WorldScore: Unified World Generation Benchmark

• WorldScore is a unified benchmark that evaluates world generation by decomposing it into sequential next-scene tasks with explicit camera trajectory specifications.
• It measures model performance across controllability, quality, and dynamics using ten normalized metrics to assess multi-scene coherence, geometric stability, and motion accuracy.
• The benchmark leverages a curated dataset of 3,000 examples covering diverse static and dynamic scenes, serving as a diagnostic framework for evolving world-generation models.

WorldScore is a unified evaluation benchmark for world generation that decomposes world generation into a sequence of next-scene generation tasks with explicit camera trajectory-based layout specifications, enabling unified evaluation of diverse approaches from 3D and 4D scene generation to video generation models. It was introduced as the first unified benchmark for world generation, with a curated dataset of 3,000 test examples spanning static and dynamic, indoor and outdoor, and photorealistic and stylized worlds, and with metrics organized around controllability, quality, and dynamics (Duan et al., 1 Apr 2025).

1. Definition and evaluation target

WorldScore was proposed to address a gap left by prior benchmarks that only evaluate single-scene or video generation models. Its stated objective is to evaluate models' abilities to create large-scale, multi-scene, diverse worlds, to enforce explicit layout and camera trajectory control, and to allow direct comparison across 3D, 4D, text-to-video (T2V), and image-to-video (I2V) methods (Duan et al., 1 Apr 2025).

The benchmark formalizes world generation as sequential next-scene generation rather than as isolated image or video synthesis. This framing is important because it makes layout adherence, multi-scene coherence, and dynamic motion part of the evaluation target. In the terminology of the benchmark, a world-generation model is judged not only on perceptual fidelity but also on whether it follows camera and object instructions and whether motion occurs where it should occur.

A later comparison in embodied AI positions WorldScore as a prior attempt at a "unified" world model evaluation, but one specifically focused on generative video quality rather than on downstream embodied functionality. That later characterization does not negate WorldScore's role; rather, it locates the benchmark within a broader progression from perceptual evaluation toward combined perceptual-functional evaluation (Shang et al., 9 Feb 2026).

2. Task formulation and layout specification

WorldScore decomposes world generation into a sequential series of next-scene generation tasks. Each task is characterized by a triplet $(\mathcal{C}, \mathcal{N}, \mathcal{L})$, where $\mathcal{C} = \{\mathbf{I}, \mathcal{P}\}$ is the current scene defined by an image $\mathbf{I}$ and text prompt $\mathcal{P}$, $\mathcal{N}$ is the next-scene text prompt describing a new scene or dynamic changes, and $\mathcal{L} = \{\mathcal{T}, \mathcal{Y}\}$ is the layout comprising a camera trajectory $\mathcal{T}$ and a textual description of the camera movement $\mathcal{Y}$ (Duan et al., 1 Apr 2025).

The benchmark writes the model interface as

$$\mathbf{V} = g_\text{world}(w_\text{proc}(\mathcal{C}, \mathcal{N}, \mathcal{L})).$$

Here, $g_\text{world}$ denotes the model and $\mathcal{C} = \{\mathbf{I}, \mathcal{P}\}$0 the model-specific preprocessing pipeline. This modular structure accommodates both static world expansions, where new scenes are added, and dynamic motion within a scene.

Layout specification is a central design choice. Camera trajectories $\mathcal{C} = \{\mathbf{I}, \mathcal{P}\}$1 are sequences of camera matrices $\mathcal{C} = \{\mathbf{I}, \mathcal{P}\}$2 that specify precise 3D movement or viewpoints for models that accept geometric inputs. Camera movement text $\mathcal{C} = \{\mathbf{I}, \mathcal{P}\}$3 describes the intended motion for T2V models that do not directly accept matrices but do respond to textual prompting. The benchmark provides 8 industry-standard camera movements, including pan, dolly, and pull-out, balanced across spatial directions and types to support both intra-scene and inter-scene transitions (Duan et al., 1 Apr 2025).

This decomposition suggests a benchmark philosophy in which control is not a peripheral attribute but part of the problem definition itself. A plausible implication is that WorldScore treats world generation as instruction-conditioned scene transition under explicit viewpoint constraints, rather than as unconstrained visual continuation.

3. Dataset composition and construction

The WorldScore dataset contains 3,000 high-quality test examples. Of these, 2,000 are static world generation examples and 1,000 are dynamic world generation examples. The static subset covers 5 indoor and 5 outdoor scene categories, and the dynamic subset covers 5 motion types: rigid, fluid, articulable, deformable, and multi-motion (Duan et al., 1 Apr 2025).

The benchmark emphasizes diversity along multiple axes. It spans indoor and outdoor settings, static and dynamic worlds, and photorealistic and stylized content. For every photorealistic example, a stylized counterpart is generated using one of 7 curated styles, including anime, impressionism, and minecraft. Source images are curated from SUN-RGBD, Matterport3D, LHQ, and Unsplash, then filtered for diversity, composition, and quality via CLIP-IQA, CLIP Aesthetic, perspective analysis, and manual review. Next-scene texts are generated using GPT-4o via an auto-regressive process so that scenes remain coherent and distinct in multi-scene expansions. Layout assignment is randomized per example, while for dynamic worlds the camera is fixed (Duan et al., 1 Apr 2025).

These construction choices define WorldScore as a benchmark over world specifications rather than over raw videos alone. The benchmark therefore evaluates whether a model can transform a structured specification—image, prompt, next-scene instruction, and layout—into a sequence that is both visually plausible and instruction-consistent.

4. Metric system and score aggregation

WorldScore aggregates 10 individual metrics under three core aspects: controllability, quality, and dynamics. Metrics are normalized using empirical bounds and mapped to $\mathcal{C} = \{\mathbf{I}, \mathcal{P}\}$4, after which overall scores are computed by averaging. WorldScore-Static is the average of controllability and quality metrics, while WorldScore-Dynamic adds the three dynamics metrics (Duan et al., 1 Apr 2025).

The metric system is summarized below.

Aspect	Metric	Brief description
Controllability	Camera Controllability	Adherence to intended camera motion from GT and estimated trajectory
Controllability	Object Controllability	Success rate of generating specified foreground objects
Controllability	Content Alignment	CLIPScore against the full next-scene prompt
Quality	3D Consistency	Geometric stability across frames via dense reconstructions and reprojection error
Quality	Photometric Consistency	Appearance stability via bidirectional optical flow and AEPE
Quality	Style Consistency	Difference between Gram matrices of first and last frame
Quality	Subjective Quality	CLIP-IQA+ and CLIP-Aesthetic calibrated with human judgments
Dynamics	Motion Accuracy	Whether motion occurs in regions that should be dynamic
Dynamics	Motion Magnitude	Median optical flow value
Dynamics	Motion Smoothness	Frame interpolation error using VFIMamba with MSE, SSIM, and LPIPS

The camera-controllability component is defined through geometric error between the ground-truth trajectory and an estimated trajectory. The benchmark gives

$\mathcal{C} = \{\mathbf{I}, \mathcal{P}\}$5

with rotation error

$\mathcal{C} = \{\mathbf{I}, \mathcal{P}\}$6

Pose estimates are obtained via DROID-SLAM (Duan et al., 1 Apr 2025).

Several metrics target distinct failure modes. Object Controllability measures whether specified foreground objects are generated at all, whereas Content Alignment uses CLIPScore to capture alignment beyond object presence. 3D Consistency and Photometric Consistency separate geometric stability from appearance stability. Motion Accuracy measures whether motion is localized to intended dynamic regions, while Motion Magnitude measures how much motion occurs. This separation is consequential because the benchmark explicitly reports that motion magnitude is not the same as motion accuracy (Duan et al., 1 Apr 2025).

5. Evaluated model classes and empirical findings

WorldScore reports extensive evaluation of 19 representative models, including 12 video models, 6 3D scene models, and 1 4D scene model. The video set includes 10 open-source and 2 commercial models; the 3D set includes models such as LucidDreamer and WonderWorld; and the 4D category is represented by 4D-fy (Duan et al., 1 Apr 2025).

Several findings recur in the benchmark analysis. First, 3D scene generation models excel in static world generation, with high camera controllability and strong geometric and appearance consistency; WonderWorld and LucidDreamer achieve the highest WorldScore-Static. Second, video generation models lag in camera controllability and are prone to layout-adherence errors, especially for long-sequence or outdoor tasks. Third, the best open-source video models, exemplified by CogVideoX-I2V, match or surpass commercial ones in overall WorldScore, although commercial models may dominate certain sub-metrics such as object and content alignment. Fourth, a trade-off exists between motion magnitude and motion smoothness: high-motion models often sacrifice temporal stability. Fifth, T2V models are easier to steer via text prompts than I2V models, which tend to anchor to the provided input-image view. Finally, 4D-fy underperforms, which the benchmark presents as evidence of fundamental challenges in dynamic world modeling (Duan et al., 1 Apr 2025).

These findings establish WorldScore as more than a leaderboard. It functions as a diagnostic framework for model categories. Its results indicate that strong perceptual output on isolated clips does not guarantee camera control, multi-scene coherence, or reliable dynamic behavior under explicit layout constraints.

6. Subsequent adoption, reinterpretation, and limitations

WorldScore has been used as a reference benchmark in later world-generation work. NeoWorld reports that it significantly outperforms existing 2D and depth-layered 2.5D methods on the WorldScore benchmark and describes an evaluation setting with 28 images spanning 7 styles, 4 generated 3D environments per image, and standardized panoramic and orbiting camera trajectories, with metrics including CIQA+, Q-Align, 3D-Const, SceneQuality, ImageCLIP, CLIP Score, IoU, and Prompt Align/VideoCLIP (Zhao et al., 29 Sep 2025). In that evaluation, NeoWorld reports the highest 3D Consistency and SceneQuality among the compared 3D world-generation methods, as affirmed by human studies (Zhao et al., 29 Sep 2025).

INSPATIO-WORLD uses WorldScore-Dynamic as a primary quantitative benchmark and reports ranking first among real-time interactive methods. Its paper states that evaluation is performed using the official set of ten core WorldScore metrics and highlights best overall camera control and photometric scores in its comparison table (Team et al., 8 Apr 2026). This later usage is significant because it shows that WorldScore-Dynamic functions as an accepted target for real-time interactive 4D simulation, not only for offline world generation.

At the same time, later benchmark work identifies scope limitations. WorldArena explicitly compares itself to WorldScore and characterizes WorldScore as focused mainly on video fidelity and as not comprehensively assessing functional or embodied roles such as use as data engines, policy evaluators, and action planners (Shang et al., 9 Feb 2026). That comparison frames a common misconception: a unified benchmark for world generation is not automatically a unified benchmark for embodied world models. WorldScore unifies evaluation across 3D, 4D, T2V, and I2V generation under explicit scene-transition specifications; it does not, on the evidence presented by later work, fully subsume downstream functionality evaluation.

The benchmark's public infrastructure includes a dataset, evaluation code, and leaderboard at https://haoyi-duan.github.io/WorldScore/ (Duan et al., 1 Apr 2025). This public release has supported comparative evaluation across model families and has made WorldScore a reference point in subsequent literature on world generation and interactive world simulation.

7. Significance within world-generation research

WorldScore's principal contribution is methodological unification. By casting world generation as next-scene generation with explicit layout specification, it makes 3D scene generators, 4D scene generators, and video generators comparable within one evaluation protocol. Its three-aspect metric design—controllability, quality, and dynamics—formalizes a view of world generation in which camera control, object realization, geometric consistency, appearance stability, and motion behavior are jointly measurable (Duan et al., 1 Apr 2025).

The benchmark also surfaces structural trade-offs across model classes. The reported results distinguish models that are visually strong but weak in control, models that are stable in static scenes but limited in dynamic worlds, and models that produce large motion without accurate localization of that motion. This suggests that the field's central difficulties are not exhausted by photorealism alone.

In the subsequent literature, WorldScore serves both as an evaluation target and as a conceptual baseline. NeoWorld uses it to substantiate improvements in explorable interactive 3D world generation (Zhao et al., 29 Sep 2025); INSPATIO-WORLD uses WorldScore-Dynamic to quantify real-time 4D interactive performance (Team et al., 8 Apr 2026); and WorldArena uses it as a point of comparison when arguing for the addition of embodied functionality metrics (Shang et al., 9 Feb 2026). Taken together, these uses indicate that WorldScore has become a reference benchmark for controllable world generation, while also clarifying the boundary between world-generation evaluation and broader embodied-world-model evaluation.