MQ66/seen2scene-FRONT-3D

Seen2Scene — 3D-FRONT Sample Scenes

A 1000-scene sample of the processed 3D-FRONT data used to train and evaluate Seen2Scene. This is enough to run the completion and generation inference commands end to end without regenerating the TSDF fusion pipeline yourself.

The full processed dataset is 4.65 TB across 6,804 scenes; this sample is 720 GB (15%).

What the scenes are

Each scene is a synthetic indoor 3D-FRONT layout that has been rendered to depth and fused into truncated signed distance fields (TSDFs) stored as sparse OpenVDB grids at voxel size 0.011 m.

The p_ value is the visibility level — the fraction of the camera trajectory used for fusion. p_1.0 fuses every frame (the most complete reconstruction, used as the generation target); p_0.01 fuses 1% of frames (a very sparse partial scan). This axis is what "visibility-guided flow matching" is trained over.

File	Read by the training/inference dataloader?	Contents
meta.json	Yes — required	Scene bounds, per-object bounding boxes and numeric category ids (the layout conditioning signal)
fusion_p_1.0_v_0.011.vdb	Yes — --latent-key tsdf_p_1.0	Full-visibility fused TSDF
fusion_p_0.1_v_0.011.vdb	Yes — --src-key tsdf_p_0.1	10%-visibility partial scan, the completion input
fusion_p_{0.01,0.05,0.2,0.5}_v_0.011.vdb	Only if you pass that key	Other visibility levels, for sweeping --src-key
scene.ply	Optional	Ground-truth scene mesh, used for validation metrics/visualization; falls back to an empty mesh if absent
cameras_p_*.ply	No	Camera positions used at each visibility level (provided for inspection only)

Category ids in meta.json are decoded through assets/front3d_mapping.csv in the code repository.

Scene selection

All 1000 scenes are held out from training. Seen2Scene splits scenes positionally — sorted by folder name, then 80% train / 5% val / 15% test — over the 6,575 scenes that pass its file-existence check. This sample is:

• 987 scenes — the entire test split
• 13 scenes from the val split, including d99124eb-078f-480a-805b-b605da337245, the scene used for the released qualitative demo outputs (remaining 12 sampled with seed 20260906)
• 0 scenes from the training split

So every scene the model was evaluated on is here, and nothing it was trained on is. The one scene the codebase excludes for data-quality reasons (979a4a2f-32c4-42fa-b078-7466c093bcf2, a two-floor layout) is not included.

The dataloader recomputes the 80/5/15 split over whatever scenes it finds on disk. If you point it at this 1000-scene sample, it will label the first 800 as "train" and only the last 150 as "test" — even though all 1000 were held out from the real training run. Target scenes explicitly with --data.scene-names <scene_id> ... rather than relying on split when working from this sample.

Download

The repository mirrors the layout the code expects by default (SEEN2SCENE_DATA_ROOT/3D-FRONT/v3/<scene_id>/), so downloading into the code repository root needs no further configuration:

pip install -U "huggingface_hub[cli]"

# From the root of your Seen2Scene checkout:
hf download MQ66/seen2scene-FRONT-3D --repo-type dataset --local-dir .

That places scenes at data/3D-FRONT/v3/<scene_id>/, which is the default SEEN2SCENE_FRONT3D_DIR. If you keep the data elsewhere, point the code at it:

export SEEN2SCENE_FRONT3D_DIR=/path/to/data/3D-FRONT/v3

Smaller download (~304 GB instead of ~720 GB)

The README's completion and generation commands only read the p_0.1 and p_1.0 levels. To skip the other four visibility levels and the camera files:

hf download MQ66/seen2scene-FRONT-3D --repo-type dataset --local-dir . \
  --include "data/3D-FRONT/v3/*/meta.json" \
            "data/3D-FRONT/v3/*/scene.ply" \
            "data/3D-FRONT/v3/*/fusion_p_0.1_v_0.011.vdb" \
            "data/3D-FRONT/v3/*/fusion_p_1.0_v_0.011.vdb"

A single scene

hf download MQ66/seen2scene-FRONT-3D --repo-type dataset --local-dir . \
  --include "data/3D-FRONT/v3/d99124eb-078f-480a-805b-b605da337245/*"

Usage

With the checkpoints downloaded as well:

# Partial-scan completion on the demo scene.
python -m seen2scene.main control task:completion \
  --ae-log 2025-12-19_01-23-28-525 \
  --gen-log 2026-02-23_16-22-25-152 \
  --ckpt-path 2026-02-26_14-01-23-930 \
  --src-key tsdf_p_0.1 \
  --latent-key tsdf_p_1.0 \
  --data.data-list 3D-FRONT \
  --data.scene-names d99124eb-078f-480a-805b-b605da337245 \
  --task.export-as bbox mesh volume \
  --slurm.cluster local

Reading a grid directly:

from seen2scene.tools.vdb_utils import loadvdb
grid = loadvdb("data/3D-FRONT/v3/d99124eb-078f-480a-805b-b605da337245/fusion_p_1.0_v_0.011.vdb")

License and attribution

These files are derived data — TSDF reconstructions computed from the 3D-FRONT dataset released by Alibaba. They contain fused scene geometry and layout bounding boxes; they do not redistribute the original 3D-FRONT JSON layouts, 3D-FUTURE furniture models, or texture assets.

3D-FRONT is distributed by its authors under their own terms, for non-commercial research use. By downloading this sample you agree to comply with the original 3D-FRONT license and terms of use, which govern the underlying data. Please cite the original 3D-FRONT papers alongside Seen2Scene:

@inproceedings{fu20213dfront,
  title={3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics},
  author={Fu, Huan and Cai, Bowen and Gao, Lin and Zhang, Ling-Xiao and Wang, Jiaming and
          Li, Cao and Zeng, Qixun and Sun, Chengyue and Jia, Rongfei and Zhao, Binqiang and Zhang, Hao},
  booktitle={ICCV},
  year={2021}
}

@misc{meng2026seen2scene,
      title={Seen2Scene: Completing Realistic 3D Scenes with Visibility-Guided Flow},
      author={Quan Meng and Yujin Chen and Lei Li and Matthias Nießner and Angela Dai},
      year={2026},
      eprint={2603.28548},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.28548},
}

The Seen2Scene code and checkpoints are MIT-licensed; that license does not extend to the underlying 3D-FRONT data.