Self-Consistency for LLM-Based Motion Trajectory Generation and Verification

Jiaju Ma, R. Kenny Jones, Jiajun Wu, Maneesh Agrawala

Keywords: verification, llm, motion graphics, animation, self-consistency vlm

TL;DR: Extending the self-consistency mechanism of LLMs to visual domain (motion trajectory)

Abstract: We study visual concept discovery for motion trajectories, where prompts often describe mid-level shape concepts such as spirals, figure-8s, and parabolas. We represent these concepts as shape families: a prototype trajectory paired with a geometric transformation group that captures allowable variation. To recover a shape family we propose an adaption of self-consistency to the visual domain by clustering samples under warp-invariant transformation groups. Our approach yields an unsupervised, training-free method that supports both improved trajectory generation and concept-based verification. On a benchmark of motion trajectory prompts, our approach improves generation accuracy by 4–6% and verification precision by 11% over strong baselines.

Submission Number: 36