Paper deep dive
VGA-BenchV2: An Expanded Unified Benchmark and Multi-Model Framework for Evaluating Video Aesthetics and Generation Quality
Longteng Jiang, DanDan Zheng, Qianqian Qiao, Heng Huang, Huaye Wang, Yihang Bo, Bao Peng, Jingdong Chen, Jun Zhou, Xin Jin
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/27/2026, 5:20:03 AM
Summary
The paper introduces VGA-BenchV2, an expanded human-aligned benchmark and optimization framework for evaluating and improving video generation quality and aesthetic value. It extends the original VGA-Bench by significantly increasing human-labeled annotations (36,000 new annotations) and introducing a hybrid evaluator architecture comprising VAQA-Net, VTag-Net, and VGQA-Net. The framework also includes an evaluation-to-optimization pipeline using reinforcement learning to fine-tune video generators based on aesthetic rewards.
Entities (9)
Relation Signals (7)
VGA-BenchV2 → extends → VGA-Bench
confidence 95% · Built upon VGA-Bench, VGA-BenchV2 preserves the original fine-grained taxonomy...
VGA-BenchV2 → uses → VAQA-Net
confidence 95% · we develop a hybrid evaluator architecture consisting of VAQA-Net for continuous aesthetic scoring
VGA-BenchV2 → uses → VTag-Net
confidence 95% · two Qwen-based Large Vision-Language Model evaluators, VTag-Net and VGQA-Net, for aesthetic tagging
VGA-BenchV2 → uses → VGQA-Net
confidence 95% · two Qwen-based Large Vision-Language Model evaluators, VTag-Net and VGQA-Net, for aesthetic tagging and generation quality assessment.
VTag-Net → basedon → Qwen
confidence 90% · two Qwen-based Large Vision-Language Model evaluators, VTag-Net and VGQA-Net
VGQA-Net → basedon → Qwen
confidence 90% · two Qwen-based Large Vision-Language Model evaluators, VTag-Net and VGQA-Net
VGA-BenchV2 → enables → Reinforcement Learning
confidence 90% · VGA-BenchV2 further introduces an evaluation-to-optimization pipeline, where the learned aesthetic evaluator serves as a reward model for reinforcement learning-based generator fine-tuning.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We introduce VGA-BenchV2, an extended human-aligned benchmark and optimization framework for jointly evaluating and improving video generation quality and aesthetic value. Built upon VGA-Bench, VGA-BenchV2 preserves the original fine-grained taxonomy with two primary dimensions-Aesthetic and Generation-and 52 sub-dimensions. Guided by this taxonomy, we curate 1,016 diverse prompts and collect over 60,000 videos generated by 12 mainstream video generation models. More importantly, VGA-BenchV2 substantially expands human-labeled supervision by adding 36,000 task-level annotations, including 16,200 for aesthetic quality, 13,200 for aesthetic tagging, and 6,600 for generation quality, corresponding to 13.46x, 11.15x, and 1.55x scale-ups over VGA-Bench, respectively. Leveraging this enlarged annotation corpus, we develop a hybrid evaluator architecture consisting of VAQA-Net for continuous aesthetic scoring and two Qwen-based Large Vision-Language Model evaluators, VTag-Net and VGQA-Net, for aesthetic tagging and generation quality assessment. Extensive experiments demonstrate strong alignment with human judgments across diverse generation models. Beyond evaluation, VGA-BenchV2 further introduces an evaluation-to-optimization pipeline, where the learned aesthetic evaluator serves as a reward model for reinforcement learning-based generator fine-tuning. This closes the loop from benchmark construction and human supervision to automated evaluation and model optimization, enabling video generators to improve not only in realism but also in aesthetic quality and human preference alignment. Resources are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2608.25452v1
- Canonical: https://arxiv.org/abs/2608.25452v1
Trouble viewing inline? Open PDF directly →
Full Text
46,002 characters extracted from source content.
Expand or collapse full text
VGA-BenchV2: An Expanded Unified Benchmark and Multi-Model Framework for Evaluating Video Aesthetics and Generation Quality Longteng Jiang 1 , Dandan Zheng 1 , Qianqian Qiao 1 , Heng Huang 1 , Huaye Wang 1 Yihang Bo 2 , Bao Peng 2 , Jingdong Chen 1 , Jun Zhou 1∗ , Xin Jin 3∗ 1 Ant Group, Beijing, China 2 Beijing Film Academy, Beijing, China 3 State Key Laboratory of General Artificial Intelligence, Beijing Institute for General Artificial Intelligence (BIGAI), Beijing, China jianglongteng.jlt, yuandan.zdd, qiaoqianqian.q, huangheng.h, wanghuaye.why@antgroup.com, boyihang@bfa.edu.cn, 3180100063@zju.edu.cn, jingdongchen.cjd@antgroup.com, jun.zhoujun@antfin.com, jinxin@bigai.ai Abstract The rapid advancement of AIGC video generation calls for evaluation frameworks that move beyond technical fidelity and incorporate human-centered aesthetic assessment. Existing benchmarks often overlook fine-grained perceptual qualities such as visual aesthetics, artistic style, and human pref- erence.To address this limitation, we intro- duce VGA-BenchV2, an extended human-aligned benchmark and optimization framework for jointly evaluating and improving video generation quality and aesthetic value. Built upon VGA-Bench, VGA-BenchV2 preserves the original fine-grained taxonomy with two pri- mary dimensions—Aesthetic and Generation—and 52 sub-dimensions.Guided by this taxonomy, we curate 1,016 diverse prompts and collect over 60,000 videos generated by 12 mainstream video generation models.More importantly, VGA- BenchV2 substantially expands human-labeled su- pervision by adding 36,000 task-level annotations, including 16,200 for aesthetic quality, 13,200 for aesthetic tagging, and 6,600 for generation quality, corresponding to 13.46×, 11.15×, and 1.55× scale- ups over VGA-Bench, respectively. Leveraging this enlarged annotation corpus, we develop a hybrid evaluator architecture consist- ing of VAQA-Net for continuous aesthetic scor- ing and two Qwen-based Large Vision-Language Model evaluators, VTag-Net and VGQA-Net, for aesthetic tagging and generation quality as- sessment.Extensive experiments demonstrate strong alignment with human judgments across diverse generation models.Beyond evaluation, VGA-BenchV2 further introduces an evaluation- to-optimization pipeline, where the learned aes- thetic evaluator serves as a reward model for re- inforcement learning-based generator fine-tuning. ∗ Corresponding author. This closes the loop from benchmark construc- tion and human supervision to automated evalua- tion and model optimization, enabling video gen- erators to improve not only in realism but also in aesthetic quality and human preference alignment. Resources are available at https://huggingface.co/ datasets/BestiVictoryLab/VGA-Bench. 1 Introduction Text-to-video generation has rapidly evolved from early proof-of-concept systems into increasingly capable gener- ative models that can synthesize coherent, temporally sta- ble, and visually compelling videos from natural language prompts [ Blattmann et al., 2023a; Blattmann et al., 2023b; Luo et al., 2023; Khachatryan et al., 2023 ] . Recent ad- vances in diffusion models [ Blattmann et al., 2023a; Song et al., 2020 ] , transformer-based architectures [ Liu et al., 2022; Selva et al., 2023 ] , and large-scale vision-language pretrain- ing [ Chen et al., 2023; Wang et al., 2023b; Dou et al., 2022 ] have further accelerated this progress, enabling state-of-the- art systems [ Wan et al., 2025; Kong et al., 2024; Tang et al., 2025; Liu et al., 2024; HaCohen et al., 2024; Team, 2024; Ma et al., 2024; Yang et al., 2024; Wang et al., 2023a; Zhang et al., 2025; Wang et al., 2025a; Guo et al., 2023; Blattmann et al., 2023a ] to support increasingly diverse cre- ative scenarios. As these models move closer to real-world use in digital art, film production, advertising, and virtual re- ality, evaluation is no longer limited to verifying whether a video is technically plausible. It must also measure whether the generated content is aesthetically appealing, stylistically controllable, and aligned with human visual preferences. However, most existing evaluation protocols are still designed around technical correctness.Metrics such as FVD [ Unterthiner et al., 2019 ] and CLIP Score [ Hessel et al., 2021 ] , together with their advanced variants [ Liu et al., 2023 ] , primarily quantify distributional similarity, temporal consistency, prompt alignment, or low-level visual artifacts. These measurements are useful for diagnosing basic genera- tion quality, but they provide limited insight into perceptual arXiv:2608.25452v1 [cs.CV] 26 Aug 2026 result Holistic evaluation of 12 models 2 Hybrid Evaluator Architecture 3 Evaluation-to-Optimization via RL Human-aligned· Specialized· Scalable Turn evaluator into reward to improve generators Prompt Suite 1,016 prompts Generated Videos 60,000+ videos from 12 models Human Annotation +36,000 new annotations Evaluators VAQA-Net / VTag-Net VGQA-Net RL Optimization Better Models Better Aesthetics AccurateInterpretable Generalizable VAQA-Net Aesthetic Scoring Continuous 0-10 score 2-stage Transfer Learning VTag-Net Aesthetic Tagging LVLM(Qwen3-VL-32B) Instruction Tuning VGQA-Net Generation Quality LVLM(Qwen3-VL-32B) Instruction Tuning Massive Human Annotations 1 Task-level annotations for training evaluators Total + 36,000 new annotations Aesthetic Tagging (Label) +13,200 new annotations Aesthetic Quality (Score) +16,200 new annotations Generation Quality (QA) +6,600 new annotations v1:1,300 17,500 Total v1:12,000 18,600 Total v1:1,300 14,500 Total 13.46x 1.55x 11.15x Figure 1: Overview of VGA-BenchV2. Left: expanded human annotations supporting evaluator training; middle: hybrid evaluator architec- ture combining VAQA-Net, VTag-Net, and VGQA-Net; right: holistic evaluation of 12 state-of-the-art video generation models, visualizing performance across 10 representative dimensions selected from the 52 sub-dimensions of VGA-Bench. The bottom panel illustrates the evaluation-to-optimization pipeline using RL. and artistic factors such as composition, lighting, color har- mony, cinematic expression, and style controllability. More importantly, conventional metrics usually serve as passive evaluation tools: they can rank or compare models, but are less effective at providing human-aligned supervision that can be directly used to improve video generators. A series of video generation benchmarks have been pro- posed to address this evaluation gap. V-Bench [ Huang et al., 2024 ] is a representative effort toward standardized multi- dimensional evaluation, but its treatment of video aesthetics remains coarse and depends heavily on off-the-shelf scoring models such as MUSIQ [ Ke et al., 2021 ] and DINO [ Caron et al., 2021 ] . VGA-Bench [ Jiang et al., 2026 ] further ad- vances this direction by introducing a fine-grained taxonomy for jointly evaluating aesthetic quality, aesthetic tags, and generation quality. Nevertheless, as video generation mod- els become stronger and more widely used, a benchmark alone is insufficient: reliable aesthetic evaluation requires substantially larger human-labeled supervision, more expres- sive evaluator architectures, and a mechanism to convert hu- man preference judgments into actionable optimization sig- nals. These requirements motivate VGA-BenchV2, which extends VGA-Bench from a fine-grained evaluation bench- mark into a human-aligned framework that unifies large-scale annotation, hybrid automated evaluation, and reinforcement learning-based model optimization. Notably, V-Bench [ Huang et al., 2024 ] represents a pi- oneering systematic effort to evaluate AIGC videos across multiple dimensions, marking a significant step toward stan- dardized evaluation. However, it reduces the multifaceted na- ture of video aesthetics into a limited set of scalar measure- ments and relies heavily on off-the-shelf scoring models, such as MUSIQ [ Ke et al., 2021 ] and DINO [ Caron et al., 2021 ] . This reliance inevitably leads to coarse granularity, poten- tial domain bias, and limited interpretability for improving generation models. VGA-Bench [ Jiang et al., 2026 ] further advances this direction by introducing a fine-grained bench- mark for jointly evaluating video aesthetic quality, aesthetic tags, and generation quality. Nevertheless, its evaluator train- ing is still limited by the scale of human-labeled supervision, and the benchmark mainly focuses on evaluation rather than forming a closed loop from human preference modeling to generative model optimization. To address these limitations, this paper introduces VGA- BenchV2 (as shown in Figure 1), an extended human- aligned benchmark and optimization framework for video aesthetic and generation quality assessment.Built upon VGA-Bench [ Jiang et al., 2026 ] , VGA-BenchV2 preserves the original fine-grained taxonomy with two primary dimen- sions, Aesthetic and Generation, and 52 sub-dimensions, while substantially extending the human annotation scale, evaluator architecture, and optimization capability. Our main contributions are summarized as follows: • Large-scale human annotation expansion.Com- pared with VGA-Bench [ Jiang et al., 2026 ] , VGA- BenchV2 substantially expands human-labeled supervi- sion by adding 36,000 newly collected task-level anno- tations, including 16,200 annotations for aesthetic qual- ity assessment, 13,200 annotations for aesthetic tagging, and 6,600 annotations for generation quality assess- ment. This corresponds to 13.46×, 11.15×, and 1.55× scale-ups over VGA-Bench for the three tasks, respec- tively, transforming the benchmark from an evaluation- oriented suite into a large-scale human-aligned training infrastructure. • Human-supervised hybrid evaluator architecture. Leveraging the enlarged annotation corpus, we de- velop a hybrid evaluator system consisting of VAQA- Net for continuous aesthetic scoring and two Qwen- based Large Vision-Language Model evaluators, VTag- Net and VGQA-Net, for aesthetic tagging and genera- tion quality assessment, respectively. This design com- bines specialized aesthetic regression with LVLM-based semantic reasoning, enabling scalable, interpretable, and human-aligned evaluation across diverse generation models. • Evaluation-to-optimization via reinforcement learn- ing. Beyond passive evaluation, VGA-BenchV2 intro- duces an evaluation-to-optimization pipeline in which the learned aesthetic evaluator serves as a reward model for reinforcement learning-based fine-tuning of video generators. By using the aesthetic score as a reward signal, the framework converts human aesthetic prefer- ences into actionable optimization objectives, demon- strating that VGA-BenchV2 can directly guide genera- tive models toward higher aesthetic quality and better human preference alignment. • Unified benchmark, evaluator, and optimization in- frastructure. VGA-BenchV2 integrates 52 fine-grained evaluation dimensions, 1,016 diverse prompts, over 60,000 videos generated by 12 mainstream models, large-scale human annotations, hybrid automated eval- uators, and reinforcement learning-based optimization into a unified framework. This enables systematic anal- ysis, fair cross-model comparison, and closed-loop im- provement of video generation models. We believe that VGA-BenchV2 serves not only as a rigor- ous evaluation platform, but also as a key infrastructure for advancing the next generation of video generation systems with enhanced aesthetic intelligence, artistic controllability, and human preference alignment. 2 Related Work Evaluating video generation models remains challenging de- spite their rapid progress. Early metrics such as FVD [ Un- terthiner et al., 2019 ] , CLIP Score [ Hessel et al., 2021 ] , and their advanced variants [ Liu et al., 2023 ] mainly measure distributional similarity, temporal consistency, prompt align- ment, or low-level visual artifacts. While useful for assessing basic technical quality, these metrics are insufficient for cap- turing fine-grained perceptual factors such as composition, lighting, color harmony, artistic style, and overall aesthetic appeal. VGA-Bench (V1) VGA-BenchV2 Human Annotations Task-level annotations (Limited scale) Evaluator Architecture Task-specific evaluators Evaluation of Generators Holistic evaluation & ranking of generation models Human Annotations Massively expanded scale Hybrid Evaluator Architecture Human-aligned · Specialized · Scalable RL Optimization Reward-driven generator improvement NEW NEW Evolution Figure 2:Comparison between VGA-Bench (V1) and VGA- BenchV2 (V2). Both include human annotations and multi-model evaluation; V2 expands annotation scale, adds a hybrid evaluator, and supports optional RL optimization. Recent benchmarks have promoted more systematic eval- uation of video generation models.V-Bench [ Huang et al., 2024 ] provides a representative multi-dimensional eval- uation protocol, and V-Bench2 [ Zheng et al., 2025 ] fur- ther extends the evaluation scope. Other benchmarks, in- cluding ChronoMagic-Bench [ Yuan et al., 2024 ] , T2V- CompBench [ Sun et al., 2025 ] , and StoryEval [ Wang et al., 2025b ] , focus on complementary aspects such as temporal coherence, compositional binding, and narrative consistency. Compared with these works, VGA-Bench [ Jiang et al., 2026 ] is distinguished by its fine-grained treatment of video aes- thetics, jointly modeling aesthetic quality, aesthetic tags, and generation quality within a unified taxonomy. VGA-BenchV2 further extends VGA-Bench [ Jiang et al., 2026 ] from a fine-grained evaluation benchmark into a human-aligned evaluation and optimization framework. Specifically, it strengthens human-labeled supervision, in- troduces a hybrid evaluator architecture that combines spe- cialized aesthetic scoring with LVLM-based tagging and generation-quality assessment, and further connects evalua- tion with reinforcement learning-based generator optimiza- tion. In this way, VGA-BenchV2 moves beyond passive benchmarking and forms a closed-loop framework that links benchmark design, human annotation, automated evaluation, and model improvement. 3 VGA-BenchV2 Construction 3.1 Inherited Evaluation Taxonomy VGA-BenchV2 inherits the fine-grained evaluation taxonomy of VGA-Bench [ Jiang et al., 2026 ] to ensure continuity, re- producibility, and cross-version comparability. Rather than redesigning the benchmark dimensions, we preserve the orig- inal taxonomy as the evaluation backbone and focus the V2 extension on human annotation expansion, hybrid evaluator training, and evaluation-to-optimization. Overall, the taxon- omy contains 52 sub-dimensions organized into two primary categories: Aesthetic and Generation. The Aesthetic category consists of two complementary perspectives. • Aesthetic Quality measures the perceptual appeal of generated videos as continuous scores, covering ten di- mensions adapted from VADB [ Qiao et al., 2025 ] : Over- all Score (Ovr), Composition (Com), Shot Size (S), Benchmark Aesthetic Dims Evaluator Architecture Annotation Usage RL Opt. V-Bench [ Huang et al., 2024 ] 1Off-the-shelf Expert ModelsAlignment Only✗ V-Bench2.0 [ Zheng et al., 2025 ] 2Off-the-shelf Expert ModelsAlignment Only✗ T2V-CompBench [ Sun et al., 2025 ] 0MLLM-basedAlignment Only✗ ChronoMagic-Bench [ Yuan et al., 2024 ] 0HybridAlignment Only✗ StoryEval [ Wang et al., 2025b ] 0MLLM-basedAlignment Only✗ VGA-Bench [ Jiang et al., 2026 ] 21Specialized NetsTraining & Alignment✗ VGA-BenchV2 (Ours)21 Hybrid Specialized Nets + LVLMs Training, Alignment & Reward ✓ Table 1: Comparison with existing video generation benchmarks. VGA-BenchV2 preserves the fine-grained aesthetic and generation taxon- omy of VGA-Bench while further extending it with hybrid human-aligned evaluators and an evaluation-to-optimization pipeline. Lighting (Lig), Visual Tone (VT), Color (Col), Depth of Field (DoF), Expression (Exp), Costume (Cos), and Makeup (Mak). • Aesthetic Tagging describes discrete visual and stylis- tic attributes as classification labels. Following pho- tographic and cinematographic principles [ Qiao et al., 2025; Matbouly, 2022; Deren, 1960; Brown, 2016 ] , it contains eleven tag dimensions: Composition Types (CT), Shot Type (ST), Number of Light Sources (NoLS), Light Source Position (LSP), Light Quality (LQ), Light Color (LC), Color Temperature (ColT), Saturation (Sat), Brightness (Bri), Contrast (Con), and Depth of Field (DoF). These 21 aesthetic-related dimensions support both quantitative model ranking and interpretable analy- sis of visual style controllability. The Generation category evaluates whether generated videos are semantically faithful, physically plausible, and vi- sually stable. Following VGA-Bench [ Jiang et al., 2026 ] and building upon V-Bench [ Huang et al., 2024 ] , VGA-BenchV2 retains 31 generation-quality sub-dimensions grouped into three aspects. • Video-Text Consistency includes Character-Text Con- sistency (Cha-T), Action-Text Consistency (Act-T), Object-Text Consistency (Obj-T), Scene-Text Consis- tency (Sce-T), Object Position-Text Consistency (Pos- T), Camera Movement-Text Consistency (Cam-T), Ob- ject Attribute-Text Consistency (Att-T), Video Content- Text Consistency (Cnt-T), Video Style-Text Consistency (Sty-T), and Video Speed-Text Consistency (Spd-T). • Reality & Plausibility includes Rigid Body Colli- sion Realism (Rig-R), Fluid Motion Realism (Flu-R), Gaseous Motion Realism (Gas-R), Gradual Change Re- alism (Gra-R), Object Trajectory Realism (Tra-R), Ac- tion Realism (Act-R), Scene Realism (Sce-R), Object Realism (Obj-R), Character Generation Quality (Cha- R), Weather Representation Realism (Wea-R), Time Period Representation Realism (Tim-R), Lighting & Shadow Realism (Lig-R), Moving Scene Reasonable- ness (Dyn-R), Textual Attribute Representation (Att-R), Overall Realism (Glo-R), and Unrealistic Description Imaginative Presentation (Ima-R). Aesthetic Quality Model：AnimateDiff-v2 Prompt：A conceptual makeup video with a strong artistic feel, using high-contrast black and white throughout to highlight the structure and ...... Dimentions：Visualtone + Costume Aesthetic Tagging Model：Latte-1 Prompt：The night market stalls are aglow with flashing lights, creating a complex lighting system from multiple sources...... Dimentions：Composition + Light Source Number + Light Source Color + ...... Generation Quality Model：HunYuan Prompt：On the Arctic glacier, reverse playback shot of an ice block falling into the water with a static camera. Dimentions：Scene-Text Consistency + Scene Realism + Video Speed-Text Consistency + Camera Movement-Text Consistency Figure 3: Prompts for the three core dimensions, their corresponding sub-dimensions, and example generated videos. • Basic Quality includes Video Noise-Free (Noi), Ab- normal Lighting Detection (AbL), Video Clarity (Cla), Static Content Stability (Sta), and Static Content Non- distortion (Dis). We use the suffix “-T” to denote video- text consistency dimensions and “-R” to denote reality or plausibility dimensions, while basic quality dimen- sions use standalone mnemonic codes. Detailed defi- nitions and evaluation criteria are provided in the Ap- pendix. 3.2 Prompt Suite and Video Pool VGA-BenchV2 inherits the prompt suite from VGA- Bench [ Jiang et al., 2026 ] to ensure cross-version compara- bility and stable dimension-wise evaluation. The prompt suite is designed under the principle of explicit attribute specifica- tion: each prompt explicitly triggers one or more target di- mensions, such that annotators and automated evaluators as- Aesthetic Quality Overall Score：5 Composition：2.5 Lighting：2 Aesthetic Tagging Number of Light Sources： Single Light Source Light Quality： Soft Light Light Color： White (Neutral) Light Generation Quality Scene-Text Consistency： 2: The entire video is not entirely the described scene or is mixed with other scenes. Scene Realism： 2: Not sufficiently realistic, some anomalies can be noticed. Composition TypesE： Rule of Thirds Light Source Position： Top Light Light Quality： Soft Light Overall Score：2 Composition：1 Color：0.5 Expression：1.5 Video Clarity： 3: The video is very clear, with only slight imperfections visible in detail- rich areas (such as hair or edges). Static Content Non-distortion： 2: Slight distortion occurs. Figure 4: Examples of human annotations for the three core dimen- sions. sess only the attributes that are semantically grounded in the prompt. This design reduces ambiguity during both human annotation and automated evaluation. The prompt suite contains 1,016 prompts, including 200 prompts for aesthetic quality, 220 for aesthetic tagging, and 596 for generation quality. Each dimension is covered by at least 50 prompts, and the suite supports both single- dimension and multi-dimensional combinations. We also re- tain two lightweight subsets with 508 and 127 prompts for efficient evaluation under different computational budgets. In VGA-BenchV2, the prompt suite serves as a unified an- chor for three downstream components: large-scale human annotation, supervised evaluator training, and evaluation-to- optimization experiments. By reusing the same dimension- aligned prompts as VGA-Bench, VGA-BenchV2 isolates the effect of expanded human supervision and improved evalua- tor design, while preserving a consistent benchmark protocol for comparing video generation models. 3.3 Large-scale Human Annotation Expansion A central goal of VGA-BenchV2 is to transform VGA-Bench from an evaluation-oriented benchmark into a human-aligned training infrastructure for automated video evaluators. To this end, we substantially expand the task-level human anno- tations over VGA-Bench [ Jiang et al., 2026 ] . Specifically, VGA-BenchV2 introduces 36,000 newly collected annota- tions, including 16,200 annotations for aesthetic quality as- sessment, 13,200 annotations for aesthetic tagging, and 6,600 annotations for generation quality assessment. These annota- tions provide direct supervision for evaluator training, align- ment verification, and reward modeling. Annotation Protocol and Quality Control We follow an expert-guided and multi-annotator annotation protocol. Domain experts first provide exemplar annotations and detailed rating guidelines for each task. Trained annota- tors then label the remaining samples according to these ex- emplars, while experts conduct batch-wise audits to ensure Task VGA-Bench Supervision New V2 Annotations V2 Total Scale-up Aesthetic Quality1,300+16,20017,50013.46× Aesthetic Tagging1,300+13,20014,50011.15× Generation Quality12,000+6,60018,6001.55× Total14,600+36,00050,6003.47× Table 2: Human annotation expansion in VGA-BenchV2. We re- port task-level supervision for aesthetic quality, aesthetic tagging, and generation quality. VGA-BenchV2 substantially increases the amount of human-labeled data used for evaluator training and align- ment. annotation consistency. If a batch fails the quality check, it is rejected and re-annotated. For aesthetic quality, each sample is scored on a 0–10 scale, and the final score is obtained by averaging multiple inde- pendent ratings. For aesthetic tagging, annotators assign dis- crete visual labels such as composition type, light source po- sition, shot type, color temperature, and contrast, with final labels determined by majority voting. For generation qual- ity, each dimension is formulated as a task-specific question with structured ordinal options, enabling the labels to capture different levels of semantic consistency, realism, and basic visual quality. Importantly, all annotation tasks follow the explicit-trigger principle inherited from the prompt suite: annotators evaluate only the dimensions explicitly specified by the prompt. This prevents subjective speculation on unmentioned attributes and ensures that each label is semantically grounded in the prompt-video pair. 3.4 Hybrid Automated Evaluation Networks Leveraging the expanded human-annotated corpus, VGA- BenchV2 moves beyond lightweight task-specific evaluators to train a robust, human-aligned hybrid evaluator system. The architecture combines specialized aesthetic regression with LVLM-based semantic reasoning, consisting of VAQA-Net for continuous aesthetic scoring, VTag-Net for aesthetic tag- ging, and VGQA-Net for generation quality assessment. VAQA-Net (Aesthetic Scoring): We adopt a two-stage transfer learning strategy. First, the video encoder is initial- ized from the VADB pre-trained model to inherit aesthetic representations. Second, the model is fine-tuned on a large- scale generated video dataset, including 8,366 videos from 12 mainstream generation models with high-quality human an- notations. Evaluation is performed on a held-out set of 1,186 generated videos to assess generalization. VTag-Net & VGQA-Net (LVLM-based Evaluators): To enable semantic reasoning and detailed assessment, we lever- age Qwen3-VL-32B [ Bai et al., 2025 ] as the backbone and apply instruction tuning for task-specific adaptation. • VTag-Net: Trained on a combination of real VADB videos and 11,100 generated videos with human- annotated aesthetic tags, outputting discrete visual la- bels. • VGQA-Net: Trained on 7,054 generated videos with human-annotated question-answer pairs, optimized for coherent reasoning and generation quality assessment. (a) VAQA-Net Architecture (Regression - Fine-tuned) Pre-trained Video Encoder MLP Scalar Aesthetic Score (0-1) Task Prompt (Text) Large Vision- Language Model (Qwen3-VL-32B) Aesthetic Tags (VTag-Net Output) Quality Assessment QA (VGQA-Net Output) (b) VTag-Net & VGQA-Net Architecture (Generation - Instruction Tuning) Figure 5: Overview of the hybrid evaluation networks. VAQA-Net predicts scalar aesthetic scores via a pre-trained encoder and regres- sion head, while VTag-Net and VGQA-Net leverage a LVLM back- bone via instruction tuning for aesthetic tagging and quality reason- ing. The training set includes representative samples from all 12 source models to ensure generalization across generation paradigms. The network architectures of VAQA-Net, VTag-Net, and VGQA-Net are illustrated in Figure 5. 3.5 Evaluation-to-Optimization Interface Beyond serving as passive evaluation metrics, the learned evaluators in VGA-BenchV2 can be used as reward mod- els for generator optimization. In particular, the aesthetic score predicted by VAQA-Net provides a scalar reward re- flecting human-aligned visual preference. This reward can be incorporated into reinforcement learning-based fine-tuning pipelines, enabling video generators to optimize toward higher aesthetic quality while maintaining stability. Formally, given a prompt p, the current policy π θ generates a video v ∼ π θ (·|p), and VAQA-Net outputs a scalar reward: r φ (v) = VAQA φ (v),(1) with φ denoting the frozen evaluator parameters. To stabilize training, KL regularization is applied: ̃r(v,p) = r φ (v)− λ KL D KL (π θ (·|p)∥π ref (·|p)),(2) and the normalized advantage is A i = ̃r i − μ B σ B + ε ,(3) where μ B and σ B denote the batch mean and standard devia- tion. For policy optimization, the unclipped and clipped advan- tages are defined as r i,t = ρ i,t A i , c i,t = clip(ρ i,t , 1− ε, 1 + ε)A i , ρ i,t = π θ (a i,t |s i,t ) π θ old (a i,t |s i,t ) , (4) and the clipped GRPO objective is L GRPO (θ) =−E i,t min(r i,t ,c i,t ) .(5) VAQA-Net (SROCC)VTag-Net (Accuracy) DimensionScoreDimension#Cls. Acc. Overall Score87.6Color Temp. (ColT)364.8 Composition (Com)86.9Saturation (Sat)371.9 Shot Size (S)87.4 Brightness (Bri)370.3 Lighting (Lig)87.2Contrast (Con)375.0 Visual Tone (VT)87.4 Light Quality (LQ)266.3 Color (Col)86.5Light Color (LC)493.6 Depth of Field (DoF) 87.4 Light Pos. (LSP)568.3* Expression (Exp)89.2Num. Lights (NoLS)283.7 Costume (Cos)88.4 Comp. Type (CT)749.8* Makeup (Mak)86.3Shot Type (ST)575.9* Depth of Field (DoF)276.1 Table 3: Validation of Aesthetic Evaluators. We report the SROCC (%) for VAQA-Net and Classification Accuracy (%) for VTag-Net. Specifically, we adopt Top-1 Accuracy for dimensions with≤4 classes and Top-2 Accuracy for those with>4 classes (marked with *) to account for varying label space sizes. Dim.Acc. Dim.Acc. Dim.Acc. Dim.Acc. Cha-T70.5Sty-T69.6Sce-R89.2Glo-R69.8 Act-T70.7Spd-T76.6Obj-R81.2Ima-R59.2 Obj-T75.2Rig-R62.2Cha-R56.8Noi58.3 Sce-T74.7Flu-R75.6Wea-R66.8AbL85.7 Pos-T64.3Gas-R78.3Tim-R83.1Cla66.4 Cam-T72.5Gra-R80.4Lig-R80.5Sta50.5 Att-T72.6Tra-R62.1Dyn-R67.0Dis71.7 Cnt-T58.4Act-R80.2Att-R78.7– Table 4: Validation of VGQA-Net. We report the Accuracy (%) across 31 sub-dimensions. The suffixes -T and -R denote Video- Text Consistency and Reality/Plausibility, respectively. This formulation encourages higher-reward samples to con- tribute stronger updates while constraining policy drift. De- tailed implementation and training parameters are described in Section 4. 4 Experiments and Results 4.1 Validation of Automated Evaluator To ensure the reliability of VGA-BenchV2, we rigorously validated our automated evaluators against human ground truth on a held-out test set. For aesthetics, Table 3 shows that VAQA-Net achieves a high SROCC of 87.6% on the overall score, demonstrating strong ranking consistency with human experts. Simultaneously, VTag-Net delivers robust classifi- cation performance, particularly in objective dimensions like Light Color (93.6%) and Number of Light Sources (83.7%), validating its precision in identifying visual elements. For generation quality, VGQA-Net (Table 4) exhibits consistent alignment across all 31 fine-grained dimensions, with an av- erage accuracy approximately 71.3%. High scores in chal- lenging categories like Scene Realism (89.2%) and Abnormal Lighting Detection (85.7%) confirm that our MLLM-based evaluators serve as effective, scalable proxies for human as- sessment. These high scores not only demonstrate the effectiveness of our models but also indicate that their predictions are closely aligned with expert human annotations, corroborating the human-aligned supervision highlighted in Table 1. 4.2 VGA-BenchV2 Evaluation Results We evaluate all generative models using the trained VAQA- Net, VTag-Net, and VGQA-Net to assess their performance across diverse aesthetic and quality dimensions. To ensure a fair and unbiased ranking, all generated videos used for evaluation are strictly held out from the evaluator training process, preventing data leakage. The evaluated models are listed chronologically by release date: Stable Video Diffusion (SVD) [ Blattmann et al., 2023a ] , AnimateDiff-v2 [ Guo et al., 2023 ] , LaVie [ Wang et al., 2025a ] , Show-1 [ Zhang et al., 2025 ] , ModelScope [ Wang et al., 2023a ] , CogVideoX [ Yang et al., 2024 ] , Latte-1 [ Ma et al., 2024 ] , Mochi [ Team, 2024 ] , LTXVideo [ HaCohen et al., 2024 ] , HunyuanVideo [ Kong et al., 2024 ] , Wan2.2 [ Wan et al., 2025 ] , and Sora2 [ Liu et al., 2024 ] . The ranking metrics are defined as follows: • Aesthetic Quality: Models are ranked by their average predicted score (higher is better). • Aesthetic Tagging: We utilize alignment accuracy as the metric (i.e., whether the generated video successfully reflects the prompted aesthetic tag), where higher mean accuracy indicates superior controllability. • Generation Quality: Models are ranked by the mean score across all sub-dimensions, where a higher average indicates better generation fidelity and consistency. The final aggregated results, obtained by normalizing and averaging scores across all sub-dimensions within the three main categories, are presented in Table 5. For detailed per- formance breakdowns of each model across specific sub- dimensions, please refer to the Appendix. 4.3 Aesthetic Optimization via Reinforcement Learning To validate the practical effect of our evaluators, we fine- tune Wan2.1 using the evaluation-to-optimization interface described in Section 3.5. As shown in Figure 6, the VAQA- Net Overall Score is used as the reward, and Flow-GRPO [ Liu et al., 2025 ] is applied with LoRA and ODE-to-SDE conver- sion. Quantitative results on a held-out test set show that the av- erage aesthetic score increases from 0.49 to 0.52. Qualitative results indicate that fine-tuned videos exhibit more appealing visual composition and higher overall aesthetic quality. These results demonstrate that VGA-BenchV2 evaluators can sup- port not only passive evaluation but also reward-driven video generator optimization. 5 Conclusion We present VGA-BenchV2, an extended human-aligned benchmark and optimization framework that builds upon Task Prompt (Text) Wan2.1 (Policy Model) VAQA-Net (Reward Model) Aesthetic Score (Reward) Flow-GRPO (RL Algorithm) Policy Update RL Optimization Figure 6: RL Aesthetic Optimization Pipeline. We utilize the Flow- GRPO algorithm to fine-tune video generation models, employing VAQA-Net as the reward model to enhance artistic quality. ModelAes. Score Tag Cla. Gen. Level SVD [ Blattmann et al., 2023a ] 0.200.600.69 AnimateDiff [ Guo et al., 2023 ] 0.360.570.67 LaVie [ Wang et al., 2025a ] 0.340.580.68 Show-1 [ Zhang et al., 2025 ] 0.290.600.04 ModelScope [ Wang et al., 2023a ] 0.310.510.64 CogVideoX [ Yang et al., 2024 ] 0.410.560.70 Latte-1 [ Ma et al., 2024 ] 0.350.600.68 Mochi [ Team, 2024 ] 0.210.680.71 LTXVideo [ HaCohen et al., 2024 ] 0.220.610.61 HunyuanVideo [ Kong et al., 2024 ] 0.450.66 0.73 Wan2.2 [ Wan et al., 2025 ] 0.460.630.71 Sora2 [ Liu et al., 2024 ] 0.500.640.80 Table 5: Performance comparison of state-of-the-art text-to-video generation models on aesthetic score (Aes. Score), tag classification accuracy (Tag Cla.), and generation level (Gen. Level) metrics. VGA-Bench by preserving 52 sub-dimensions, 1,016 prompts, and over 60,000 videos while introducing three key advancements: expanded human-labeled supervision for aesthetic quality, aesthetic tagging, and generation quality; a hybrid evaluator architecture combining VAQA-Net and LVLM-based VTag-Net and VGQA-Net for accurate, in- terpretable, and scalable assessment; and an evaluation-to- optimization pipeline enabling RL-based fine-tuning of video generators to align outputs with human aesthetic prefer- ence. Experiments demonstrate that VGA-BenchV2 provides more precise, human-aligned evaluation and facilitates multi- dimensional optimization. In the future, it can be extended to higher-resolution videos, additional aesthetic dimensions, cross-modal generation, and controllable creative applica- tions, supporting next-generation AIGC video systems that harmonize fidelity, artistry, and human preference. Ethical Statement We strictly ensure that all prompts and generated videos within this benchmark have been rigorously screened to ex- clude pornographic, violent, or otherwise offensive content. Furthermore, all human annotation procedures comply with ethical guidelines and have been reviewed and approved by the relevant Institutional Review Board (IRB). Acknowledgments This work was supported by the Ant Group Research Fund, the National Natural Science Foundation of China under Grant No.62072014, and the Opening Project of the State Key Laboratory of General Artificial Intelligence, BIGAI/Peking University, Beijing, China (Project No.SKLAGI2025OP01). References [ Bai et al., 2025 ] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Ke- qin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan Liu, Dunjie Lu, Ruilin Luo, Chenxu Lv, Rui Men, Lingchen Meng, Xuancheng Ren, Xingzhang Ren, Sibo Song, Yu- chong Sun, Jun Tang, Jianhong Tu, Jianqiang Wan, Peng Wang, Pengfei Wang, Qiuyue Wang, Yuxuan Wang, Tian- bao Xie, Yiheng Xu, Haiyang Xu, Jin Xu, Zhibo Yang, Mingkun Yang, Jianxin Yang, An Yang, Bowen Yu, Fei Zhang, Hang Zhang, Xi Zhang, Bo Zheng, Humen Zhong, Jingren Zhou, Fan Zhou, Jing Zhou, Yuanzhi Zhu, and Ke Zhu.Qwen3-vl technical report. arXiv preprint arXiv:2511.21631, 2025. [ Blattmann et al., 2023a ] Andreas Blattmann, Tim Dock- horn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. [ Blattmann et al., 2023b ] Andreas Blattmann, Robin Rom- bach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis.Align your latents: High- resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 22563–22575, 2023. [ Brown, 2016 ] Blain Brown. Cinematography: theory and practice: image making for cinematographers and direc- tors. Routledge, 2016. [ Caron et al., 2021 ] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ́ e J ́ egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vi- sion transformers. In Proceedings of the IEEE/CVF in- ternational conference on computer vision, pages 9650– 9660, 2021. [ Chen et al., 2023 ] Fei-Long Chen, Du-Zhen Zhang, Ming- Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38–56, 2023. [ Deren, 1960 ] Maya Deren. Cinematography: the creative use of reality. Daedalus, 89(1):150–167, 1960. [ Dou et al., 2022 ] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to- fine vision-language pre-training with fusion in the back- bone. Advances in neural information processing systems, 35:32942–32956, 2022. [ Guo et al., 2023 ] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. [ HaCohen et al., 2024 ] Yoav HaCohen,Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. [ Hessel et al., 2021 ] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 conference on empirical methods in natural language processing, pages 7514–7528, 2021. [ Huang et al., 2024 ] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video gen- erative models. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 21807–21818, 2024. [ Jiang et al., 2026 ] Longteng Jiang, DanDan Zheng, Qian- qian Qiao, Heng Huang, Huaye Wang, Yihang Bo, Bao Peng, Jingdong Chen, Jun Zhou, and Xin Jin.Vga- bench: A unified benchmark and multi-model framework for video aesthetics and generation quality evaluation. arXiv preprint arXiv:2604.10127, 2026. [ Ke et al., 2021 ] Junjie Ke, Qifei Wang, Yilin Wang, Pey- man Milanfar, and Feng Yang. Musiq: Multi-scale im- age quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148– 5157, 2021. [ Khachatryan et al., 2023 ] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi.Text2video-zero: Text-to-image diffusion models are zero-shot video generators.In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954–15964, 2023. [ Kong et al., 2024 ] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A system- atic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. [ Liu et al., 2022 ] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202– 3211, 2022. [ Liu et al., 2023 ] Yuanxin Liu, Lei Li, Shuhuai Ren, Run- dong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou.Fetv: A benchmark for fine-grained evalua- tion of open-domain text-to-video generation. Advances in Neural Information Processing Systems, 36:62352–62387, 2023. [ Liu et al., 2024 ] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. [ Liu et al., 2025 ] Jie Liu, Gongye Liu, Jiajun Liang, Yang- guang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang.Flow-grpo: Training flow matching models via online rl. arXiv preprint arXiv:2505.05470, 2025. [ Luo et al., 2023 ] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. arXiv preprint arXiv:2303.08320, 2023. [ Ma et al., 2024 ] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. [ Matbouly, 2022 ] Mustafa Yousry Matbouly. Quantifying the unquantifiable: the color of cinematic lighting and its effect on audience’s impressions towards the appearance of film characters. Current Psychology, 41(6):3694–3715, 2022. [ Qiao et al., 2025 ] Qianqian Qiao, DanDan Zheng, Yihang Bo, Bao Peng, Heng Huang, Longteng Jiang, Huaye Wang, Jingdong Chen, Jun Zhou, and Xin Jin. Vadb: A large-scale video aesthetic database with profes- sional and multi-dimensional annotations. arXiv preprint arXiv:2510.25238, 2025. [ Selva et al., 2023 ] Javier Selva, Anders S Johansen, Sergio Escalera, Kamal Nasrollahi, Thomas B Moeslund, and Al- bert Clap ́ es. Video transformers: A survey. IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 45(11):12922–12943, 2023. [ Song et al., 2020 ] Yang Song,Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole.Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. [ Sun et al., 2025 ] Kaiyue Sun, Kaiyi Huang, Xian Liu, Yue Wu, Zihan Xu, Zhenguo Li, and Xihui Liu.T2v- compbench: A comprehensive benchmark for composi- tional text-to-video generation.In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8406–8416, 2025. [ Tang et al., 2025 ] Shixiang Tang, Yizhou Wang, Lu Chen, Yuan Wang, Sida Peng, Dan Xu, and Wanli Ouyang. Human-centric foundation models: Perception, generation and agentic modeling. arXiv preprint arXiv:2502.08556, 2025. [ Team, 2024 ] Genmo Team. Mochi 1. https://github.com/ genmoai/models, 2024. [ Unterthiner et al., 2019 ] ThomasUnterthiner,Sjoerd Van Steenkiste, Karol Kurach, Rapha ̈ el Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. [ Wan et al., 2025 ] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. [ Wang et al., 2023a ] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. [ Wang et al., 2023b ] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggar- wal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 19175–19186, 2023. [ Wang et al., 2025a ] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al.Lavie: High-quality video generation with cascaded latent diffu- sion models. International Journal of Computer Vision, 133(5):3059–3078, 2025. [ Wang et al., 2025b ] Yiping Wang, Xuehai He, Kuan Wang, Luyao Ma, Jianwei Yang, Shuohang Wang, Simon Shaolei Du, and Yelong Shen. Is your world simulator a good story presenter? a consecutive events-based benchmark for fu- ture long video generation. In Proceedings of the Com- puter Vision and Pattern Recognition Conference, pages 13629–13638, 2025. [ Yang et al., 2024 ] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. [ Yuan et al., 2024 ] Shenghai Yuan, Jinfa Huang, Yongqi Xu, Yaoyang Liu, Shaofeng Zhang, Yujun Shi, Rui-Jie Zhu, Xinhua Cheng, Jiebo Luo, and Li Yuan. Chronomagic- bench: A benchmark for metamorphic evaluation of text- to-time-lapse video generation. Advances in Neural Infor- mation Processing Systems, 37:21236–21270, 2024. [ Zhang et al., 2025 ] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. Inter- national Journal of Computer Vision, 133(4):1879–1893, 2025. [ Zheng et al., 2025 ] Dian Zheng, Ziqi Huang, Hongbo Liu, Kai Zou, Yinan He, Fan Zhang, Lulu Gu, Yuanhan Zhang, Jingwen He, Wei-Shi Zheng, et al. Vbench-2.0: Advanc- ing video generation benchmark suite for intrinsic faithful- ness. arXiv preprint arXiv:2503.21755, 2025.