Multi-human image generation typically focuses on preserving the identity of the person being personalized, but other people in the same image may also be affected. A new paper on the arXiv characterizes this overlooked failure mode, showing that identity changes can occur in people outside the nominal target of the generative operation.

The study uses UniPortrait as a controlled testbed, deliberately avoiding the introduction of a new preservation algorithm. Instead, the authors measure how source-coupled multi-human generation leads to cross-person identity changes, providing a baseline for future work that aims to prevent such drift.

The findings highlight that identity preservation in multi-human settings is not just about the intended subject. The paper's emphasis on a controlled study suggests the effect is systematic and worth accounting for in the design of personalization systems.