CADFS: A Big CAD Program Dataset and Framework for Computer-Aided Design with Large Language Models

Vladislav Pyatov; Gleb Bobrovskikh; Saveliy Galochkin; Nikita Boldyrev; Oleg Voynov; Alexander Filippov; Gonzalo Ferrer; Peter Wonka; Evgeny Burnaev

doi:10.48550/arXiv.2605.01925

CADFS: A Big CAD Program Dataset and Framework for Computer-Aided Design with Large Language Models

Vladislav Pyatov^1,2
Gleb Bobrovskikh¹
Saveliy Galochkin^3,1
Nikita Boldyrev¹
Oleg Voynov^1,3
Alexander Filippov²
Gonzalo Ferrer¹
Peter Wonka⁴
Evgeny Burnaev^1,3

¹Applied AI Institute
²AI Foundation and Algorithm Lab
³AXXX
⁴KAUST

CVPR 2026

Abstract

We introduce CADFS, a data-centric framework that enables large vision-language models to generate complex CAD design histories. Existing generative CAD systems are restricted to sketch-extrude operations due to simplified representations and limited datasets. We address this by introducing a FeatureScript-based representation and constructing a dataset of 450k real-world CAD models spanning 15 modeling operations. We obtain the dataset via a new pipeline that reconstructs clean, executable FeatureScript programs and provides multimodal annotations. Fine-tuning a VLM on this representation yields state-of-the-art results in text-conditioned CAD generation and image-based reconstruction, producing more accurate, diverse, and feature-rich designs than prior frameworks. Ablations show that each individual component of our framework, i.e., the FeatureScript representation, the extended operation set, and representation-aligned textual descriptions, significantly improves performance. Our framework substantially broadens the complexity and realism achievable in generative CAD.

Method

1. Reconstruct executable CAD programs

We recover compact FeatureScript programs from Onshape models by making sketch parameters explicit, standardizing units, and simplifying queries, identifiers, and operation definitions. Each program is executed and checked against its source model, and only successful reconstructions are retained.

2. Annotate the design history

An LLM annotator describes the construction process from the recovered code. A second LLM reviews the draft against the code and FeatureScript documentation, correcting terminology, operation order, and references to geometric entities.

3. Train for text and image inputs

Text- and image-conditioned CAD training pipeline

We fine-tune Qwen2-VL to generate FeatureScript from textual descriptions or multi-view images with dimensional information. The generated program compiles into CAD geometry while preserving an editable sequence of modeling operations.

Text-conditioned generation

We evaluate text-conditioned generation on the CADFS test set across a broad range of modeling operations, and on the DeepCAD test set for designs built using sketch-and-extrude operations.

Bevel selected edges.

Reference

Download STEP

CADFS

Download STEP

Drag any model to rotate both views together. Use + / − to zoom, or Reset view to return to the starting angle.

CADFS input text

Step 1 - Sketch Create a new sketch on the default Top plane. Sketch the rectangular outline consisting of five line segments: (0 mm, -50 mm) to (10 mm, -50 mm), up to (10 mm, -10 mm), across to (50 mm, -10 mm), up to (50 mm, 10 mm), and back to (0 mm, 10 mm). Mirror this outline to the left side producing matching segments from (0 mm, -50 mm) to (-10 mm, -50 mm), up to (-10 mm, -10 mm), across to (-50 mm, -10 mm), up to (-50 mm, 10 mm), and back to (0 mm, 10 mm). Add two vertical lines at x = -30 mm and x = 30 mm spanning y = -10 mm to y = 10 mm, and place circles of radius 2.25 mm at (0,0), (20,0), (40,0), (0,-20), (0,-40), (-40,0) and (-20,0). Solve the sketch. Step 2 - Extrude NEW Select the imprint face on the positive side of the bottom edge (the face adjacent to edge from (0, -50) to (10, -50)). Extrude this face a blind depth of 4 mm to create a solid plate following the outer contour of the sketch. Step 3 - Chamfer EQUAL_OFFSETS Apply a 2 mm equal-offset chamfer with tangent propagation to all cap edges of the extruded plate: the outer rectangular edges on both sides and the two central vertical edges at x = -30 mm and x = 30 mm. Step 4 - Chamfer EQUAL_OFFSETS Apply a 1.45 mm equal-offset chamfer with tangent propagation to the cap edges of all seven circles created in the sketch (circles at (0,0), (20,0), (40,0), (-40,0), (-20,0), (0,-20) and (0,-40)), rounding the edges of the resulting holes.

1 / 20

Drag any model to rotate all four views together. Use + / − to zoom, or Reset view to return to the starting angle.

CADFS input text

Step 1 - Sketch Create a sketch on the default Top plane. Draw a circle centered at the origin with a radius of 25.4 mm. Then draw a second circle with a radius of 11.43 mm whose centre is at (0 mm, -11.43 mm). Solve the sketch. Step 2 - Extrude NEW Select the interior region of the 25.4 mm circle and extrude it upward by 5.08 mm, creating a solid cylindrical disc. Step 3 - Extrude ADD Select the interior region of the 11.43 mm circle and add an extrusion upward by 2.54 mm, extending the previous cylinder. Step 4 - Sketch On the top face of the body created in Step 2, start a new sketch. Place a construction circle of radius 7.62 mm centred at (0 mm, 11.5 mm). Then draw a closed six-segment outline: a horizontal line from (3.67 mm, 3.87 mm) to (-3.67 mm, 3.87 mm); a line to (-8.25 mm, 9.61 mm); a line to (-6.61 mm, 16.77 mm); a line to (0 mm, 19.95 mm); a line to (6.61 mm, 16.77 mm); a line to (8.25 mm, 9.61 mm); and finally back to the starting point (3.67 mm, 3.87 mm). Add a point at (0 mm, 3.87 mm). Solve the sketch. Step 5 - Extrude ADD Select the region on the opposite side of the horizontal segment (the area below the line at y = 3.87 mm) from the sketch in Step 4. Extrude this region upward by 25.4 mm with a 3° inward draft (draft pulled toward the sketch plane). Step 6 - Extrude ADD Using the cap face generated by the extrusion in Step 5, extrude the same region outward by 5.08 mm with a 30° outward draft. Step 7 - Extrude ADD Again on that cap face, extrude the same region outward by 25.4 mm with a 45° inward draft (draft pulled toward the sketch plane). Step 8 - Sketch Create a sketch on the default Front plane. Draw the following geometry: - A horizontal segment from (-3.67 mm, 5.08 mm) to (0 mm, 5.08 mm). - A sloping segment from (-3.67 mm, 5.08 mm) to (-3.54 mm, 10.16 mm). - A horizontal segment from (-3.54 mm, 10.16 mm) to (0 mm, 10.16 mm). - A vertical segment from (0 mm, 10.16 mm) down to (0 mm, 5.08 mm). Mirror these three segments to the right side, producing symmetric lines: - From (3.54 mm, 10.16 mm) to (0 mm, 10.16 mm). - From (3.67 mm, 5.08 mm) to (3.54 mm, 10.16 mm). - From (3.67 mm, 5.08 mm) to (0 mm, 5.08 mm). Place a point at (1.83 mm, 5.08 mm). Around this point draw a centered rectangle: bottom edge from (-1.83 mm, 5.08 mm) to (1.83 mm, 5.08 mm); top edge from (-1.83 mm, 8.9 mm) to (1.83 mm, 8.9 mm); left vertical from (-1.83 mm, 5.08 mm) to (-1.83 mm, 8.9 mm); right vertical from (1.83 mm, 5.08 mm) to (1.83 mm, 8.9 mm). Solve the sketch. Step 9 - Extrude ADD Select two faces from the sketch in Step 8: the face on the right side of the mirrored sloping line and the rectangular face defined by the four edges of the centered rectangle. Extrude both faces opposite the sketch normal by 5.08 mm, then extrude an additional 2.54 mm in the same opposite direction, creating a block with a tapered extension.

Image-conditioned reconstruction

We also evaluate our method on image-conditioned reconstruction of CAD models featuring complex modeling operations from the CADFS test set.

Bevel selected edges.

Reference

Download STEP

CADFS

Download STEP

Cadrille

Download STEP

Cadrille RL

Download STEP

Drag any model to rotate all four views together. Use + / − to zoom, or Reset view to return to the starting angle.

Input views

The source image contains multiple views of the same object.

Acknowledgments

We are grateful to Onshape for providing public access to a vast library of CAD designs.

BibTeX

@InProceedings{pyatov2026cadfs,
    author    = {Pyatov, Vladislav and Bobrovskikh, Gleb and Galochkin, Saveliy and Boldyrev, Nikita and Voynov, Oleg and Filippov, Alexander and Ferrer, Gonzalo and Wonka, Peter and Burnaev, Evgeny},
    title     = {CADFS: A Big CAD Program Dataset and Framework for Computer-Aided Design with Large Language Models},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {10176-10186}
}

CADFS: A Big CAD Program Dataset and Framework for Computer-Aided Design with Large Language Models

Abstract

Method

1. Reconstruct executable CAD programs

2. Annotate the design history

3. Train for text and image inputs

Text-conditioned generation

Reference

CADFS

Reference

CADFS

Text2CAD

Cadrille

Image-conditioned reconstruction

Reference

CADFS

Cadrille

Cadrille RL

Input views

Acknowledgments

BibTeX