MIRAGE‑CAD generates executable CAD

MIRAGE-CAD is a new research system aimed at a specific CAD problem: not just generating a 3D shape, but recovering an executable parametric program that can rebuild that shape. The work appears in an arXiv paper titled “MIRAGE-CAD: Construction-Mediated Multimodal Generation of Executable CAD Programs,” posted in late August 2026. (arxiv.org) The authors say the system accepts four input types — natural-language descriptions, rendered images, point clouds and STEP/B-Rep geometry — and maps them into a shared construction representation before generating code. The distinction matters because conventional generative 3D systems often stop at geometry. (github.com) MIRAGE-CAD instead targets a program that can be run again, edited and exported through a CAD kernel. The paper says the resulting Python CAD code is executed by an OpenCASCADE kernel to build the solid and export it as a STEP file. ### Why are researchers emphasizing “executable CAD programs” instead of just shapes? The arXiv paper says recovering an executable parametric CAD program from an observed object is “fundamentally ambiguous” because the same final geometry can come from different construction procedures. (arxiv.org) That is a CAD-specific problem: two parts can look identical as solids while encoding different design intent, feature order or editable parameters. MIRAGE-CAD addresses that by routing every input modality through what the authors call a shared construction intermediate representation and an explicit construction-plan interface. (arxiv.org) In practice, that means the system is trying to infer not only what the object is, but how it was built in CAD terms. ### What kinds of inputs can it work from? The paper lists four sources: text prompts, rendered images, point clouds and STEP/B-Rep models. That gives the project a broader scope than systems limited to text-to-3D or image-to-mesh generation. (arxiv.org) The GitHub repository uses similar language, describing MIRAGE-CAD as a multimodal system that generates executable parametric CAD programs by routing each input modality through a shared construction representation and an explicit construction plan. The repository also says the tools folder contains executable code for “editable CAD program generation.” (arxiv.org) ### What role does OpenCASCADE play here? OpenCASCADE is the geometric kernel behind execution. The MIRAGE-CAD paper says the generated Python code is run by an OpenCASCADE kernel, which then constructs the solid and exports it as STEP. (arxiv.org) Open Cascade describes OCCT3D Technology as an open-source 3D modeling kernel and software development kit for CAD data, including surface and solid modeling, visualization and data exchange. That makes it the infrastructure layer that turns MIRAGE-CAD’s generated program into an actual CAD object rather than a static prediction. (github.com) ### How does this fit into the broader CAD-AI race? Another 2026 paper, “Zero-to-CAD,” also framed CAD generation around interpretable, executable construction sequences rather than end-state geometry alone. That suggests MIRAGE-CAD is part of a broader research push toward systems that recover construction history and design intent, not just surfaces. (arxiv.org) The MIRAGE-CAD paper itself stays focused on its own pipeline, but its framing is clear: the target output is editable, executable CAD logic. That is closer to how engineers work in parametric CAD systems, where sketches, features and constraints are often as important as the finished part. (opencascade.com) ### Where can readers check the work themselves? The paper is available on arXiv under identifier 2608.28669, and the GitHub repository is published under the Cad-Kernel organization. The repository says it includes code, pipeline scripts, analysis scripts and run reports for MIRAGE-CAD, along with tooling for experiment scales including a smoke test and a larger benchmark setup. (arxiv.org) As of this week, arXiv lists the submission as a recent paper in computer vision and related categories, and the public repository remains accessible for readers who want to inspect the implementation details. (arxiv.org 1) (arxiv.org 2)