Paired Multi-Identity Data for Group Photos: Construction and Effective Exploitation

Construct large-scale datasets of group photos that provide multiple reference images per identity, and develop training and inference methods that effectively exploit such paired multi-identity data to enable controllable, identity-consistent image generation while mitigating copy-paste artifacts.

Background

The paper identifies a widespread failure mode in identity customization models: excessive similarity to a single reference image leads to copy-paste artifacts that suppress natural variations in pose, expression, and lighting. A key cause is the scarcity of datasets that provide multiple reference images per identity, especially in multi-person scenes, which forces most methods into reconstruction-based training that reinforces copying behavior.

To address this gap, the authors introduce MultiID-2M, a large-scale dataset with paired references and group photos, and WithAnyone, a training paradigm that leverages contrastive identity losses and extended negatives. Despite these contributions, the broader challenge of building such datasets at scale and designing algorithms that fully exploit multi-reference, multi-identity data for controllable generation remains explicitly noted by the authors as open.