de233/AnchorSplat-3DGS-SR-Train-15K

AnchorSplat 3DGS-SR Train 15K

This dataset release contains the 14,525 training scenes used for AnchorSplat 3DGS-SR training.

Each scene contains a low-resolution 3DGS input at 256 resolution and high-resolution supervision at 1024 resolution.

Files

• parts/*.tar.zst: independent tar+zstd archives. Each part can be downloaded and extracted separately.
• parts/*.tar.zst.sha256: per-part SHA256 checksum.
• SHA256SUMS.txt: checksums for all archive parts.
• MANIFEST.jsonl: one line per scene.
• PARTS.json: scene ranges and source sizes for each part.

Layout

After extraction:

3dgs-sr/train/<scene_id>/
  256/
  1024/

Download

Download all files:

huggingface-cli download de233/AnchorSplat-3DGS-SR-Train-15K \
  --repo-type dataset \
  --local-dir data/downloads/3dgs-sr-train-15k

Download one archive part:

huggingface-cli download de233/AnchorSplat-3DGS-SR-Train-15K \
  parts/3dgs_sr_train_15k_part_000.tar.zst \
  parts/3dgs_sr_train_15k_part_000.tar.zst.sha256 \
  --repo-type dataset \
  --local-dir data/downloads/3dgs-sr-train-15k

Verify

cd data/downloads/3dgs-sr-train-15k
sha256sum -c SHA256SUMS.txt

For a single downloaded part:

sha256sum -c parts/3dgs_sr_train_15k_part_000.tar.zst.sha256

Extract

mkdir -p /path/to/data

for f in data/downloads/3dgs-sr-train-15k/parts/3dgs_sr_train_15k_part_*.tar.zst; do
  tar --zstd -xf "$f" -C /path/to/data
done

Set data/3dgs-sr/train as the training data path in the AnchorSplat repository.

Citation

@article{zhu2026anchorsplat,
  title={AnchorSplat: Fast and Structure Consistent Detail Synthesis for Gaussian Splatting},
  author={Zhu, Dexu and Shao, Jiangnan and Wang, Xiaofeng and Duan, Junxian and Cao, Jie and Zhu, Zheng and Huang, Huaibo},
  journal={arXiv preprint arXiv:2607.01290},
  year={2026}
}