KV-Control: Parameter-Efficient K/V Injection for Trajectory-Controlled Text-to-Motion
Researchers have developed KV-Control, a new method for parameter-efficient control of text-to-motion generation. This technique injects trajectory constraints as memory into the self-attention layers of frozen motion transformers. By co-designing a part-tokenized motion substrate called PartVQ, KV-Control allows for precise tracking of root and multi-joint trajectories while preserving the quality of text-conditioned motion. AI
IMPACT Enables more precise and efficient control over AI-generated motion, crucial for animation and embodied agents.