Introducing SGS-1: The First Generative Model for Structured CAD Spectral Labs unveils SGS-1, a foundation model designed to generate fully manufacturable and parametric 3D geometry. SGS-1 produces CAD Boundary Representation (B-Rep) parts in STEP format from images or 3D meshes. Unlike existing generative models, SGS-1 outputs accurate, editable CAD files compatible with traditional CAD software. --- Key Features Inputs: Image or "dumb" 3D mesh Outputs: Parametric B-Rep CAD parts (STEP files) Compatibility: Easily editable in standard CAD software like Fusion360 Supports generation of complex and diverse CAD shapes Real-world engineering applications: designing parts within assembly contexts --- Research Preview A research preview of SGS-1 is available online. Example shown: SGS-1 used in Fusion360 CAD software to create brackets for a roller assembly. Demonstrated ability to generate parametric brackets based on partial context and text descriptions. --- Comparison to Prior Models SGS-1 was benchmarked against: GPT-5 Thinking: A large multimodal reasoning model from OpenAI generating CadQuery parametric geometry code. HoLa: A latent diffusion model trained to generate B-Rep geometry from images. Benchmark Details Dataset: 75 images of medium to high complexity parametric geometry (varied sources: CAD renders, engineering sketches, AI-generated images) Evaluation: Success ratio based on valid watertight solids matching input images using distance metrics Experiment: 10 runs per model, scores recorded for all runs and best output Results SGS-1 outperformed GPT-5 and HoLa, achieving success in almost all cases except the most complex. GPT-5 showed poor spatial understanding; its geometry outputs were often unusable or too simple. SGS-1 accurately captured geometry and fitted parts into assemblies with parametric editability. Visual comparisons illustrate SGS-1's precise geometry versus GPT-5's flawed results. --- Generating Parametric Geometry in Assembly Context SGS-1 can generate new parts within an existing assembly. Input: partial CAD assembly + text description or image of part Output: downloadable parametric B-Rep (STEP) file that fits the assembly Users can import and adjust dimensions easily to ensure fit. SGS-1 can produce diverse designs for parts like brackets. --- Converting Sketches and Engineering Drawings to B-Rep SGS-1 transforms simple freehand sketches and detailed engineering drawings into parametric 3D CAD geometry. Enables powerful design workflows directly from hand sketches or drawings. Suitable for structured engineering drawings. --- Automating Reverse Engineering and STL to STEP Conversion Converts scans and mesh files (like STL) to parametric STEP CAD files without manual intervention. Automates reverse engineering of parts and shapes. --- Limitations Primarily designed for structured, engineering-focused geometry. Struggles with: Creative assets or organic shapes with complex curves Very thin structural elements Generating full assemblies in a single step Future model generations aim to address these issues. --- Next Steps & Vision SGS-1 opens major advancements in foundation models for 3D engineering tasks. Planned future enhancements: Native multimodal support Larger spatial context and complex physical system modeling Advanced physical reasoning with long-range planning Reinforcement learning guided by physical simulation for improved reasoning Collaboration and deployment inquiries welcomed via a contact form on Spectral Labs' website. The team is hiring AI researchers and engineers with backgrounds from institutions such as Autodesk Research, Samsung Research, CMU, and Meta. --- Additional Resources Try SGS-1 demo: Hugging Face Space Contact and research