sampled re- sponses toN = 8and the KL penalty weight to ÎČ = 0.001. The model is then trained for 5 epochs on 8 NVIDIA A100 GPUs with a learning rate of 1Ă 10 â6 . In the motion-aware representations, the number of learnable queriesKis set to 8. For the multi-dimensional scoring reward, the asymmetric penalty coefficientαis set to 1.5. For RGB-frame inputs, MotionInsight samples frames at a stride of 16 frames. F MotionInsight for Generation We further explore the potential of MotionInsight as a reward for improving video generation through preference optimization. We use Wan2.1-T2V- 1.3B (Wan et al., 2025) as the base generator. For each prompt, we sample eight candidate videos and use Qwen-3-VL-8B-Instruct (Bai et al., 2025) to identify salient moving objects in each video. For each detected objecto i , MotionInsight predicts three scoress OC i ,s MC i , ands P P i . We compute the object-level motion reward by multiplying the normalized scores across the three dimensions: r i = Y dâOC,MC,P P s d i 5 .(7) We then obtain the video-level object-motion re- ward by taking the minimum reward over all de- tected salient moving objects: R(v) = N min i=1 r i .(8) For each prompt, we select the candidate with the highestR(v)as the positive sample and the can- didate with the lowestR(v)as the negative sam- ple, forming a preference pair for DPO training. We then fine-tune the base generator using these MotionInsight-derived preference pairs. For com- parison, we construct another set of preference pairs using VideoPhy2 (Bansal et al., 2025) scores MethodOCâMCâPâ Motion-Aware Rep.0.761 0.633 0.727 w/ Temporal Shuffle 0.161 0.109 0.113 w/o Camera Motion0.697 0.580 0.683 Table 7: Ablation on the design of motion-aware rep- resentations for GRPO, reported with SRCC. Rep. de- notes representations. w/ Temporal Shuffle randomly permutes the temporal order of the motion represen- tations, while w/o Camera Motion removes camera- motion tokens and keeps only object-motion features. OC: Object Consistency, MC: Motion Continuity, and P: Physical Plausibility. under the same candidate pool and train a corre- sponding DPO baseline. We recruit 20 volunteers to conduct a 2AFC hu- man study. As shown in Table 9, DPO training with MotionInsight-derived preference pairs is strongly preferred over both the original Wan 2.1 model and the VideoPhy2-DPO baseline. Figure 6 further shows that MotionInsight-DPO reduces typical ob- ject motion artifacts, including deformation during bow drawing and foot penetration when the horse crosses a hurdle. These results demonstrate that the diagnostic signals from MotionInsight are use- ful not only for evaluation, but also for improving video generation, leading to better object motion fidelity while preserving overall video quality. G Ablation on Motion-Aware Representations Table 7 further validates the effectiveness of the pro- posed motion-aware representation. When the tem- poral order of the motion embeddings is randomly shuffled, performance drops substantially across all three dimensions, indicating that the model relies on temporally structured motion form rather than merely benefiting from additional features. More- over, retaining only object-motion tokens also un- derperforms the full model, showing that both ob- ject motion and camera poses are important for dis- entangling object dynamics from camera-induced motion. Figure 5: Distribution of target objects in VidMotion. The objects are grouped into six coarse categories: Hu- man, Animal, Vehicle, Toy and Ball, Tool, and Other. The distribution highlights the diversity of object-centric motion scenarios covered by the dataset. DimensionKrippendorffâs α Object Consistency0.7981 Motion Continuity0.6846 Physical Plausibility0.7432 Table 8: Inter-annotator reliability for the three scalar motion-quality dimensions in VidMotion. Krippen- dorffâsαis used to measure agreement among annota- tors, with higher values indicating stronger consistency. MethodMotion Fidelity Overall Quality over Wan 2.192.86%86.67% over VideoPhy2-DPO87.56%87.50% Table 9: 2AFC human study results for DPO-based video generation. We compare the model trained with MotionInsight-derived preference pairs against the orig- inal Wan2.1 and the VideoPhy2-DPO (Bansal et al., 2025) baseline. MotionInsight-DPO (Ours) VideoPhy2-DPOWan2.1 A black-and-white dog weaves through colorful poles on green turf, tail raised andbody stretched mid-run . Prompt A rider in a black helmet and white shirt jumps a brown horse over an obstacle. A small windmill spins in the wind, its blades turning steadily in an outdoor setting. A woman draws a bow, releases the arrow, and lowers the bow after the shot. 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. Question: Give the target object " skateboard " in this video three continuous scores from 1 to 5 for the following aspects in order: object consistency, motion continuity, and physicalplausibility. Object Consistency MotionContinuity PhysicalPlausibility Object consistency is generally well maintained, with the skateboard remaining visually stable and showing no obvious deformation or shape inconsistency. Motion continuity is high, as the skateboard moves smoothly and continuously across the sequence without abrupt jumps. Physical plausibility is weaker, because the skateboard appears to start moving without any clear applied force , which makes the motion less physically realistic. Motion Continuity Object Consistency Physical Plausibility GPT-5.4 Human MotionInsight Question: Give the target object " brown ball " in this video three continuous scores from 1 to 5 for the following aspects in order: object consistency, motion continuity, and physicalplausibility. Object Consistency MotionContinuity PhysicalPlausibility The brown ball remains largely well-formed throughout the video, with no obvious deformation, but its position behaves unnaturally. After being hit, the ball suddenly flashes back to its starting point , breaking the continuity of its trajectory. This abrupt return also weakens the physical plausibility of the scene, since it does not match how a real ball would move in field hockey. Motion Continuity Object Consistency Physical Plausibility GPT-5.4 Human MotionInsight Figure 9: Additional qualitative results of MotionInsight. For each example, we show the sampled video frames, the modelâs reasoning for the designated object, and the predicted scores across the three motion-quality dimensions. Suppose you are an expert in judging and evaluating object-centric motion quality in real or AI-generated videos. Given the sampled frames of a video and the target object to be evaluated, please carefully analyze the motion of the target object. You should evaluate the target objectâs motion from three dimensions: (1) object consistency: whether the target object preserves a stable and coherent appearance, identity, and structure during motion; (2) motion continuity: whether the target objectâs motion is temporally smooth, continuous, and natural; (3) physical plausibility: whether the target objectâs motion follows intuitive physical causality and common-sense dynamics. Please first provide your reasoning process within<thinking> </thinking>tags, and then output only the three scores within<answer> </answer>tags.Each score should be in [1, 5], with one decimal place. For each dimension, output a score from [1, 5], where 1 means very poor, 3 means minor but noticeable artifacts, and 5 means excellent. Use the following exact JSON format: <thinking> reasoning process here </thinking> <answer> "object_consistency": 3.2, "motion_continuity": 2.8, "physical_plausibility": 3.5 </answer> Be strict and conservative in scoring. A mostly static object is not necessarily poor if it is consistent with the scene. For this video, the target object is âtarget_objectâ. The sampled video frames are as follows: Table 10: Prompting template used for multi-dimensional scoring reward in MotionInsight GRPO training and VLM baselines. Real Videos Video Models Target: Swing Target: Golf ball Real Videos Generated Videos Caption & Target Extraction Video Models VidMotion Human Annotations Caption: Gently putt the golf ball with a golf club. Target: Golf Ball Generated Videos Real Videos Score Object Consistency: Motion Continuity: Physical Plausibility: 5 3 2 Failure Causes Distortion Teleporting Splitting Disappearance ... Confidence ... 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.