PR-CAD: Progressive Refinement for Unified Controllable and Faithful Text-to-CAD Generation with Large Language Models
New AI system treats CAD design as a single, iterative conversation, boosting modeling efficiency.
A research team led by Jiyuan An has introduced PR-CAD, a novel framework that merges the traditionally separate tasks of generating and editing Computer-Aided Design (CAD) models using large language models (LLMs). The core innovation is a progressive refinement approach, treating CAD modeling as a unified, iterative conversation rather than disjoint operations. To train this system, the team curated a comprehensive dataset spanning the full CAD lifecycle, which includes multiple CAD representations and both qualitative (e.g., 'make it sleeker') and quantitative (e.g., 'increase radius by 5mm') descriptions. This data systematically defines edit operations to generate highly human-like interactions.
PR-CAD's power comes from its reinforcement learning-enhanced reasoning framework. This single AI agent integrates three critical capabilities: understanding user intent, estimating precise numerical parameters, and localizing exactly which part of a 3D model to edit. This 'all-in-one' architecture allows for seamless transitions from initial generation ('design a simple gear') to complex refinements ('make the teeth 20% thicker and add a chamfer'). Extensive experiments show strong mutual reinforcement between generation and editing tasks. On public benchmarks, PR-CAD achieves state-of-the-art results in both controllability (following user commands accurately) and faithfulness (producing geometrically correct models), significantly improving CAD modeling efficiency and user-friendliness.
- Unifies generation and editing into a single, iterative LLM-powered framework, moving beyond disjoint tools.
- Uses a reinforcement learning-enhanced agent to integrate intent understanding, parameter estimation, and precise edit localization.
- Trained on a novel, high-fidelity CAD interaction dataset and achieves state-of-the-art benchmark performance for controllability and faithfulness.
Why It Matters
This could dramatically lower the barrier to 3D design, allowing engineers and designers to iterate on complex models through natural language conversation.