A2Z-10M+ | CVPR 2026

A2Z-10M+: The largest multi-modal annotation resource for BRep learning, unifying 10 million annotations across 1 million ABC CAD models. By tightly integrating high-resolution scan-like meshes, 3D hand-drawn sketches, explicit geometric–topological BRep descriptors (co-edges, corners, surfaces), and rich mechanical-language captions and tags, A2Z bridges raw geometry with semantic understanding. The scale, diversity, and structural depth of A2Z unlock an unprecedented foundation for next-generation representation learning, cross-modal reasoning, and intelligent X-to-CAD systems.

Contribution

We propose A2Z-10M+, The largest dataset with multimodal annotations to advance X-to-BRep representation learning and precision engineering for numerous applications like Product Design, Retrieval, Language-guided CAD retrival genration within eXtended Reality (XR) environment, and many others. Our main contributions are:

Highly realistic 3D Scan pairs of 1 million low Poly CAD models under ABC dataset.
Multi-Level CAD Sketches obtained by simulating different skill levels of drawing artists. Realistic deformation, tapering, opennings, hand-drawing wobble, curve openning, and stroke patterns are applied on the Normalized arc-length parameterization of BRep co-edge curves.
BRep Annotations on Scans and Sketches are trasferred with precision using multi-layered distance transforms between noisy scan and CAD (.STEP) .
Textual Caption and Tags generated using BRep metadata injection to a jury system of multiple Large Vision Language Models (Qwen3 and InternVL3). These 1 million textual captions and tags allows retrieval through system-level information, functional intents, forms and usability as input query.
Foundation Model for BRep Boundary and Corner Detection is trained on two on 300K data samples that out performs state-of-the-art methods by large margin. The model is tested on numerous first-had industry grade scan data for precision engineering.

Data and Annotation Visualizer

Scan Annotated by BRep Features: CAD trimesh vertices colored by BRep face membership. Toggle between face partition, boundary curves (per edge ID), wire loops, and corner/junction overlays. Drag to rotate • scroll to zoom • right-drag to pan.

Select a model to begin…

◆ Face

IDs Types Off

∿ Boundary

Edges Vertices Corners

💬 Language

Caption + Tags

↔

Base

Annotated

3D Sketches: Multi-level 3D hand-drawn strokes simulating five skill levels — from beginner wobble to expert precision. Each level applies progressive deformation, tapering, curve openings, and stroke patterns on arc-length parameterized BRep co-edges. Drag to rotate • scroll to zoom.

Select a model to begin…

2D Sketches: Projected 2D representations of 3D BRep geometry with full annotation support for cross-modal reasoning, retrieval, and AI-assisted reverse engineering tasks.

Interactive Visualizer — Coming Soon

Tags and Caption: Rich mechanical-language captions and semantic tags bridging raw geometry with natural language understanding, enabling cross-modal retrieval, language-guided CAD generation, and XR-based design workflows. Turntable with Caption

Results

Quantitative Results: Benchmark evaluation metrics comparing A2Z-10M+ trained models against state-of-the-art methods on BRep boundary detection, corner detection, and cross-modal retrieval tasks.

Tags & Captions

Detailed Analysis

Acknowledgement

This project was primarily funded by BITS Pilani Hyderabad’s NFSG Grant (Reference N4/24/1033). We are thankful to Vinci4D.ai for providing some CAD model samples of electronic enclosures.

Citation

If you find our work useful, please cite:

@inproceedings{AliA2ZCVPR26,
    title     = {A2Z-10M+: Geometric Deep Learning with A-to-Z BRep Annotations for AI-Assisted CAD Modeling and Reverse Engineering},
    author    = {Jena, Pritham Kumar and Baburaj, Bhavika and Anand, Tushar and Dutta, Vedant and Ulavala, Vineeth and Ali, Sk Aziz},
    booktitle = {AxXiv},
    year      = {2026}
}

A2Z-10M+: Geometric Deep Learning with A-to-Z BRep Annotations for AI-Assisted CAD Modeling and Reverse Engineering

Pritham Kumar Jena^1,2 • Bhavika Baburaj^1,2 • Tushar Anand² • Vedant Dutta²
Vineeth Ulavala² • Sk Aziz Ali^1,2

IEEE CVPR 2026

Teaser

Contribution

Data and Annotation Visualizer

Results

Acknowledgement

Citation

Get in Touch