{"id":63587,"date":"2026-07-16T07:25:38","date_gmt":"2026-07-16T11:25:38","guid":{"rendered":"https:\/\/overcentral.com\/en\/?p=63587"},"modified":"2026-07-16T07:25:38","modified_gmt":"2026-07-16T11:25:38","slug":"mit-gift-2d-to-3d-cad","status":"publish","type":"post","link":"https:\/\/overcentral.com\/en\/mit-gift-2d-to-3d-cad\/","title":{"rendered":"MIT Researchers Develop GIFT System to Turn 2D Designs into Accurate 3D CAD Models"},"content":{"rendered":"

Engineers<\/a> working on everything from aircraft components to automotive parts routinely rely on computer-aided design (CAD) software to create 3D models that can be tested in virtual crash simulations or durability assessments. The challenge has been that vision-language models (VLMs<\/a>), which can generate new designs from 2D images and text prompts, often produce CAD code that is too imprecise for real engineering work. Researchers at MIT, in collaboration with colleagues from Red Hat and IBM, have developed a system called GIFT (Geometric Inference Feedback Tuning) that addresses this gap by teaching VLMs to convert 2D designs into accurate, functional CAD programs with dramatically less computation than competing methods.<\/p>\n

GIFT, presented at the International Conference on Machine Learning, introduces a data augmentation framework that is both model-aware and task-aware. Rather than relying on randomly tweaked training data\u2014the conventional approach to data augmentation\u2014the system actively tests a VLM on CAD generation problems, identifies where it fails, and generates new training examples that target those specific weaknesses. The result is a self-improving loop that does not require human intervention to correct mistakes.<\/p>\n

How GIFT Converts 2D Images into Executable CAD Code<\/h2>\n

The core problem the researchers identified is a lack of diverse, high-quality datasets for training VLMs on CAD generation. Existing models can produce code that is almost<\/em> correct, but generating perfectly executable code that produces a geometrically accurate 3D model remains difficult. GIFT tackles this by asking the model to solve the same CAD generation problem multiple times in parallel, checking each guess for correctness. When the model generates a mix of successes and near-misses\u2014rather than all correct or all wrong answers\u2014those in-between cases become valuable training material.<\/p>\n

GIFT adjusts the near-miss solutions to make them fully correct, then saves both the corrected versions and the successful solutions into a new dataset. This dataset teaches the model how to overcome the specific problems it struggles with. “If we sample the model 10 times and it generates 10 correct answers to the same problem, then there is not much for it to learn,” says lead author Giorgio Giannone, a research affiliate in the DeCoDE Lab at MIT and a principal research scientist at Red Hat. “We care about the in-between cases, where the model might only solve the problem 50 percent of the time.”<\/p>\n

Inference-Time Scaling Reduces Computational Cost<\/h2>\n

A key innovation in GIFT is its use of inference-time scaling, which allows a pre-trained VLM to generate better outputs without the expense of full retraining. This approach lets users control how much computation they invest in the tuning process, making it adaptable to different time and budget constraints. The results are striking: GIFT outperformed several competing techniques, producing CAD programs that were more accurate and better aligned with the shapes of ground-truth models, while using only about 20 percent as much computation.<\/p>\n

For engineering teams, this efficiency matters because it brings the prospect of AI-assisted design closer to practical, everyday use. “Nearly every physical product around us, from airplanes to appliances, begins its life as a CAD model,” says co-senior author Faez Ahmed, associate professor of mechanical engineering at MIT and leader of the DeCoDE Lab. “Industry teams are eager for AI that can help speed up the creation of these designs, but today’s models often produce simple shapes inadequate for practice. What excites me about this work is that it gives many image-to-CAD-code models a way to improve themselves, learning from their own errors rather than waiting for more human-made data.”<\/p>\n

What Is GIFT and How Does It Differ from Standard Data Augmentation?<\/h2>\n

GIFT stands for Geometric Inference Feedback Tuning. Unlike standard data augmentation, which creates new training samples by randomly altering properties like color, size, or shape, GIFT generates data that is specifically designed to improve a model’s performance on a target task. It develops an understanding of the model’s strengths and weaknesses by testing it, then uses that knowledge to produce augmentations that address the model’s actual failure modes. Because the system is fully automatic, it can scale without requiring engineers to manually correct errors or curate training data.<\/p>\n

Why This Matters for Rapid Prototyping and Design Exploration<\/h2>\n

The practical implications for engineering workflows are significant. Currently, converting a 2D concept into a functional 3D CAD model that can be tested in simulation is a manual, time-intensive process. GIFT automates a large part of that pipeline, and its ability to learn from its own<\/a> mistakes means it can improve over time without additional human-labeled data. The researchers also note that the system could help engineers identify design choices they might otherwise overlook, expanding the range of solutions explored during the prototyping phase.<\/p>\n

The team is already planning next steps. They aim to extend GIFT so that it can teach models to generate CAD programs that optimize for performance and manufacturability, not just geometric accuracy. They also plan to apply the framework to larger models and more diverse CAD generation tasks.<\/p>\n

Who Should Try This Now<\/h2>\n

GIFT is a research system, not a commercial product, but the paper and methodology are publicly available. Engineering teams and AI researchers working on CAD generation, design automation, or vision-language models can study the framework and adapt it to their own models. For teams building AI-assisted design tools, the key takeaway is that inference-time scaling combined with model-aware data augmentation offers a path to significantly better CAD code generation without the prohibitive cost of retraining large models. Those interested can review the full paper<\/a> on arXiv and consider how a similar self-improvement loop might apply to their own design-to-CAD pipelines.<\/p>\n","protected":false},"excerpt":{"rendered":"

Engineers working on everything from aircraft components to automotive parts routinely rely on computer-aided design (CAD) software to create 3D models that can be tested in virtual crash simulations or durability assessments. The challenge has been that vision-language models (VLMs), which can generate new designs from 2D images and text prompts, often produce CAD code […]<\/p>\n","protected":false},"author":7,"featured_media":84086,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"fifu_image_url":"https:\/\/cards.overcentral.com\/cards\/en\/63587.png","fifu_image_alt":"MIT Researchers Develop GIFT System to Turn 2D Designs into Accurate 3D","footnotes":""},"categories":[349],"tags":[],"class_list":["post-63587","post","type-post","status-publish","format-standard","has-post-thumbnail","category-articles"],"fifu_image_url":"https:\/\/cards.overcentral.com\/cards\/en\/63587.png","fifu_image_alt":"MIT Researchers Develop GIFT System to Turn 2D Designs into Accurate 3D","_links":{"self":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/63587","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/comments?post=63587"}],"version-history":[{"count":0,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/63587\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media\/84086"}],"wp:attachment":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media?parent=63587"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/categories?post=63587"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/tags?post=63587"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}