DyMD: Preserving Interaction Dynamics through Distribution Matching Distillation in Few-Step Video World Models

TL;DR

DyMD improves few-step video generation in world models by refining distribution matching distillation through adaptive teacher supervision and dynamic fake-score tracking. It enhances motion fidelity in robot-object interactions without sacrificing visual quality, enabling more effective downstream planning.

Why It Matters

Few-step video generation is critical for real-time interactive systems like robotics, but existing methods struggle with motion accuracy. DyMD enables high-quality, fast video prediction that supports better action planning, making it practical for embodied AI applications.

Key Findings

• Weak re-noising in standard DMD concentrates the teacher posterior around motion-poor rollouts, limiting motion recovery guidance.
• Stronger-motion rollouts incur higher fake-score fitting errors, which disrupts learning of interaction dynamics in the student model.
• DyMD improves performance by using temporal affinity to adapt re-noising schedules and a noise-conditioned predictor to upweight difficult rollouts in critic training.