CADMAGEDDON: Open-Source Mesh to Editable CAD

Introduction
Converting unorganized polygonal geometry into clean, parametric boundary representations has long been a notorious bottleneck in digital asset creation. Today, researchers unveiled CADMAGEDDON, an open-source reconstruction agent engineered to reverse raw meshes into fully editable, parametric CAD assemblies entirely through local inference.
Unlike SaaS offerings that demand enterprise teams upload unreleased geometry to remote cloud servers, CADMAGEDDON runs client-side on single-GPU workstations and modern laptops. For technical artists and hard-surface modelers sitting on high-density photogrammetry scans, messy sculpts, or proprietary weapon and vehicle concepts, this release changes the operational calculus of asset reverse-engineering.
What Shipped
CADMAGEDDON delivers a complete local inference framework tailored for lightweight hardware environments. The key deliverables include:
- Local Agentic Pipeline: An autonomous agent combining geometric heuristics with compact neural vision-language reasoning to extract sketch profiles, extrusions, and fillets.
- Standardized CAD Output: Emits structured step-by-step constructive solid geometry (CSG) commands exportable to standard STEP and BREP interchange formats, ensuring compatibility with standard DCCs and parametric kernels.
- Low Compute Footprint: Quantized and architected to run within a modest 8 GB VRAM envelope or standard Apple Silicon unified memory without off-device API calls.
- Open Repository: Public model weights, inference scripts, and benchmark evaluation suites published under an open research license.
How it works
Reconstructing parametric CAD from arbitrary triangulated meshes usually suffers from topological ambiguity. CADMAGEDDON bypasses brute-force regression by structuring the challenge as an iterative agentic loop.
The framework first ingests an arbitrary target mesh—whether generated via photogrammetry, neural surface synthesis, or legacy polygon exports. A spatial decomposition module sections the mesh into primitive geometric zones (planar faces, cylindrical cuts, and transitional chamfers).
Next, a localized reasoning agent proposes discrete modeling operations: defining reference sketch planes, tracing closed 2D curves, and applying Boolean operations (extrude, cut, revolve). Rather than guessing the entire construction tree in one pass, CADMAGEDDON queries an embedded geometric engine locally, executing the proposed operations and evaluating the Hausdorff distance and normal deviation against the reference polygon mesh. When discrepancies exceed tight tolerances, the agent refines sketch parameters and constraints until convergence, producing a fully editable history tree.
Run it
Setting up CADMAGEDDON locally requires a basic Python environment with CUDA acceleration or MPS support on Apple Silicon:
git clone https://github.com/cadmageddon/cadmageddon.git
cd cadmageddon
pip install -r requirements.txtYou can feed an untextured .obj or .ply mesh into the reconstruction agent via the CLI:
python reconstruct.py \
--input_mesh ./assets/hard_surface_part.obj \
--output_dir ./exports/ \
--format step \
--device cuda \
--max_iterations 12For embedded Python workflows inside DCC automation tools:
from cadmageddon import CADReconstructor
agent = CADReconstructor(model_precision="fp16", local_weights="./weights/cadmageddon-base")
result = agent.fit_mesh(
mesh_path="engine_bracket_scan.ply",
tolerance=0.002,
export_csg=True
)
result.save_step("engine_bracket_editable.step")
print(f"Reconstructed in {len(result.history_tree)} parametric operations.")Why game devs should care
Game studios frequently wrestle with dense, non-manifold triangulated meshes. Whether you are buying kitbash libraries, digitizing industrial props via laser scanning, or using generative 3D tools that spit out unmanageable poly soup, standardizing those assets for runtime optimization is painful.
CADMAGEDDON lets technical artists ingest arbitrary triangulated objects and output native parametric geometry. Once an asset lives as a STEP or BREP entity, generating LOD zero through LOD four becomes deterministic: you can tessellate directly to exact target triangle budgets with flawless edge bevels and curvature-aligned vertex normals. No more manual retopology for mechanical props, vehicles, or weapon chassis.
Critically, keeping inference local eliminates the NDAs and IP leakage risks associated with pushing work-in-progress game assets through third-party cloud generators.
Caveats
While CADMAGEDDON represents a major technical step forward, engineers must keep several constraints in mind:
- Organic Geometry Limits: This tool is designed strictly for manufactured, prismatic, and mechanical parts. Sculpted character faces, cloth folds, and foliage will fail or generate nonsensical CSG trees.
- Inference Latency: Because the agent relies on an iterative propose-and-verify feedback loop, reconstructing complex components with dozens of intersecting cuts can take several minutes per part on modest laptop GPUs.
- Noisy Point Clouds: Severely broken, non-manifold meshes with missing surfaces or extreme noise require cleaning before ingestion; the segmentation module expects coherent surface approximations.
Conclusion
Parametric reconstruction is shifting out of closed enterprise silos and into the open-source toolkit. CADMAGEDDON proves that high-fidelity, editable reverse engineering does not require cloud clusters or proprietary subscriptions—just disciplined agent architecture running locally on everyday hardware. For game developers building hard-surface asset pipelines, this is a tool to test immediately.
Original source: Hugging Face Papers
—Titus
