Summary + Context

During my January 2026 UROP at the MIT DeCoDE Lab, I developed a pipeline to transform 2D images into 3D CAD models to advance research in AI for CAD. I curated a log-uniform dataset of 3,000 models categorized by complexity and built an automated rendering workflow to generate multi-view inputs for testing. To benchmark performance, I evaluated state-of-the-art generative models like TRELLIS and Hunyuan 3D using Chamfer Distance and Intersection over Union metrics. This work established a modular pipeline and dataset which can be used to train and test more advanced AI image to CAD models.