CADENA: Stepwise CAD Reverse Engineering

Reference implementation of stepwise inference for CADENA, a model that reconstructs a 3D mesh as a parametric CAD program by emitting one operation at a time and comparing the target with the currently built geometry at every step.

• Paper: arXiv:2608.00799
• Model weights: https://huggingface.co/kulibinai/cadena
• Benchmark: https://huggingface.co/datasets/kulibinai/cadena-bench

What is here

Path	Contents
cadgen/	The DSL runtime — the operations a CADENA program calls when it executes
inference/	Stepwise rollout: vLLM serving, per-operation expansion, prefix selection
utils/	Geometry helpers and the metrics, including GMS (utils/gms.py)
rl/	Sandboxed prefix rendering (cad_pool.py) and the reward/scoring used to rank prefixes
dataset/	Building an inference dataset from a directory of meshes
visualization.py	Rendering the multi-view input images

The rule-based program generator that synthesised the training corpus is not part of this release. cadgen/ contains the execution half of the DSL only.

Install

Requires Python 3.10+, a CUDA GPU for serving, and OpenCASCADE via CadQuery.

conda env create -f environment.yml -n cadena && conda activate cadena

Serving the policy needs an NVIDIA GPU (vLLM). The DSL runtime, the metrics and dataset building are CPU-only. To reproduce the paper's numbers bit-for-bit on Linux/x86-64, use the fully pinned environment.lock.yml instead.

Checkpoints

Both stages live in one repository, as subfolders:

Subfolder	Stage
sft	Supervised, final checkpoint of the second stage
rl	Reinforcement learning against executed geometry — the paper's CADENA-RL

hf download kulibinai/cadena --include 'rl/*' --local-dir ./ckpt

Or load directly:

from transformers import AutoModelForVision2Seq
model = AutoModelForVision2Seq.from_pretrained("kulibinai/cadena", subfolder="rl")

Run

Build a dataset from a directory of target meshes, then roll out:

MESH_ROOT=data/meshes OUT_DIR=data/stepwise_hf python dataset/create_hf_dataset.py

MODEL_PATH=./ckpt/rl DATASET_PATH=data/stepwise_hf/meshes ./inference/run.sh

run.sh starts vLLM, waits for it, rolls out, and shuts the servers down. Its defaults reproduce the main-results setting of the paper: greedy decoding and an operation budget of 20. A step is kept only if it improves IoU against the target, so the returned program is the best prefix and extra steps cannot make it worse. For each part it writes the program, the input image and the built STL at every step. Every parameter is an environment variable — see the header of the script.

For the sampling configuration reported alongside greedy, set MODE=beam_topk with TEMPERATURE=1.0 and NUM_EXPANSIONS=12 (E, the candidates drawn per operation; the best is kept by IoU).

The DSL

A generated program opens with the prelude in utils.pipeline.CODE_PREFIX, which imports the operations it may call:

import cadquery as cq
from cadgen.selectors import PointOnEdgeSelector
from cadgen.extrude import extrude
from cadgen.shell import shell
from cadgen.hole import hole
from cadgen.revolve import revolve
from cadgen.orto_cut import orto_cut
from cadgen.sweep import sweep
from cadgen.gear import gear
from cadgen.spring import spring
from cadgen.sweep_adv import sweep_adv
from cadgen.loft import loft
r = None

Each emitted line applies one operation to r. Executing a prefix yields the solid built so far, which is what the model is conditioned on at the next step.

GMS

utils/gms.py implements the metric used in the paper. A predicted point matches a target point only when it is both close enough and has a consistent normal — ‖p − q‖ ≤ τ and n_p · n_q ≥ cos α — so the score reflects whether the surface itself was reconstructed, not merely whether volume was filled. GMS is the normalised area under the F1 curve as the angular tolerance sweeps to α_max = 25°, at fixed τ = 0.05 with 8192 sampled points.

Notes

• A prediction is invalid if it fails to build or is not watertight. Invalid predictions count toward the invalid rate and are excluded from metric means.

Citation

@article{cadena2026,
  title         = {CADENA: Stepwise CAD Reverse Engineering},
  author        = {Kabisov, Soslan and Savrasov, Gennadiy and Elistratov, Maksim and
                   Rodriguez, Antonio and Ignatiev, Daniil and Gavrilov, Nikita and
                   Uzdenov, Rustam and Boyko, Alexey I. and Pasechnik, Igor and
                   Konushin, Anton and Kuznetsov, Andrey and Zhemchuzhnikov, Dmitrii},
  year          = {2026},
  eprint        = {2608.00799},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2608.00799}
}

License

MIT — see LICENSE.