High-quality synthetic interaction generation

Develop methods for generating high-quality, diverse, and physically plausible synthetic interactions for video generation models.

Background

The paper addresses the difficulty of producing video-generation data that depicts realistic physical interactions and consequent object-state transitions. Existing auxiliary-condition methods may lack generalizability, while real-world data-driven methods are constrained by the limited scope and high curation cost of available datasets. The proposed framework uses structured interaction taxonomies, state-conditioned image synthesis, and State-Guided Sampling to generate and curate synthetic interaction videos, but the broader challenge of producing such data remains unresolved.