Turn Sketches Into CAD Code Instantly With This Open-Source AI Tool

Key takeaways

Useful thresholds

What is CAD-Coder and why use it now?

CAD-Coder is an open-source Vision-Language Model (VLM) that converts hand-drawn sketches or diagrams into editable CadQuery Python code with 92% syntax validity and 88% 3D solid similarity, as validated in its latest benchmarks. It reduces design iteration time from hours to seconds for simple parts by automating parametric geometry generation.

The tool fine-tunes a multimodal transformer on the CADSketch Dataset (100K+ annotated sketches) and Fusion 360 Gallery (50K parametric models) to infer geometric constraints (radii, hole placements) from visual input. Unlike proprietary tools like Autodesk’s "Sketch to 3D" (which outputs locked meshes), CAD-Coder generates human-readable, modifiable code compatible with version control and automated workflows. This suits prototyping, DIY manufacturing, and small-scale production where flexibility outweighs precision.

Under Apache 2.0 license, CAD-Coder permits commercial use without royalties, unlike closed tools (e.g., Shapr3D at $300/year or Fusion 360 at $60/month). Hardware requirements: NVIDIA GPU with ≥8GB VRAM (e.g., RTX 3060) for real-time processing; CPU fallback slows inference 3–5x. Ubuntu 22.04 outperforms Windows 11 by ~20% due to CUDA driver optimizations, with 1.5s average latency for 2D sketches on an RTX 4070.

In 2026, CAD-Coder’s adoption among its user base spans prototyping (78%), furniture design (65%), and DIY manufacturing (55%). Aerospace and automotive sectors avoid it due to sub-0.1mm precision needs. Limitations include poor performance on organic shapes and 3D sketches, though 2D orthographic views (front/top/side) convert to 3D CAD with 85% accuracy. A 2026 update added PCB layout support and a "sketch cleanup" preprocessor, reducing error rates by 15%.

Common pitfalls: treating output as production-ready (30% of users report manufacturing errors), ignoring unit consistency (mm vs. inches), and overloading sketches with >50 entities (crashes parser). The model flags "low-confidence" outputs in 90% of ambiguous cases but may still generate incorrect code. For niche workflows (e.g., jewelry design), fine-tuning requires ≥1,000 labeled sketches and a PyTorch environment; templates are available in the GitHub repo.

Install CAD-Coder from GitHub, verify GPU meets 8GB VRAM minimum, and test with a labeled 2D sketch (e.g., a bracket with "R10" radius). Use Hugging Face’s inference API for low-resource devices (requires internet). Commercial use must comply with Apache 2.0 terms: attribution is mandatory, and reselling the model as a service requires open-sourcing modifications.

Who can use it—license and commercial rules

CAD-Coder is free for all users, including commercial businesses, under the Apache 2.0 license. No royalties, subscription fees, or per-use charges apply—only attribution is required.

Apache 2.0 permits modification, distribution, and commercial use of the software and its outputs. This includes selling products designed with CAD-Coder, integrating it into proprietary workflows, or offering it as a service. Legal constraints are: (1) including the original license and copyright notice in all copies, (2) documenting changes to the source code, and (3) not using the project’s name or contributors’ names to endorse derived works without permission. Unlike Fusion 360 ($60/month) or Shapr3D ($300/year), CAD-Coder has no financial barriers.

Commercial use does not require open-sourcing your product—only modifications to CAD-Coder itself. If reselling the model as a service (e.g., a cloud-based sketch-to-CAD API), changes to the core model must be released under Apache 2.0. This applies only to the model’s code, not to generated CAD files or products. Example: A furniture startup can sell chairs designed with CAD-Coder without publishing proprietary designs, but must open-source changes to CAD-Coder’s inference engine.

Apache 2.0 is globally recognized, but some jurisdictions impose additional requirements. UAE: Selling products may require a DET trade license or eTrader permit for home-based businesses. U.S.: No federal license is needed, but state-level business registrations (e.g., LLC formation) may apply. The tool’s local processing complies with GDPR/CCPA, as no data is transmitted to external servers by default.

Common mistakes: (1) removing attribution, (2) patenting CAD-Coder’s algorithms without permission, or (3) using it in safety-critical applications (e.g., aerospace, medical devices) without independent validation. Apache 2.0 disclaims warranties; users bear liability for defects. A 2026 case study found 30% of users who skipped manual validation encountered production failures (e.g., misaligned holes, incorrect thread pitches).

To comply: Download from the official GitHub repository, retain the LICENSE file, and add a notice like "This product uses CAD-Coder (Apache 2.0) by Annie Doris et al." to documentation or credits. For modifications, document changes in CHANGES.md and release under Apache 2.0. For commercial services, use the Hugging Face inference API (auto-attribution) or self-host with the provided Docker container.

What file formats work for sketches and CAD?

CAD-Coder accepts raster images (PNG, JPG) and vector sketches (SVG) as input, outputting CadQuery Python code. Sketches must be 2D orthographic views (front, top, side); 3D isometric or perspective sketches are not supported.

Raster images require 300–600 DPI for optimal accuracy—72 DPI phone photos reduce dimension detection by 40%. SVG files improve edge detection for sharp corners and precise radii but require clean paths (no overlapping lines or open loops). JPG compression artifacts may increase misinterpretation of thin lines (<0.5mm). PNG is preferred for hand-drawn sketches (lossless compression); SVG is ideal for digital sketches from Inkscape or Adobe Illustrator.

Output is CadQuery Python code, requiring a compatible environment (FreeCAD, Blender with CadQuery plugin) to generate 3D models. STEP/DXF/STL exports need manual execution of the generated code or third-party converters (e.g., FreeCAD’s "Export" function). DWG, IGES, or OBJ files are not directly supported—conversion requires intermediate software (e.g., LibreDWG for DWG). The 2026 v1.2 update added KiCad-compatible Python script support for PCB layouts, but Gerber or ODB++ exports still require external tools.

Unit specification (mm/inches) in the CadQuery script is mandatory to avoid scaling errors. Unsupported inputs include 3D STL files, grayscale sketches with low contrast (accuracy drops 30%), and assumptions of tolerance inclusion (manual validation required). Preprocess sketches with CAD-Coder’s "sketch cleanup" tool (reduces noise by 15%) and annotate critical dimensions (e.g., "R10", "Ø20"). For the Hugging Face API, upload PNG/SVG files under 5MB; larger files may time out. Offline batch-processing is available via the GitHub repository.

What hardware do you need to run it?

CAD-Coder requires an NVIDIA GPU with ≥8GB VRAM for real-time processing or a CPU fallback (3–5x slower). Ubuntu 22.04 is recommended, outperforming Windows 11 by ~20% due to CUDA optimizations.

The 1.2B-parameter transformer model demands parallel processing. An RTX 3060 (8GB VRAM) processes simple 2D sketches in 1.5s; PCB layouts or 3D orthographic conversions need 12–16GB VRAM (RTX 4070/4070 Ti) to avoid bottlenecks. System RAM: ≥16GB (32GB ideal for batch processing). Local execution is default; Hugging Face’s inference API (requires internet) offloads computation for sub-minimum-spec devices.

Exceptions: Hugging Face API enables CPU-only devices (e.g., MacBook Air M1) at 5–10x slower speeds (15–30s latency). AMD GPUs may work via ROCm on Linux but lack official support. Apple Silicon (M1/M2) may work with manual PyTorch setup but lacks official support.

Common mistakes: Underestimating VRAM for >50-entity sketches (may crash 8GB GPUs). Integrated graphics (e.g., Intel UHD) lack CUDA cores and fail to initialize. Power supply: RTX 4070 draws 215W under load—may exceed prebuilt office PC wattage.

Hardware matching table:

Test with "sample_bracket.png" before production. Latency >30s on recommended GPUs indicates driver conflicts or thermal throttling.

How accurate is it and what mistakes happen?

CAD-Coder achieves 92% syntax validity and 88% 3D solid similarity for simple parts, with accuracy declining for complex or ambiguous inputs. Benchmark tests show common errors include misinterpreted dimensions (12% of cases) and incorrect hole placements (8%).

Accuracy derives from training on the CADSketch Dataset (100K+ labeled sketches) and Fusion 360 Gallery (50K parametric models), emphasizing geometric constraints like radii, hole placements, and symmetry. Explicit labels (e.g., "R10", "Ø20") improve accuracy by 40% over freehand drawings. The transformer architecture maps visual patterns to code templates but cannot infer intent—ambiguous lines (e.g., overlapping edges, incomplete circles) cause errors in 12% of cases, often misinterpreting dimensions or omitting features.

What to do next

Now that you’ve seen how CAD-Coder transforms sketches into editable CAD code, here’s a clear roadmap to put it into action—step by step.

Also worth reading: We Found The Best Open Source Software For CAD Model Conversion · CAD Coder Brings Open Source Vision Language Models to Computer Aided Design · Effortlessly Turn DXF Into G Code · Building with Code Open Source Tools Redefine Design Workflows

Sources: arxiv, github, themoonlight, mit, researchgate