WorldMark: A Unified Benchmark Suite for Interactive Video World Models

Abstract

Interactive video generation models—Genie, YUME, HY-World, Matrix-Game, among others—are advancing rapidly, yet every model is evaluated on its own proprietary benchmark with bespoke scenes and trajectories, making fair cross-model comparison impossible. Existing public benchmarks offer useful metrics such as trajectory error, aesthetic scores, and VLM-based judgments, but none supplies the standardized test conditions—identical scenes, identical action sequences, and a unified control interface—needed to make those metrics comparable across models with heterogeneous inputs.

We introduce WorldMark, the first benchmark that provides such a common playing field for interactive Image-to-Video world models. WorldMark contributes: (1) a unified action-mapping layer that translates a shared WASD-style action vocabulary into each model's native control format, enabling apples-to-apples comparison across six major models on identical scenes and trajectories; (2) a hierarchical test suite of 500 evaluation cases covering first- and third-person viewpoints, photorealistic and stylized scenes, and three difficulty tiers from Easy to Hard spanning 20–60s; and (3) a modular evaluation toolkit for Visual Quality, Control Alignment, and World Consistency, designed so that researchers can reuse our standardized inputs while plugging in their own metrics as the field evolves. We release all data, evaluation code, and model outputs to facilitate future research.

Introduction Video

Overview

Current interactive world models each define their own control interfaces, scenes, and evaluation protocols, making cross-model comparison impossible. WorldMark resolves this fragmentation through three pillars: (a) A curated test suite of 500 cases spanning first- and third-person views with photorealistic and stylized scenes. (b) A unified action-mapping layer translating a shared WASD vocabulary into each model's native format. (c) Same reference image + same action sequence reveals striking quality divergence across models.

Why WorldMark?

Existing benchmarks lack standardized scenes, unified cross-model action mapping, keyboard-interactive support, or difficulty hierarchies. WorldMark is the first to provide all of these.

Image Suite

50 diverse reference images spanning three scene categories (Nature, City, Indoor), two visual styles (Real, Stylized), and two viewpoints (First-person, Third-person). The stylized subset covers oil painting, Ukiyo-e, cyberpunk, and Minecraft aesthetics, ensuring evaluation is not biased toward photorealistic domains alone. Cross-viewpoint pairs yield 100 test images total.

Action Suite

15 standardized action sequences of increasing complexity, expressed in a shared vocabulary of WASD movement plus L/R yaw rotation:

Combined with 100 test images, this yields approximately 500 standardized evaluation cases. A VLM-based scene-aware filtering step ensures only physically plausible actions are assigned to each image.

Quantitative Results

WorldMark evaluates generated videos along three complementary dimensions with eight metrics:

First-Person Real

First-Person Stylized

Third-Person

Only three models natively support third-person perspective: MatrixGame 2.0, HY-World, and Genie 3.

Key Takeaways

@inproceedings{worldmark2026,
  title={WorldMark: A Unified Benchmark Suite for Interactive Video World Models},
  author={Anonymous},
  journal={arXiv preprint},
  year={2026}
}