PRIMFLOW: Learning Trajectory Primitives for Few-Step Distillation

Key points

• Existing approaches construct within-step dynamics using specific trajectory parameterizations or discrete interval predictions, tying the learned update to a prescribed temporal structure.1
• We propose \method, which directly learns a continuous trajectory primitive representing cumulative displacement along the teacher trajectory.1
• By supervising finite differences of the primitive with local teacher velocities, \method learns the accumulated motion directly, while obtaining global updates through endpoint differences.1
• We further establish a local-to-global error bound to characterize how local regression errors affect full-interval displacement accuracy, and introduce endpoint absorption to eliminate the auxiliary primitive branch while preserving one-network-evaluation inference.1

Sources (1)

• [1]PRIMFLOW: Learning Trajectory Primitives for Few-Step Distillation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 06:40 PM

  Few-step trajectory distillation aims to learn accurate global updates from dense local supervision along a teacher's sampling trajectory.

  Existing approaches construct within-step dynamics using specific trajectory parameterizations or discrete interval predictions, tying the learned update to a prescribed temporal structure.

Extractive summary: sentences quoted from the sources.