AnimationBench: Are Video Models Good at Character-Centric Animation?

Leyi Wu*, Pengjun Fang*, Kai Sun*, Yazhou Xing, Yinwei Wu, Songsong Wang, Ziqi Huang, Dan Zhou, Yingqing He, Ying-Cong Chen, Qifeng Chen

The Hong Kong University of Science and Technology, The Hong Kong University of Science and Technology (Guangzhou), Nanyang Technological University, New AI Labs, Pearl Studio

Project Page | Paper

AnimationBench is a benchmark for evaluating image-to-video models on character-centric animation. Unlike realism-oriented video benchmarks, AnimationBench focuses on what matters most in animation: stylized character appearance, expressive motion, animation principles, and long-range character consistency.

The benchmark organizes evaluation into three major pillars: Disney's 12 Principles of Animation, IP / Character Preservation, and Broader Generative Video Quality. It provides a more informative way to compare modern video generation models with respect to animation-specific behavior and is designed to better reflect human preferences in character-centric animation scenarios.

demo.mp4

Overview

• A dedicated benchmark for character-centric animation evaluation.
• Covers 19 evaluation dimensions across animation principles, character preservation, and broader video quality.
• Includes a diverse benchmark suite with customized prompts, curated character assets, and animation-style test cases.
• Supports both standardized evaluation and flexible diagnostic analysis.
• Shows strong alignment with human preference while exposing quality gaps that realism-focused benchmarks often miss.

What AnimationBench Evaluates

AnimationBench goes beyond generic video quality metrics and focuses on animation-specific properties that are central to character performance:

• Animation Principles: e.g., squash and stretch, anticipation, slow in and slow out, arcs, secondary action, appeal, exaggeration, and related motion cues.
• IP / Character Preservation: whether a generated character remains on-model in appearance, behavior, personality, and style.
• Broader Quality Dimensions: semantic consistency, motion rationality, camera motion consistency, and other general video quality factors.

Evaluation Results

We report results following our top-down hierarchy:

• IP Preservation
• Animation Principles
• Broader Quality Dimensions.

Abbrev.: Antic. = Anticipation; FTOA = Follow Through and Overlapping Action; SI/SO = Slow In and Slow Out; S&S = Squash and Stretch; DC = Distinctive Content; SD = Solid Drawing; DD = Dynamic Degree; SE = Semantic Extension; Sem. = Semantic Consistency; MR = Motion Rationality; CMC = Camera Motion Consistency.

Human Alignment Results

Main Findings

• AnimationBench reveals clear differences among modern video models on character-centric animation tasks.
• Current models are relatively strong on basic semantic alignment and camera motion consistency while fine-grained character behavior, expressive motion, and long-term IP preservation remain challenging.
• The benchmark provides a more discriminative view of animation quality than realism-oriented evaluation settings.

Citation

If you find AnimationBench useful in your research, please consider citing:

@misc{wu2026animationbenchvideomodelsgood,
      title={AnimationBench: Are Video Models Good at Character-Centric Animation?}, 
      author={Leyi Wu and Pengjun Fang and Kai Sun and Yazhou Xing and Yinwei Wu and Songsong Wang and Ziqi Huang and Dan Zhou and Yingqing He and Ying-Cong Chen and Qifeng Chen},
      year={2026},
      eprint={2604.15299},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.15299}, 
}