TREND-10K: A Comprehensive Dataset for Next-Generation Video Quality Assessment Based on Preference-Driven Media

Ziheng Jia; Zicheng Zhang; Junqi Zhang; Jiaying Qian; Jiarui Wang; Yushuo Zheng; Xiongkuo Min

cs.AIAug 11, 2026

TREND-10K: A Comprehensive Dataset for Next-Generation Video Quality Assessment Based on Preference-Driven Media

Authors: Ziheng Jia, Zicheng Zhang, Junqi Zhang, Jiaying Qian, Jiarui Wang, Yushuo Zheng, Xiongkuo Min

Organizations: Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory

Abstract

The increasing prominence of short-video platforms, coupled with the advanced commercialization of AI-generated content (AIGC) videos, has led to a shift in the types of video media trend consumed by users in their daily lives. Traditional user-generated content (UGC) is gradually being replaced by professional short dramas and AIGC entertainment. Consequently, VQA for contemporary media content has become increasingly important. This requires a unified evaluation framework that can handle diverse video content and evolving media trends. In this context, we introduce TREND-10K, a next-generation comprehensive VQA dataset consisting of the trend-driven part and the static part, containing $10,000$ videos across a wide spectrum of content types. The trend-driven part is based on the TREND-Search framework, which captures user preference profiles from trending lists on online platforms and formulates sampling strategies based on these profiles. The static part, on the other hand, is composed of supplementary samples selected from publicly available datasets. To support unified evaluation for various video types, we incorporate three evaluation dimensions: technical, aesthetic, and AIGC-trace. Experiments show that our dataset ensures high annotation quality and exhibits remarkable generalization across multiple content categories. In conclusion, our work presents a robust framework for advancing VQA, addressing challenges caused by the temporal evolution of user perceptual habits and preferences.

Figures & tables

Figure 1: The motivation of TREND-10K is to enable VQA to continuously adapt to evolving media-consumption trends, thereby better serving as feedback for media technologies. We design the TREND-Search system to construct the trend-driven part through three major procedures, while introducing the static part for balancing. The resulting dataset contains videos with diverse content.

Figure 2: Visualization of annotated data analysis.

Models	Technical	Aesthetic	AIGC-trace	AVG.
FAST-VQA-B	0.833 / 0.856	0.801 / 0.818	0.846 / 0.858	0.835
DOVER	0.838 / 0.851	0.821 / 0.833	0.858 / 0.907	0.851
Mini-VQA-VII	0.820 / 0.844	0.796 / 0.803	0.851 / 0.897	0.835
Mini-VQA-IX	0.825 / 0.867	0.813 / 0.821	0.845 / 0.891	0.844
KVQ	0.850 / 0.879	0.805 / 0.817	0.880 / 0.913	0.857
Q-Align	0.856 / 0.877	0.812 / 0.821	0.866 / 0.905	0.856

Table 1: Intra-dataset results ( PLCC

\uparrow

/ SRCC

\uparrow

). Mini-VQA refers to Minimalistic-VQA . [Per column: best in boldface , second in italics .]

Trained on LSVQ (train) ( $28,056$ videos)
Datasets	YT-UGC		YT-Gaming		CGVDS		KVQ		FINEVD		AIGVQA		AGAV		GAIA		AVG.
# test videos	1,098		600		357		2,926		1,016		2,808		622		9,180
Models	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$
\rowcolor light-gray0 FAST-VQA-B	0.730	0.743	0.636	0.675	0.727	0.750	0.524	0.529	0.617	0.669	0.533	0.589	0.315	0.315	0.228	0.228	0.550
Mini-VQA-VII	0.748	0.756	0.621	0.693	0.725	0.769	0.612	0.654	0.550	0.632	0.625	0.634	0.277	0.281	0.207	0.199	0.561
\rowcolor light-gray0 Mini-VQA-IX	0.761	0.768	0.675	0.734	0.767	0.806	0.629	0.668	0.625	0.682	0.643	0.644	0.268	0.279	0.175	0.175	0.581

Table 2: Cross-dataset performance of VQA models pre-trained from scratch on different sources. Test datasets marked in italics denote out of distribution (OOD) . When trained on LSVQ , all models use publicly available weights, except VITAL , which uses its dedicatedly pretrained VITAL-Base . [Per column: highest in boldface , second in italics , third underlined .]

PreTrained on LSVQ (train)+Finetuned on KVQ
Datasets	LSVQ (test)		LSVQ (1080p)		KoNViD-1K		YT-UGC		YT-Gaming		KVQ		FINEVD		AIGVQADB		GAIA
# test videos	7,182		3,573		1,200		1,098		600		2,926		1,016		2,808		9,180
Models	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$
FAST-VQA-B	0.689	0.688	0.532	0.557	0.542	0.556	0.657	0.662	0.678	0.703	-	-	0.698	0.719	0.296	0.378	0.110	0.107
\rowcolor light-gray0 DOVER	0.789	0.771	0.677	0.703	0.663	0.667	0.774	0.773	0.686	0.736	-	-	0.704	0.735	0.317	0.319	0.188	0.185
KVQ	0.764	0.762	0.662	0.716	0.702	0.720	0.769	0.765	0.714	0.732	-	-	0.734	0.756	0.323	0.323	0.210	0.195

Table 3: Transfer learning performance of VQA models trained on top of LSVQ (train) on different source datasets and evaluated on multiple test datasets. Intra-dataset evaluation is not conducted.

Models	YT-UGC	YT-Gaming	FINEVD	AIGVQA
DOVER++ (Ref.)	0.696 / 0.705	0.638 / 0.688	0.656 / 0.700	0.714 / 0.704
\rowcolor light-gray0 DOVER (Tech.)	0.787 / 0.787	0.701 / 0.771	0.719 / 0.742	0.754 / 0.738
DOVER (Aes.)	0.762 / 0.760	0.673 / 0.733	0.681 / 0.712	0.728 / 0.716
\rowcolor light-gray0 DOVER++ (All)	0.802 / 0.805	0.746 / 0.811	0.741 / 0.784	0.782 / 0.768

Table 4: Evaluation on the joint training effects.

Training	# Videos	YT-UGC	YT-Gaming	FINEVD	AIGVQADB
L1	$5,772$	0.799 / 0.806	0.719 / 0.753	0.727 / 0.766	0.741 / 0.758
\rowcolor light-gray0 L2+L3	$2,227$	0.679 / 0.665	0.647 / 0.658	0.573 / 0.551	0.664 / 0.673
L4+L5	$2,001$	0.713 / 0.723	0.689 / 0.734	0.641 / 0.622	0.788 / 0.791
\rowcolor light-gray0 L1+L2+L3	$7,999$	0.793 / 0.801	0.725 / 0.781	0.721 / 0.752	0.732 / 0.708
L1+L4+L5	$7,773$	0.776 / 0.785	0.716 / 0.770	0.698 / 0.766	0.778 / 0.770
\rowcolor light-gray0 All	$10,000$	0.802 / 0.805	0.746 / 0.811	0.741 / 0.784	0.782 / 0.768

Table 5: Performance of training using DOVER++ setting with different AIGC-trace levels ( L1

\rightarrow

L5 : low

\rightarrow

high).

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Datasets for UGC/PGC-VQA	Year	# Videos	# MOS	Description
LIVE-VQA	2010	160	160	Full-reference video quality rating
CVD2014	2016	234	234	Quality assessment of video captured by cameras
LIVE-Qualcomm	2017	208	208	Mobile in-capture video quality rating
KoNViD-1K	2017	1,200	1,200	Unified UGC video quality rating
LIVE-NFLX-I	2017	558	558	Quality-of-experience (QoE) rating of hand-craft streaming videos
LIVE-VQC	2018	585	585	Quality rating of real world UGC videos

Appendix

Table 6: Summary of existing UGC/PGC VQA MIDBs.

Datasets for AIGC VQA	Year	# Videos	# MOS	Description
T2VQA-DB	2024	10,000	10,000	Subjective-aligned quality assessment of text-to-video generated videos
LGVQ	2024	1,800	5,400	Multi-dimensional quality assessment of AIGC videos
AIGVQA-DB	2024	2,808	8,424	Large-scale perceptual quality assessment of AI-generated videos
AGAVQA-3k	2024	3,382	10,146	Multi-dimensional audio-visual quality assessment
GAIA	2024	9,180	9,180	Action quality assessment of AI-generated videos
TDVE-DB	2025	3,857	11,571	Quality assessment of text-driven video editing

Appendix

Table 7: Summary of existing AIGC VQA datasets.

Dataset part	Video source	# Videos
Trend-driven part (7000)	YouTube 2025.10	588
	YouTube 2025.11	667
	YouTube 2025.12	675
	YouTube 2026.1	1,023
	YouTube 2026.2	1,125
	YouTube 2026.3	1,860

Appendix

Table 8: Composition of the TREND-10K dataset.

Figure 5: Subjective Experiment Guideline, part 1

Figure 6: Subjective Experiment Guideline, part 2

Figure 7: Details for trending profile and sampling plan construction process, part 1

Figure 8: Details for trending profile and sampling plan construction process, part 2

Figure 10: Word cloud for supplementary T2V generation prompts.

Figure 11: Histogram of Word Count in generated prompts

Figure 12: All videos duration histogram

Figure 17: Details for Online Search and Video Crawling Process, part 1.

Figure 18: Details for Online Search and Video Crawling Process, part 2.

Figure 19: Details for T2V Generation Process.

Figure 20: Visualization for examples of multiple AI trace-Technical dimensions levels

Figure 21: Visualization for examples of multiple AI trace-Aesthetic dimensions levels

Figure 22: Aesthetic score distribution for all annotators

Figure 23: Technical score distribution for all annotators

Figure 24: AIGC-trace score distribution for all annotators

Video Quality Assessment

Video quality assessment measures perceived visual and temporal fidelity in captured, compressed, streamed, and generated videos.

See all 52 papers

VGA-Bench: A Unified Benchmark and Multi-Model Framework for Video Aesthetics and Generation Quality Evaluation
Apr 11, 2026Longteng Jiang, DanDan Zheng, Qianqian Qiao +7Video Generation Evaluation Video Quality Assessment
Ant Group · Beijing Film Academy · State Key Laboratory of General Artificial Intelligence, BIGAI
Comparison Drives Preference: Reference-Aware Modeling for AI-Generated Video Quality Assessment
Apr 18, 2026Minghao Zou, Gen Liu, Guanghui Yue +5Video Quality Assessment Video Generation Evaluation
Cardiff University Cardiff, UK · Shandong University of Science and Technology Qingdao, China · Shenzhen University Shenzhen, China +2
Analytic Score Optimization for Multi Dimension Video Quality Assessment
Feb 18, 2026Boda Lin, Yongjie Zhu, Wenyu Qin +2Video Quality Assessment
Kling Team, Kuaishou Technology, Beijing, China.

Apr 11, 2026J/K move · Enter open · S save

VGA-Bench: A Unified Benchmark and Multi-Model Framework for Video Aesthetics and Generation Quality Evaluation

Longteng Jiang, DanDan Zheng, Qianqian Qiao, Heng Huang+6

Ant Group · Beijing Film Academy · State Key Laboratory of General Artificial Intelligence, BIGAI

Video Generation Evaluation Video Quality Assessment

Open paper AskPDF arXiv

The rapid advancement of AIGC-based video generation has underscored the critical need for comprehensive evaluation frameworks that go beyond traditional generation quality metrics to encompass aesthetic appeal. However, existing benchmarks remain largely focused on technical fidelity, leaving a significant gap in holistic assessment-particularly with respect to perceptual and artistic qualities. To address this limitation, we introduce VGA-Bench, a unified benchmark for joint evaluation of video generation quality and aesthetic quality. VGA-Bench is built upon a principled three-tier taxonomy: Aesthetic Quality, Aesthetic Tagging, and Generation Quality, each decomposed into multiple fine-grained sub-dimensions to enable systematic assessment. Guided by this taxonomy, we design 1,016 diverse prompts and generate a large-scale dataset of over 60,000 videos using 12 video generation models, ensuring broad coverage across content, style, and artifacts. To enable scalable and automated evaluation, we annotate a subset of the dataset via human labeling and develop three dedicated multi-task neural assessors: VAQA-Net for aesthetic quality prediction, VTag-Net for automatic aesthetic tagging, and VGQA-Net for generation and basic quality attributes. Extensive experiments demonstrate that our models achieve reliable alignment with human judgments, offering both accuracy and efficiency. We release VGA-Bench as a public benchmark to foster research in AIGC evaluation, with applications in content moderation, model debugging, and generative model optimization.

Abstract

Video Quality Assessment

Explore similar work

VGA-Bench: A Unified Benchmark and Multi-Model Framework for Video Aesthetics and Generation Quality Evaluation

Comparison Drives Preference: Reference-Aware Modeling for AI-Generated Video Quality Assessment

Analytic Score Optimization for Multi Dimension Video Quality Assessment