This paper presents HVG-4D, an extension of the Human Video Generation in 4D (HVG) model, to handle multi-character scenarios with complex interactions. HVG-4D introduces innovations like Multi-Character Pose Modulation, Interaction-Aware Temporal Alignment, and Progressive Multi-Character Sampling, achieving state-of-the-art performance in multi-character 4D video generation.
Key findings
HVG-4D extends HVG to multi-character 4D scene composition.
Introduces MCPM for handling inter-character spatial relationships.
Presents IATA for intra-character and inter-character interaction coherence.
Develops PMCS for consistency across varying character counts.
Enables flexible character editing in scenes with the Compositional Scene Graph module.
Limitations & open questions
The computational cost of generating multi-character scenes must be addressed for practical deployment.