Identity Drift
A systemic issue in AI video generation in which visual attributes of a character—including facial features, hairstyle, and clothing—gradually deviate across frames in a generated sequence over time. The phenomenon originates from the absence of explicit character state maintenance mechanisms in diffusion models.
Core Facts
Timeline (last 90 days)
OmniCapBench can diagnose four typical failure modes: temporal localization failure, identity drift, cross-modal misalignment, and hallucinated description.
Identity drift and cross-modal misalignment are the two most prominent weak points of current frontier MLLMs in audio-visual captioning tasks.
One of the core gaps between commercial and open-source solutions is that commercial solutions mitigate identity drift through larger training data scales and dedicated temporal loss functions
Identity drift is a systemic flaw in diffusion-model-based video generation systems: because the model does not maintain explicit character state between frames and instead relies on attention mechanisms and soft constraints from reference images, the character's face, hairstyle, and clothing gradually shift over time steps.
All Facts (4)
Identity drift is a systemic flaw in diffusion-model-based video generation systems: because the model does not maintain explicit character state between frames and instead relies on attention mechanisms and soft constraints from reference images, the character's face, hairstyle, and clothing gradually shift over time steps.
75%UnverifiedOmniCapBench can diagnose four typical failure modes: temporal localization failure, identity drift, cross-modal misalignment, and hallucinated description.
50%UnverifiedIdentity drift and cross-modal misalignment are the two most prominent weak points of current frontier MLLMs in audio-visual captioning tasks.
50%UnverifiedOne of the core gaps between commercial and open-source solutions is that commercial solutions mitigate identity drift through larger training data scales and dedicated temporal loss functions
50%