| --- |
| license: mit |
| library_name: transformers |
| pipeline_tag: image-to-text |
| base_model: Qwen/Qwen2-VL-2B-Instruct |
| tags: |
| - arxiv:2608.00799 |
| - cad |
| - reverse-engineering |
| - cadquery |
| - vision-language |
| --- |
| |
| # CADENA |
|
|
| CADENA reconstructs a 3D mesh as a parametric CAD program. Instead of emitting |
| the whole program in one pass, it grows the operation sequence one step at a |
| time, executing the partial program and comparing the target with the currently |
| built geometry before choosing the next operation. |
|
|
| - Paper: [arXiv:2608.00799](https://arxiv.org/abs/2608.00799) |
| - Code: <https://github.com/zhemdi/cadena> |
| - Benchmark: <https://huggingface.co/datasets/kulibinai/cadena-bench> |
|
|
| ## Checkpoints |
|
|
| This repository holds both training stages as subfolders. |
|
|
| | Subfolder | Stage | |
| | --- | --- | |
| | `sft` | Supervised, final checkpoint of the second stage | |
| | `rl` | Reinforcement learning against executed geometry — the paper's CADENA-RL | |
|
|
| ```python |
| from transformers import AutoModelForVision2Seq, AutoProcessor |
| |
| model = AutoModelForVision2Seq.from_pretrained("kulibinai/cadena", subfolder="rl") |
| processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct") |
| ``` |
|
|
| Or fetch one stage only: |
|
|
| ```bash |
| hf download kulibinai/cadena --include 'rl/*' --local-dir ./ckpt |
| ``` |
|
|
| Both are Qwen2-VL-2B policies, 4.4 GB each in bfloat16. |
|
|
| ## Input |
|
|
| The model observes eight renders of the target tiled into a single image: six |
| axis-aligned views and two isometric ones. The target occupies the green |
| channel and the geometry built so far the red channel, so the residual — what |
| is still missing — is what the policy actually reads. |
|
|
| ## Results |
|
|
| CADENA-RL outperforms prior methods on DeepCAD, Fusion 360, MCB and |
| CADENA-Bench. The margin grows with geometric complexity: it is modest on |
| sketch–extrude corpora and substantially larger on real mechanical parts. |
| Reinforcement learning against executed geometry improves reconstruction |
| accuracy while lowering the invalid rate, since an operation that fails to |
| build receives no reward. |
|
|
| See the paper for the full tables. |
|
|
| ## Usage |
|
|
| Inference is stepwise and needs the DSL runtime, so use the repository rather |
| than a bare `generate` call: |
|
|
| ```bash |
| git clone https://github.com/zhemdi/cadena && cd cadena |
| pip install -r requirements.txt |
| MODEL_PATH=./ckpt/rl DATASET_PATH=data/stepwise_hf/wrapped/meshes ./inference/run.sh |
| ``` |
|
|
| Defaults reproduce the paper's main-results setting: greedy decoding, an |
| operation budget of 20, `E = 12` candidates per operation, and the returned |
| program is the prefix with the highest IoU against the target. |
|
|
| ## Limitations |
|
|
| - Programs are CadQuery, so CADENA inherits its limitations; in corner cases a |
| construction tree valid in CadQuery does not transfer to industrial CAD. |
| - Accuracy is uneven across part families — revolved and patterned geometry |
| (gears, bearings, springs, fasteners) remains the hardest. |
|
|
| ## License |
|
|
| MIT. |
|
|
