CADFit — Loop classifier (model_mlp_med.pt)

This repo hosts the trained loop-filtering model used by CADFit, our framework for precise mesh-to-CAD program generation via hybrid optimization (ICML 2026).

• 📄 Project page: https://ghadinehme.github.io/cadfit.github.io/
• 💻 Code: https://github.com/ghadinehme/CADFit
• 📊 The model files in this repo:
  • model_mlp_med.pt — 3-layer MLP classifier weights + checkpoint metadata (≈ 8.7 MB)
  • normalization.npz — input feature mean/std for standardization (≈ 33 KB)

What CADFit does

CADFit recovers complex, editable CAD construction sequences from watertight meshes by incrementally fitting and validating parametric operations using geometric feedback. Instead of predicting a program in one forward pass, CADFit formulates reconstruction as an explicit IoU-driven optimization over executable CAD programs, supporting a rich operator set (extrusions, revolutions, fillets, chamfers) composed with Boolean union and cut.

Pipeline:

1. Slice the mesh at axis-aligned and detected planar slabs to extract 2D contour loops (extract_planar_contours.py, extract_sketches.py).
2. Filter loops — this model scores every candidate loop and discards ones unlikely to be useful sketch primitives.
3. Analyze translations and revolves — for each surviving loop, find the extrusion height ranges and the symmetry axes that align with the underlying CAD program.
4. Greedy reconstruction — at each step add the candidate operation that yields the largest IoU gain against the input mesh, accepting/rejecting via CadQuery kernel evaluation. Continue until residual volume falls below tolerance.
5. Iterative residual refinement — reconstruct positive and negative residuals with the same pipeline and compose via Boolean union/cut.

The result is a clean, editable, executable CAD program (Python via CadQuery).

The classifier

model_mlp_med.pt is a 3-layer MLP that takes:

• A 4 × 1024-d DINOv2 ViT-L/14 embedding of 4 rendered views per loop (isometric mesh, isometric mesh + loop overlay, face-on mesh, face-on mesh + loop overlay), 224 × 224 px each.
• A 17-d hand-crafted geometric descriptor (loop area, perimeter, point count, bbox extents, plane origin, plane normal, axis-alignment, …).

Total input: 4 × 1024 + 17 = 4113-d. Standardized with (mu, sd) from normalization.npz, then passed through:

Linear(4113 → h₁) → GELU → Dropout
Linear(h₁ → h₂)  → GELU → Dropout
Linear(h₂ → 1)   → sigmoid

(h₁, h₂, and dropout p are stored in ckpt["arch"].)

The output sigmoid probability ≥ threshold (default 0.425, also stored in the checkpoint) is the keep decision.

In CADFit the filter slots in between loop extraction and translation/revolve analysis, shrinking the candidate search space and keeping the greedy reconstruction focused on loops that have a chance of corresponding to a real CAD sketch primitive.

Typical impact

• Drops roughly half of the raw extracted loops.
• Reduces translation/revolve analysis runtime ~2×.
• Improves greedy reconstruction quality (slightly higher mean IoU on ABC-Easy / Medium / Hard).

How to use

From CADFit (recommended)

# Install CADFit and the optional filter dependencies
pip install -r requirements.txt
pip install torch pyvista Pillow huggingface-hub

# Headless Linux only — install Xvfb for off-screen PyVista rendering:
sudo apt-get install xvfb

# Run the pipeline with the filter enabled.
# Model weights auto-download from this repo on first call.
python run_pipeline.py mesh_folder/ --output-folder runs/ --filter-sketches

Stand-alone

from huggingface_hub import hf_hub_download
from filter_loops import filter_jsons_inplace

# Filter a loops_split.json in place; weights cached after first call.
filter_jsons_inplace(
    stl_path="mesh.stl",
    json_paths=["loops_split.json"],
    threshold=0.425,
)

Direct download

from huggingface_hub import hf_hub_download

mlp = hf_hub_download("ghadinehme/CADFit", "model_mlp_med.pt")
norm = hf_hub_download("ghadinehme/CADFit", "normalization.npz")

Citation

@article{nehme2026cadfit,
  title   = {CADFit: Precise Mesh-to-CAD Program Generation with Hybrid Optimization},
  author  = {Nehme, Ghadi and Whalen, Eamon and Ahmed, Faez},
  journal = {arXiv preprint arXiv:2605.01171},
  year    = {2026}
}

License

Released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Free for academic and research use; commercial licensing inquiries to ghadi@mit.edu.

Contact

Questions, feedback, or collaboration: ghadi@mit.edu.