Towards Customized Multimodal Role-Play

Researchers introduced Customized Multimodal Role-Play, a new task for making AI characters consistent in both text and images.

They also released RoleScape-20, a dataset with 20 characters designed for few-shot character customization and cross-modal alignment.

The proposed UniCharacter framework uses a two-stage training process to learn persona, style, and visual identity from limited examples.

UniCharacter outperforms prior methods on the benchmark, improving consistency and immersion for multimodal role-play agents.