Reproduce “ViSculpt: Visual-Centric Agentic Geometry Editing” 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
Bo Pang, Jiaqi Pan, Xiaocheng Zhang, Jiacheng Xu, Guoping Wang, Peng-Shuai Wang
Researchers from Peking University developed ViSculpt, a visual-centric agentic framework that allows language-driven, in-place 3D geometry editing of existing meshes within professional software like Blender. This system achieved an average user study score of 7.53, comparable to human-created edits at 7.20, by emulating an artist's iterative workflow with continuous visual feedback.
Key insights
- A visual-centric, feedback-driven paradigm can effectively bridge high-level semantic intent with low-level GUI operations for localized 3D geometry editing.
- Abstracting continuous mouse trajectories into compact primitive actions significantly simplifies control for LLM agents while maintaining expressiveness for complex edits.
- A multi-agent system incorporating planning, action execution, and visual reflection, augmented by external knowledge, can effectively emulate a human artist's iterative workflow in professional 3D software.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.