TL;DR

While flagship AI models like Sora and Kling can generate breathtaking visuals, they often fail the "physicality test." HumanScore, a new benchmark from Stanford University, moves past pixel-realism to evaluate whether AI-generated humans actually move like biological beings. By extracting 3D skeletons and applying biomechanical laws, it provides a rigorous ranking of current SOTA models, showing that even the best generators still struggle with basic anatomical consistency.

The "Glitch" in the Matrix: Why Visual Realism Isn't Enough

If you look at a modern AI video of a ballerina, the lighting and textures are stunning. However, look closer at the limbs: do the bones stay the same length during a turn? Does the knee bend at an angle that would snap a human tendon?

Traditional benchmarks like VBench or FVD (Fréchet Video Distance) are "appearance-centric." They check if the pixels look right, but they don't understand that a human has a rigid skeleton and mass. HumanScore addresses this by shifting the evaluation from 2D pixels to 3D structural dynamics.

Methodology: The Three-Tier Biomechanical Hierarchy

The authors propose a structured evaluation that mimics the hierarchy of human movement science:

1. Anatomical Correctness (The Foundation): Uses artifacts detectors (like HADM) to find "ghost" limbs and analyzes if bone segments remain rigid over time.
2. Kinematic Correctness (The Geometry): Maps video motion to an OpenSim skeleton to check if joint angles stay within biological limits (Range of Motion) and uses 3D mesh intersection tests to detect self-collision (e.g., an arm passing through a torso).
3. Kinetic Correctness (The Physics): Since we can't measure muscle force directly from video, the authors use $F=ma$ (Newton's Second Law). They track velocity and "jerk" (the rate of change of acceleration) to penalize unnatural jitters and physics-defying speed spikes.

Fig 1: The HumanScore pipeline: From motion curation and prompt design to multifaceted biomechanical evaluation.

Exploring the Leaderboard: Who Moves Best?

The study benchmarked 13 models, including proprietary giants and open-source contenders.

• The Leaders: Seedance 1.0 Pro and HunyuanVideo 1.5 co-lead the board with an overall score of 91.1.
• The Gap: Real human videos score 94.3, highlighting a persistent gap.
• Specific Strengths: KlingAI showed the highest Kinetic and Kinematic scores, meaning its motions are physically "smoother," while HunyuanVideo 1.5 excelled in Anatomy, generating fewer "spurious limbs."

Table 1: The HumanScore Leaderboard. Note how real videos still lead, but Top-AI models are closing the gap.

Key Insights & Failure Modes

The study identified several recurring "AI-isms":

• Temporal Drift: Parts of the body (like feet) tend to "slide" or change scale when the model focuses on complex movements.
• The "Intensity" Penalty: As the motion becomes more intense (e.g., from walking to sprinting), all models experience a significant drop in score. Harder motions like ballet or parkour act as the ultimate "stress test" for generative physics.
• Prompting isn't a silver bullet: Even with highly detailed "prompt engineering" to specify full-body, static camera shots, the underlying generative engines still struggle to maintain physical constraints.

Conclusion and Future Outlook

HumanScore is a wake-up call for the generative AI community. It proves that to move from "fun creative tools" to "reliable simulation tools" for industries like sports science, medicine, or high-end filmmaking, we need physics-informed architectures.

The modular design of HumanScore allows it to evolve as better 3D pose estimators (like PromptHMR) emerge. For researchers, the goal is now clear: don't just generate a human—generate a human that obeys the laws of biology.

Note: This benchmark provides the first technical "compass" for navigating the complex intersection of human motion reconstruction and video generation.