Chen8566/UniCAD-Synthetic-Views

UniCAD Synthetic Multi-View CAD Images

This dataset contains the synthetic multi-view image data created for training UniCAD. It provides four 1024 x 1024 RGB views of CAD objects in two visual styles:

• sketch: hand-drawn-style images.
• realistic: photorealistic-style images.

The release contains only the UniCAD team's generated images and lightweight provenance metadata. It does not redistribute Text2CAD descriptions, Text2CAD CadQuery programs, or the original DeepCAD renders. Records retain the upstream UID so that users who obtain those datasets under their own terms can join them locally.

Dataset size

Style	Train CADs / images	Validation CADs / images	Test CADs / images	Total images
Sketch	67,457 / 269,828	5,690 / 22,760	7,105 / 28,420	321,008
Realistic	69,144 / 276,576	5,849 / 23,396	7,934 / 31,736	331,708
Total	136,601 / 546,404	11,539 / 46,156	15,039 / 60,156	652,716

Some UIDs occur in both visual styles. Each CAD has exactly four views within a style and split. The source PNG WebDataset shards total 723.61 GiB.

Format

The images are stored as WebDataset tar shards. Each image is one WebDataset sample with the following pair of members:

<uid>_<view>.png
<uid>_<view>.json

The JSON metadata fields are:

• uid: the upstream CAD identifier.
• split: train, validation, or test.
• style: sketch or realistic.
• view_index: integer from 1 through 4.
• source_dataset: upstream source family used for provenance.
• source_sha256: checksum of the generated source PNG.
• image_sha256: checksum of the released image bytes.
• encoding, width, and height: released image properties.

Four records with the same (uid, split, style) form one multi-view CAD sample. Shard manifests and SHA-256 checksums are included in manifests/ and checksums/.

Loading

The dataset can be streamed with the Hugging Face datasets library or any WebDataset-compatible reader:

from datasets import load_dataset

dataset = load_dataset(
    "webdataset",
    data_files={
        "train": "hf://datasets/Chen8566/UniCAD-Synthetic-Views/data/sketch/train/*.tar"
    },
    split="train",
    streaming=True,
)

Creation and provenance

The image-editing model was Qwen/Qwen-Image-Edit. Archived generation scripts use seed 0 and true_cfg_scale=4.0. The realistic prompt was Transform the image into photorealistic style. The sketch generation scripts used either Transform the image into hand-drawn sketch style. or a longer prompt requesting a faithful hand-drawn rendition without adding parts. Inference steps vary across generation batches (6, 8, 10, or 50). Per-image generation parameters were not logged, so the exact batch settings cannot be recovered for every individual image.

UID selection and train/validation/test membership follow Text2CAD annotations. The underlying CAD render family comes from DeepCAD. Invalid, undecodable, and all-black images were excluded by the UniCAD training data loader; a CAD is included only when all four views for that style can be decoded and are not all-black.

Upstream resources and attribution

• DeepCAD: https://github.com/rundiwu/DeepCAD
• Text2CAD: https://github.com/SadilKhan/Text2CAD
• Qwen-Image-Edit: https://huggingface.co/Qwen/Qwen-Image-Edit

Please cite and comply with the applicable upstream terms when combining this release with upstream descriptions, CAD programs, or original renders.

License

The UniCAD-generated portion of this release is distributed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). See LICENSE.md for attribution and scope.

Limitations

• These are synthetic image edits rather than photographs or human drawings.
• Generated images may contain image-model artifacts or geometric deviations.
• Coverage differs between the sketch and realistic styles.
• The dataset does not by itself contain target CAD code or natural-language descriptions; join by uid with properly obtained upstream resources.