Is Visual Realism Enough? Evaluating Gait Biometric Fidelity in Generative AI Human Animation

TL;DR

The paper investigates whether state-of-the-art GenAI-based human animation preserves gait-based behavioral biometrics, beyond mere visual realism. It compares four GenAI animation models across two tasks—gait restoration and zero-shot identity transfer—using four gait-recognition models and two datasets to assess biometric fidelity. The findings reveal a clear gap: high visual fidelity does not translate into robust gait-identification cues, with many models relying on appearance rather than temporal dynamics, and only a few approaches show limited motion transfer under controlled conditions. The work highlights the need for biometric-aware training objectives and temporal-consistency strategies to align visual quality with faithful, secure gait biometrics in synthetic videos.

Abstract

Generative AI (GenAI) models have revolutionized animation, enabling the synthesis of humans and motion patterns with remarkable visual fidelity. However, generating truly realistic human animation remains a formidable challenge, where even minor inconsistencies can make a subject appear unnatural. This limitation is particularly critical when AI-generated videos are evaluated for behavioral biometrics, where subtle motion cues that define identity are easily lost or distorted. The present study investigates whether state-of-the-art GenAI human animation models can preserve the subtle spatio-temporal details needed for person identification through gait biometrics. Specifically, we evaluate four different GenAI models across two primary evaluation tasks to assess their ability to i) restore gait patterns from reference videos under varying conditions of complexity, and ii) transfer these gait patterns to different visual identities. Our results show that while visual quality is mostly high, biometric fidelity remains low in tasks focusing on identification, suggesting that current GenAI models struggle to disentangle identity from motion. Furthermore, through an identity transfer task, we expose a fundamental flaw in appearance-based gait recognition: when texture is disentangled from motion, identification collapses, proving current GenAI models rely on visual attributes rather than temporal dynamics.

Figures (6)

• Figure 1: Motivation. We examine whether synthetic videos generated through state-of-the-art GenAI animation models successfully preserve the identity-specific behavioral biometric traits (gait). Both visual and biometric fidelity are analyzed in the study.
• Figure 2: Overview of the generation and evaluation pipeline. The pipeline extracts different motion guidance (Skeleton, SMPL, DensePose) from a driving video to condition four state-of-the-art animation models. To assess biometric fidelity, we extract specific gait modalities (Silhouettes, Skeleton Maps, RGB) from the generated outputs and subject them to four different gait recognition models. This allows us to quantify identity preservation and visual quality.
• Figure 3: Overview of the two primary evaluation tasks comprising different scenarios. Task 1 evaluates gait restoration across (A) Baseline, (B) Off-Pose, and (C) In-the-Wild scenarios. Task 2 examines zero-shot identity transfer. The Evaluation Method block details the dual validation protocol against reference (appearance) and driving (motion) identities.
• Figure 4: Failures produced by different GenAI models across different scenarios. We visualize the specific artifacts that degrade biometric fidelity. We highlight the visual inconsistencies, and the behavioral inconsistencies.
• Figure 5: Task 1: Gait Restoration Biometric Performance. The comparison highlights a modality gap: AnimateAnyone excel in texture preservation (BigGait), while MagicAnimate dominates in motion fidelity (SkeletonGait). Performance consistently degrades in the unconstrained In-the-Wild scenario.