Computer-Aided Design (CAD) relies on structured and editable geometric representations, yet existing generative methods are constrained by small annotated datasets with explicit design histories or boundary representation (BRep) labels. Meanwhile, millions of unannotated 3D meshes remain untapped, limiting progress in scalable CAD generation. To address this, we propose DreamCAD, a multi-modal generative framework that directly produces editable BReps from point-level supervision, without CAD-specific annotations. DreamCAD represents each BRep as a set of parametric patches (e.g., Bézier surfaces) and uses a differentiable tessellation method to generate meshes. This enables large-scale training on 3D datasets while reconstructing connected and editable surfaces. Furthermore, we introduce CADCap-1M, the largest CAD captioning dataset to date, with 1M+ descriptions generated using GPT-5 for advancing text-to-CAD research. DreamCAD achieves state-of-the-art performance on ABC and Objaverse benchmarks across text, image, and point modalities, improving geometric fidelity and surpassing 75% user preference. Code and dataset will be publicly available.
Figures & tables
Figure 1: Our proposed DreamCAD (left) is a multimodal generative framework that can reconstruct CAD models from text, images, and point clouds using parametric patches. CADCap-1M (right) provides 1M+ GPT-5–generated captions.
Dataset
Samples
CADCap
Dataset
Samples
CADCap
ABC [ 23 ]
757,433
✓
3D-Future [ 13 ]
16,990
✓
Automate [ 19 ]
380,124
✓
ModelNet [ 63 ]
12,308
✓
ShapeNet [ 4 ]
52,458
✗
ABO [ 5 ]
7,944
✗
CADParser [ 71 ]
40,989
✓
Fusion360 [ 59 ]
4,603
✓
HSSD [ 22 ]
30,078
✗
Toys4K [ 53 ]
3,482
✗
Total
1,306,409
Table 1: Training datasets for DreamCAD comprising 1.3M 3D meshes. ✓ denotes if the dataset is included in proposed CADCap-1M dataset.
Figure 2: Bézier surface representation and differentiable tessellation.
Figure 3: DreamCAD Overview: (A). Sparse Transformer VAE from an input mesh, active voxels vi are generated with local features fi , from DINOv2 [ 45 ] embeddings, normal images, and SDF values and encodes it to generate structured latents zi . These are then decoded into parametric (rational bézier) surfaces and optimized using Chamfer loss. (B). Initial C0 -continuous Parametric Surface generation from sparse voxels via flood-fill and quad conversion using grid control points and unit weights. (C). Multi-modal CAD generation from images, or points using a flow-matching framework from coarse voxel grid to parametric surface refinement.
Figure 4: Examples of metadata-augmented captions from CADCap-1M showing object type, part names, and hole counts.
Figure 5: Qualitative comparison on Point2CAD (Top-Right), Image2CAD (Bottom-Left) and Text2CAD (Right) tasks. For each task, the first four examples are from the ABC dataset, while the last two from Objaverse dataset. ✗ indicates invalid models.
Task
Models
ABC
Objaverse
F1 ↑
NC ↑
CD ↓
HD ↓
JSD ↓
MMD ↓
F1 ↑
NC ↑
CD ↓
HD ↓
JSD ↓
MMD ↓
Point2CAD
DeepCAD [ 61 ]
19.31
0.49
51.10
0.37
783.94
29.63
7.05
0.48
320.33
0.41
855.14
34.62
CAD-Recode [ 49 ]
75.99
0.79
3.73
0.13
271.89
2.94
53.24
0.66
7.92
0.19
479.50
6.27
Cadrille [ 24 ]
78.86
0.80
2.98
0.12
236.10
2.51
57.49
0.67
6.28
0.17
445.23
5.24
DreamCAD (Ours)
92.12
0.94
0.93
0.06
96.13
0.84
87.31
0.89
1.25
0.11
189.12
1.86
Img2CAD
GPT ↑
User ↑
CD ↓
HD ↓
JSD ↓
MMD ↓
GPT ↑
User ↑
CD ↓
HD ↓
JSD ↓
MMD ↓
Table 2: Quantitative comparison across the Point2CAD, Img2CAD, and Text2CAD tasks on the ABC and Objaverse datasets. For readability, F1 is scaled by 102 and CD, JSD, and MMD by 103 . For text- and image-to-CAD, GPT and User ratings measure visual alignment.
Table 3: Ablation studies on regularization (left), and voxel resolution choice (right).
Figure 6: Examples showing VAE reconstructions for different regularizers.
Figure 7: Left: Captions from GPT-5 with and without metadata-augmented prompting. Right: Images from pretrained vs. fine-tuned SD-3.5 given the same prompts.
Table 4: Ablation studies on initialization quality and coarse-to-fine generation.
Figure 8: ( Left ) Topology recovery from patch-based outputs to hybrid NURBS CAD representation [ 54 ] . ( Right ) For multimodal inputs, DreamCAD reconstructions and the corresponding recovered topology.
Figure 9: Wordcloud of captions from CADCap-1M
Mean Length
Unigram
Bigram
Trigram
19.63
21042
446,061
2,368,731
Table 5: Summary statistics of the CADCap-1M dataset, including average caption length and diversity measured through unigram, bigram, and trigram counts.
Figure 10: Distribution of patch sizes in training datasets.
VAE
Sparse Structure
Coarse
Fine-Grained
Image
Point
Image
Point
71M
133M
354M
280M
400M
325M
Table 6: Parameter counts for all components of DreamCAD.
Task
Models
DeepCAD
F1 ↑
NC ↑
CD ↓
HD ↓
JSD ↓
MMD ↓
Point
DeepCAD
23.18
0.53
47.63
0.34
729.11
26.80
CAD-Recode
90.95
0.92
1.38
0.06
126.15
1.13
Cadrille
92.12
0.93
0.30
0.05
112.31
1.05
DreamCAD
94.91
0.94
0.12
0.03
98.51
0.90
Img
GPT ↑
User ↑
CD ↓
HD ↓
JSD ↓
MMD ↓
Table 7: Quantitative comparison on Point2CAD, Img2CAD, and Text2CAD tasks over the DeepCAD test set. F1 is scaled by 102 , while CD, JSD, and MMD by 103 . For text - and image-to-CAD , GPT and User ratings measure visual alignment.
Figure 11: Some failure cases showcasing missing geometric details and invalid models.
Table 8: Ablation studies on initialization quality and coarse-to-fine generation.
Figure 12: Examples of captions for complex (top) and simple (bottom) CAD parts.
Figure 13: Examples of captions for different types of fasteners .
Figure 14: Examples of captions for letter -based (top) and engraving -style (bottom) CAD models.
Figure 15: More qualitative results for the point-to-CAD reconstruction task using DreamCAD. Each row shows input point clouds (top) and the corresponding reconstructed CAD models (bottom) generated by DreamCAD across a wide range of shapes, including mechanical components, furniture, utensils, and free-form surfaces. As seen in the examples, DreamCAD successfully recovers clean parametric geometry from input point clouds, preserving fine structural details (e.g., tubular connectors, circular cutouts, brackets ), smooth surfaces (e.g., bowls, vases, seats ), and complex topologies (e.g., multi-part assemblies and chair frames ).
Figure 16: More qualitative results for the image-to-CAD generation task using DreamCAD. Each row shows the ground-truth reference image (top) and the CAD reconstruction produced by DreamCAD (bottom) across a wide range of object categories, including furniture, mechanical parts, consumer products, and free-form designs. As seen in the examples, DreamCAD consistently recovers accurate 3D geometry from a single image, capturing fine structural details (e.g., chair backrests, table legs, nozzle openings ), complex mechanical features (e.g., gear teeth, connectors, threads ), and overall proportions with high fidelity.
Figure 17: More qualitative text-to-CAD results on the test set. Each row shows an input text description (bottom) and the corresponding CAD geometry generated by DreamCAD (top). DreamCAD successfully reconstructs shapes from complex text prompts, ranging from mechanical parts ( gear assemblies, flange plates, couplings ), tools ( pyramidal axe head ), and structural elements ( DIN rail, stepped platforms ), to free-form designs ( seat frames, turbine blades ), and consumer objects ( USB cable ).
Figure 18: Text-to-CAD generation results on GPT-generated prompts. Each example shows a text prompt produced by GPT-5 (bottom) and the corresponding CAD reconstruction (top) by DreamCAD. As shown, DreamCAD generalizes beyond dataset-style prompts and reliably interprets free-form, open-vocabulary instructions which includes consumer electronics ( smartphone , computer mouse ), household items ( frying pan, stool, water cooler bottle ), tools ( screwdriver, crowbar, scissors ), mechanical components ( V-groove pulley ), and articulated objects ( desk lamp arm, quadcopter drone ).
Multi-modal approaches used for 3D CAD generation require substantial computational resources, necessitating efficient training. To address this, we propose GuideCAD, which leverages semantically rich visual-textual representations having only a small number of trainable parameters to generate 3D CAD models. Specifically, GuideCAD uses a mapping network that converts image embeddings into prefix embeddings, enabling a pretrained large language model to integrate visual and textual information. As a result, a transformer-based decoder predicts the construction sequence using the visual-textual embeddings in order to generate the 3D CAD model. For experimental evaluation, we construct a new dataset, referred to as GuideCAD, which consists of text-image pairs. Each pair includes a text prompt that represents a 3D CAD construction sequence and its corresponding 3D CAD image. Our experimental results show that GuideCAD generates comparably high-quality 3D CAD models while using approximately four times fewer parameters and achieving twice the training efficiency compared to fine-tuning approaches. We have released the source code and dataset for our method at: https://github.com/mskimS2/GuideCAD
Minseong Kim, Jinyeong Park, Sungho Park +1
Convergence Research Center for Insect Vectors Incheon, South Korea · Department of Computer Science and Engineering Incheon National University Incheon, 22012, South Korea · Department of Computer Science and Engineering Center for Brain-Machine Interface Incheon National University Incheon, South Korea
Computer-Aided Design (CAD) models are defined by their construction history: a parametric recipe that encodes design intent. However, existing large-scale 3D datasets predominantly consist of boundary representations (B-Reps) or meshes, stripping away this critical procedural information. To address this scarcity, we introduce Zero-to-CAD, a scalable framework for synthesizing executable CAD construction sequences. We frame synthesis as an agentic search problem: by embedding a large language model (LLM) within a feedback-driven CAD environment, our system iteratively generates, executes, and validates code using tools and documentation lookup to promote geometric validity and operation diversity. This agentic approach enables the synthesis of approximately one million executable, readable, editable CAD sequences, covering a rich vocabulary of operations beyond sketch-and-extrude workflows. We also release a curated subset of 100,000 high-quality models selected for geometric diversity. To demonstrate the dataset's utility, we fine-tune a vision-language model on our synthetic data to reconstruct editable CAD programs from multi-view images, outperforming strong baselines, including GPT-5.2, and effectively bootstrapping sequence generation capabilities without real construction-history training data. Zero-to-CAD bridges the gap between geometric scale and parametric interpretability, offering a vital resource for the next generation of CAD AI.
Mohammadmehdi Ataei, Farzaneh Askari, Kamal Rahimi Malekshan +1
The field of Computer-Aided Design (CAD) generation has made significant progress in recent years. Existing methods typically fall into two separate categories: parametric CAD modeling and direct boundary representation (B-Rep) synthesis. In modern feature-based CAD systems, parametric modeling and B-Rep are inherently intertwined, as advanced parametric operations (e.g., fillet and chamfer) require explicit selection of B-Rep geometric primitives, and the B-Rep itself is derived from parametric operations. Consequently, this paradigm gap remains a critical factor limiting AI-driven CAD modeling for complex industrial product design. This paper presents FutureCAD, a novel text-to-CAD framework that leverages large language models (LLMs) and a B-Rep grounding transformer (BRepGround) for high-fidelity CAD generation. Our method generates executable CadQuery scripts, and introduces a text-based query mechanism that enables the LLM to specify geometric selections via natural language, which BRepGround then grounds to the target primitives. To train our framework, we construct a new dataset comprising real-world CAD models. For the LLM, we apply supervised fine-tuning (SFT) to establish fundamental CAD generation capabilities, followed by reinforcement learning (RL) to improve generalization. Experiments show that FutureCAD achieves state-of-the-art CAD generation performance. Code and dataset are available at: https://github.com/JohanStackk/FutureCAD