MIT’s GIFT sharpens 2D-to-3D CAD with less compute

Researchers from MIT, Red Hat, and IBM have introduced a system called GIFT that improves how vision-language models turn 2D designs into executable CAD programs for 3D models. The team says the method produces more accurate and functional results than competing approaches while using only about 20 percent as much computation.

The work targets a persistent problem in AI-assisted engineering. Models can generate design concepts for parts such as airplane or automobile components, but converting those ideas into reliable CAD code for simulation and prototyping is much harder. According to the researchers, one major bottleneck is the lack of diverse, high-quality CAD training data.

GIFT, short for Geometric Inference Feedback Tuning, tackles that by generating new training data from a model’s own attempts. Rather than relying on standard data augmentation tricks, it tests a pre-trained model on CAD generation tasks, identifies where it succeeds and where it nearly succeeds, and uses those outcomes to build a better dataset.

“We want engineers to be able to point our framework at an underperforming CAD model, set a compute budget, and let the system take over—turning the model’s own mistakes into better training data.”

How GIFT improves image-to-CAD models

The framework asks a model to generate CAD query code for the same problem multiple times in parallel. It then checks which outputs are correct and which are close. Those near-misses are adjusted into successful solutions and stored alongside correct answers, creating training examples focused on the model’s actual weak points.

Recommended reading

FBI eyes Nvidia and Google TPU AI systems

That makes the system both model-aware and task-aware. It also uses inference-time scaling, which lets a static model improve outputs without the cost of retraining the entire system. Users can set a compute budget based on time or cost constraints.

Giannone said the hardest part is getting code that is not just almost right, but fully executable in CAD software. The team found the most useful cases were the middle ground, where a model solves a problem only 50 percent of the time, rather than always succeeding or always failing.

ICML presentation and next steps

The paper was recently presented at the International Conference on Machine Learning. The authors are Giorgio Giannone, Anna Claire Doris, Amin Heyrani Nobari, Kai Xu, and co-senior authors Akash Srivastava and Faez Ahmed.

Ahmed said current systems often produce shapes too simple for real engineering use, and that this work pushes AI design tools closer to everyday practice by letting models improve from their own errors instead of waiting for more human-made data.

For now, the focus is on geometry. Next, the researchers want to expand GIFT so models can generate CAD programs that also improve performance and manufacturability, and apply the approach to larger models and a wider range of CAD tasks.

The publication is “GIFT: Bootstrapping Image-to-CAD Program Synthesis via Geometric Feedback” on arXiv (2026), with DOI 10.48550/arxiv.2603.27448.