GroupVideo: Multi-Identity Customized Text-to-Video Generation

Existing text-to-video models often struggle when tasked with generating videos featuring multiple distinct characters. Previous attempts to extend single-identity frameworks to multi-identity settings typically result in the copy-paste phenomenon, where characters appear unnaturally superimposed, exhibit rigid expressions, or suffer from identity confusion. This paper asks how to effectively maintain identity fidelity, motion naturalness, and spatial-temporal consistency for multiple characters simultaneously.

The authors introduce GroupVideo, a framework built upon Video Diffusion Transformers (DiTs). To solve the multi-identity challenge, the model employs three primary innovations:

1. Multimodal Identity Alignment: Instead of simply concatenating face images, the model decouples identity features into visual and semantic branches. Visual tokens are integrated into the latent space, while a semantic perceiver fuses identity embeddings with textual descriptions to ensure motion coherence.
2. ID Localization Module: This module learns to predict spatial-temporal masks that assign specific identity features to corresponding regions in the video, preventing identity blending.
3. Progressive Training & Dataset: The authors curated a new high-quality dataset of 20,000 multi-person videos and implemented a two-stage training strategy using bounding box constraints and mask regularization to improve training efficiency and identity preservation.

Multi-identity video generation is a critical frontier for personalized content creation. By moving beyond the limitations of simple condition-stacking, GroupVideo provides a more scalable and robust solution for complex scenes involving multiple characters. The release of the 20,000-video dataset also provides a valuable resource for future research in this domain, addressing a significant data bottleneck in the field.