dscdyc/partversexl

Dataset

4

stars

26

commits

1

linked in READMEs

Mar 26, 2026

updated

README

PartVerse-XL (FullPart) — 3D Parts Dataset

A large-scale 3D parts dataset. It contains textured part meshes, whole-object meshes, annotations, captions. Related dataset: PartVerse.

Contents

  • Split archives (parts):
    • textured_part_glbs.tar.gz00 … textured_part_glbs.tar.gz66 (67 parts)
    • normalized_glbs.tar.gz00 … normalized_glbs.tar.gz08 (9 parts)
  • Other files:
    • anno_infos.tar.gz
    • text_captions.json
    • train.csv
    • val.csv
    • metadata.csv

What’s inside (after extraction)

  • textured_part_glbs/: GLB meshes for each decomposed part with textures.
  • normalized_glbs/: GLB meshes for full, normalized objects.
  • anno_infos/: auxiliary files for part statistics and processing.
  • text_captions.json: auto-generated captions for parts (from a VLM).
  • train.csv / val.csv: instance lists for train/eval.
  • metadata.csv: metadata aligned with Objaverse-XL.

Expected directory layout:

dataset/
├── textured_part_glbs/
├── normalized_glbs/
├── anno_infos/
├── text_captions.json
├── train.csv
├── val.csv
└── metadata.csv

Download

Using huggingface_hub:

from huggingface_hub import snapshot_download

local_dir = "dataset"
snapshot_download(
    repo_id="dscdyc/partversexl",
    repo_type="dataset",
    local_dir=local_dir,
    local_dir_use_symlinks=False
)

Join split archives and extract

cd dataset
cat ${TARGET_PACKAGE}.tar.gz* | tar xzf - -C /path/to/target/

Intended uses

  • Part-level generation, assembly, retrieval, and analysis
  • Whole-object understanding and alignment with parts
  • Text–shape alignment and retrieval

Citation

@misc{ding2025fullpart,
  title        = {FullPart: Generating each 3D Part at Full Resolution},
  author       = {Lihe Ding and Shaocong Dong and Yaokun Li and Chenjian Gao and Xiao Chen and Rui Han and Yihao Kuang and Hong Zhang and Bo Huang and Zhanpeng Huang and Zibin Wang and Dan Xu and Tianfan Xue},
  year         = {2025},
  eprint       = {2510.26140},
  archivePrefix= {arXiv},
}

Changelog

  • Initial release. Some split archives are still uploading; will update once completed.

Contributors

dscdyc

26 commits

dscdyc/partversexl

Dataset

4

stars

26

commits

1

linked in READMEs

Mar 26, 2026

updated

README

PartVerse-XL (FullPart) — 3D Parts Dataset

A large-scale 3D parts dataset. It contains textured part meshes, whole-object meshes, annotations, captions. Related dataset: PartVerse.

Contents

  • Split archives (parts):
    • textured_part_glbs.tar.gz00 … textured_part_glbs.tar.gz66 (67 parts)
    • normalized_glbs.tar.gz00 … normalized_glbs.tar.gz08 (9 parts)
  • Other files:
    • anno_infos.tar.gz
    • text_captions.json
    • train.csv
    • val.csv
    • metadata.csv

What’s inside (after extraction)

  • textured_part_glbs/: GLB meshes for each decomposed part with textures.
  • normalized_glbs/: GLB meshes for full, normalized objects.
  • anno_infos/: auxiliary files for part statistics and processing.
  • text_captions.json: auto-generated captions for parts (from a VLM).
  • train.csv / val.csv: instance lists for train/eval.
  • metadata.csv: metadata aligned with Objaverse-XL.

Expected directory layout:

dataset/
├── textured_part_glbs/
├── normalized_glbs/
├── anno_infos/
├── text_captions.json
├── train.csv
├── val.csv
└── metadata.csv

Download

Using huggingface_hub:

from huggingface_hub import snapshot_download

local_dir = "dataset"
snapshot_download(
    repo_id="dscdyc/partversexl",
    repo_type="dataset",
    local_dir=local_dir,
    local_dir_use_symlinks=False
)

Join split archives and extract

cd dataset
cat ${TARGET_PACKAGE}.tar.gz* | tar xzf - -C /path/to/target/

Intended uses

  • Part-level generation, assembly, retrieval, and analysis
  • Whole-object understanding and alignment with parts
  • Text–shape alignment and retrieval

Citation

@misc{ding2025fullpart,
  title        = {FullPart: Generating each 3D Part at Full Resolution},
  author       = {Lihe Ding and Shaocong Dong and Yaokun Li and Chenjian Gao and Xiao Chen and Rui Han and Yihao Kuang and Hong Zhang and Bo Huang and Zhanpeng Huang and Zibin Wang and Dan Xu and Tianfan Xue},
  year         = {2025},
  eprint       = {2510.26140},
  archivePrefix= {arXiv},
}

Changelog

  • Initial release. Some split archives are still uploading; will update once completed.

Contributors

dscdyc

26 commits