CamWorldQA: Perceptual Quality Assessment of Camera-Controlled World Video Generation

Yunhe Li; Likun Wu; Sijing Wu; Xinyu Tian; Huiyu Duan; Yixuan Gao; Yunhao Li; Guangtao Zhai

cs.CVAug 19, 2026

CamWorldQA: Perceptual Quality Assessment of Camera-Controlled World Video Generation

Authors: Yunhe Li, Likun Wu, Sijing Wu, Xinyu Tian, Huiyu Duan, Yixuan Gao, Yunhao Li, Guangtao Zhai

Organizations: Shanghai Jiao Tong University · Eindhoven University of Technology

Abstract

Recent advances in generative video models have enabled camera-controlled world video generation, allowing models to synthesize videos under user-defined camera trajectories. However, existing video quality assessment (VQA) methods are mainly developed for natural videos and fail to capture the unique perceptual characteristics of camera-controlled generation, such as viewpoint consistency, motion coherence, and content preservation. In this work, we introduce CamWorldQA, the first benchmark for perceptual quality assessment of camera-controlled world video generation. CamWorldQA contains 720 generated videos produced by 6 representative generation methods from 20 diverse source videos under 6 camera trajectories, where each video is annotated with a human-rated perceptual quality score through subjective experiments. Furthermore, we propose CWQA, a no-reference quality assessment network with three complementary branches that extract spatial features, temporal motion features and optical flow features to jointly predict quality scores. Extensive experiments demonstrate that CWQA achieves superior performance over existing quality assessment methods on the CamWorldQA dataset.

Figures & tables

Figure 1. Overview of camera-controlled world video quality assessment, including camera trajectory definition, CamWorldQA dataset construction, and the proposed CWQA model.

Figure 2. MOS variations across (a) generation methods and (b) camera trajectories.

Figure 3. MOS distributions across (a) generation methods, (b) camera trajectories, and (c) video content categories.

Figure 4. MOS distribution of CamWorldQA.

Method	Mean MOS $\uparrow$	Std.	Min	Max	Rank
GEN3C ( Ren et al., 2025 )	61.88	10.49	23.73	77.59	1
ReDirector ( Park et al., 2026 )	56.31	10.59	23.63	73.09	2
SierpinskiCam ( Wizadwongsa et al., 2026 )	54.49	9.92	24.83	71.86	3
ReCamMaster ( Bai et al., 2025a )	53.84	10.74	24.08	69.90	4
Diffusion as Shader ( Gu et al., 2025 )	38.07	11.40	20.59	69.90	5
TrajectoryCrafter ( Yu et al., 2025 )	36.15	7.87	20.58	55.99	6

Table 1. MOS statistics of videos generated by the six video generation methods in CamWorldQA.

Figure 5. Overview of the proposed CWQA method.

Method	SRCC $\uparrow$	PLCC $\uparrow$	KRCC $\uparrow$
$\diamondsuit$ RAPIQUE ( Tu et al., 2021b )	0.6592	0.6545	0.4722
$\diamondsuit$ VIDEVAL ( Tu et al., 2021a )	0.3491	0.3677	0.2386
$\clubsuit$ SimpleVQA ( Sun et al., 2022 )	0.1816	0.1906	0.1212
$\clubsuit$ FastVQA ( Wu et al., 2022 )	0.4662	0.4615	0.3112
$\clubsuit$ DOVER ( Wu et al., 2023 )	0.2098	0.2590	0.1428
$\heartsuit$ Qwen3-VL (4B) ( Bai et al., 2025b )	0.2342	0.2526	0.1852

Table 2. Overall performance comparison on the CamWorldQA dataset.

\diamondsuit

\clubsuit

, and

\heartsuit

denote traditional VQA methods, learning-based VQA methods, and multimodal large language models, respectively. † indicates models trained on the CamWorldQA dataset. The best and runner-up performances are bold and underlined, respectively.

	Animals			Human			Objects			Scene
Method	SRCC $\uparrow$	PLCC $\uparrow$	KRCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	KRCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	KRCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	KRCC $\uparrow$
$\diamondsuit$ RAPIQUE ( Tu et al., 2021b )	0.7051	0.7430	0.5352	0.7096	0.7780	0.4970	0.5003	0.7114	0.3492	0.7156	0.6880	0.5116
$\diamondsuit$ VIDEVAL ( Tu et al., 2021a )	0.2948	0.2662	0.2135	0.3206	0.4011	0.2182	0.1970	0.5798	0.1429	0.5184	0.5230	0.3693
$\clubsuit$ DOVER ( Wu et al., 2023 )	0.1844	0.2516	0.1380	0.0392	0.0237	0.0242	0.3689	0.6359	0.2751	0.4226	0.3792	0.3063
$\clubsuit$ SimpleVQA ( Sun et al., 2022 )	0.2836	0.3574	0.2164	0.0460	0.3703	0.0364	0.0317	0.0480	0.0159	0.3968	0.4080	0.3171
$\clubsuit$ FastVQA ( Wu et al., 2022 )	0.3117	0.3893	0.2006	0.3138	0.3723	0.1855	0.4165	0.5740	0.2698	0.7097	0.8089	0.4957

Table 3. Category-wise performance comparison on the CamWorldQA dataset.

\diamondsuit

\clubsuit

, and

\heartsuit

Method	SRCC $\uparrow$	PLCC $\uparrow$	KRCC $\uparrow$
w/o Spatial Branch	0.7572	0.7553	0.5712
w/o Temporal Branch	0.7248	0.7099	0.5308
w/o Optical Flow	0.6263	0.6260	0.4478
w/o Adaptive Gating	0.7315	0.6636	0.5374
Full CWQA (Ours)	0.7804	0.7848	0.5739

Table 4. Ablation study of CWQA on CamWorldQA.

Figure 6. Qualitative results of the proposed CWQA method.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Canonical Motion	$t_{x}$	$t_{y}$	$t_{z}$
Truck Left	$-0.25$	$0$	$0$
Truck Right	$+0.25$	$0$	$0$
Pedestal Up	$0$	$+0.25$	$0$
Pedestal Down	$0$	$-0.25$	$0$
Dolly In	$0$	$0$	$-0.30$
Dolly Out	$0$	$0$	$+0.30$

Appendix

Table A . Canonical camera motions and their corresponding camera parameters.

Method	Year	Resolution	Frames (FPS)	Type
ReCamMaster	2025	1280 $\times$ 1280‡	81‡/15‡	Diff. (3D)
TrajectoryCrafter	2025	Inherited ∗	49†/10†	Diffusion
ReDirector	2025	832 $\times$ 480†	81†/30†	Diffusion
GEN3C	2025	704 $\times$ 1280‡	121‡/24†	3D world
Diffusion as Shader	2025	720 $\times$ 480‡	49 ${}^{*}/8{\dagger}$	3D-aware
SierpinskiCam	2026	832 $\times$ 480†	81†/12†	Diffusion

Appendix

Table B. Overview of the six camera-controlled video generation methods.

Figure A . Illustration of the GUI used in the subjective study

Panel of Figure 14 — Figure B . Demo videos of different mean opinion scores. For each video, its content category, camera trajectory, and MOS are annotated above the frames.

Video Quality Assessment

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

See all 52 papers

WorldJen: An End-to-End Multi-Dimensional Benchmark for Generative Video Models
May 5, 2026Karthik Inbasekar, Guy Rom, Omer ShlomovitsHuman Preference Evaluation Video Quality Assessment
WorldMark: A Unified Benchmark Suite for Interactive Video World Models
Apr 23, 2026Xiaojie Xu, Zhengyuan Lin, Kang He +5Action-Conditioned World Models World Model Evaluation
Alaya Studio, Shanda AI Research Tokyo · The University of Tokyo · Shanghai Innovation Institute
WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors
May 11, 2026Keming Wu, Yijing Cui, Wenhan Xue +11World Models Video Reasoning
Tsinghua University · Nanyang Technological University · Hong Kong University of Science and Technology (Guangzhou) +2

May 5, 2026J/K move · Enter open · S save

WorldJen: An End-to-End Multi-Dimensional Benchmark for Generative Video Models

Karthik Inbasekar, Guy Rom, Omer Shlomovits

Human Preference Evaluation Video Quality Assessment Vision-Language Model-as-a-Judge Video Generation Evaluation

Open paper AskPDF arXiv

Evaluating generative video models remains an open problem. Reference-based metrics such as Structural Similarity Index Measure (SSIM) and Peak Signal to Noise Ratio (PSNR) reward pixel fidelity over semantic correctness, while Frechet Video Distance (FVD) favors distributional textures over physical plausibility. Binary Visual Question Answering (VQA) based benchmarks like VBench~2.0 are prone to yes-bias and rely on low-resolution auditors that miss temporal failures. Moreover, their prompts target a single dimension at a time, multiplying the number of videos required while still not guaranteeing reliable results. WorldJen addresses these limitations directly. Binary VQA is replaced with Likert-scale questionnaires graded by a VLM that receives frames at native video resolution. Video generation costs are addressed by using adversarially curated prompts that are designed to exercise up to 16 quality dimensions simultaneously. The framework is built around two interlocking contributions. First, A blind human preference study is conducted, accumulating (2,696 pairwise annotations from 7 annotators with 100% pair coverage over 50 of the curated prompts

\times

6 state-of-the-art video models. A mean inter-annotator agreement of 66.9% is achieved and the study establishes a human ground-truth Bradley-Terry (BT) rating with a three-tier structure. Second, A VLM-as-a-judge evaluation engine using prompt-specific, dimension-specific Likert questionnaires (10 questions per dimension, 47,160 scored responses) judges the videos and reproduces the human-established three-tier BT rating structure independently. The VLM achieves a Spearman

\hatρ=1.000,~p=0.0014

that is interpreted as tier agreement with the human results. Six focused ablation studies validate the robustness of the VLM evaluation framework. Project page: https://moonmath.ai/worldjen/

Abstract

Video Quality Assessment

Explore similar work

WorldJen: An End-to-End Multi-Dimensional Benchmark for Generative Video Models

WorldMark: A Unified Benchmark Suite for Interactive Video World Models

WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors