Reproduce “Iterative Motion Editing with Natural Language” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Purvi Goel, Kuan-Chieh Wang, C. Karen Liu, Kayvon Fatahalian
Researchers at Stanford University developed a system for iteratively editing existing character animations using natural language instructions. The approach translates user commands into kinematic motion editing operators (MEOs) via a large language model and employs a specialized diffusion model to integrate edits while preserving motion realism. A user study demonstrated superior fidelity and structural similarity compared to text-to-motion baselines.
Key insights
- Decomposing natural language motion editing into LLM-driven program synthesis using Motion Editing Operators (MEOs) and diffusion-model-based infilling enables precise, iterative control.
- A constrained set of Kinematic Motion Editing Operators (MEOs) provides a predictable intermediate representation, bridging high-level language and low-level motion manipulation.
- Framing motion editing as a "motion-infilling" problem, guided by a specialized diffusion model and seeded with splines, ensures edited motions remain globally harmonized and physically plausible.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.