SceneEdit3D-15K

SceneEdit3D-15K is a large-scale paired 3D scene-editing dataset introduced in JointEdit3D: Feed-Forward 3D Scene Editing in a Unified Latent Space. It contains 15,319 Blender-rendered indoor-scene editing samples with paired source and edited renderings, natural-language edit instructions, edited reference frames, edit masks, depth maps, camera intrinsics, and camera poses.

The dataset covers five edit groups: Add, Delete, Move, Appearance, and Multi-op. It is constructed from composable Imaginarium indoor scenes. Candidate edits are proposed from scene content and layout, executed in Blender, and rendered before and after editing under the same camera trajectory. This provides paired multi-view supervision and renderer-provided 3D annotations for both training and evaluation.

Dataset splits

SceneEdit3D-15K contains 13,799 training samples and a scene-disjoint 1,520-sample held-out (test/val) split. The 100-sample SceneEdit3D-Bench benchmark is curated from this held-out split. In this release, the benchmark is provided as test; the remaining 1,420 held-out samples are provided as validation.

released split	samples	description
train	13,799	Training split from 135 scenes.
validation	1,420	Non-benchmark portion of the scene-disjoint held-out (test/val) split. It is not part of the training split.
test	100	SceneEdit3D-Bench, a fixed held-out benchmark from 15 scenes.

The full split composition is:

split	Add	Delete	Move	Appearance	Multi-op
train (13,799)	4,990	4,987	1,884	1,017	921
held-out test/val (1,520)	600	600	186	24	110

SceneEdit3D-Bench

SceneEdit3D-Bench is the 100-sample evaluation benchmark used in JointEdit3D. It is stratified across edit operations and includes both regular and challenging edits, such as partially occluded objects, small edited regions, large edited regions, and multi-operation edits. It supports evaluation of edit fidelity, background preservation, and 3D structure under a common paired rendering protocol.

Add	Delete	Move	Appearance	Multi-op
29	29	14	14	14

The renderer-defined edited-area distribution is 11 samples below 1%, 47 in 1–5%, 29 in 5–15%, and 13 at or above 15% of image pixels.

Contents

metadata/train.jsonl
metadata/validation.jsonl
metadata/test.jsonl
shards/train_rgb_mask-*.tar
shards/train_geometry-*.tar
shards/validation_rgb_mask-*.tar
shards/validation_geometry-*.tar
shards/test_rgb_mask-*.tar
shards/test_geometry-*.tar

Each sample is stored under a common sample ID in two tar-shard families:

<id>/before/frames/000000.png
<id>/before/masks/000000.png
<id>/after/frames/000000.png
<id>/after/masks/000000.png

<id>/before/depth/000000.exr
<id>/before/intrinsics.txt
<id>/before/poses.txt
<id>/after/depth/000000.exr
<id>/after/intrinsics.txt
<id>/after/poses.txt

*_rgb_mask-*.tar contains source/edited RGB frames and edit masks. *_geometry-*.tar contains depth maps and camera parameters. The tar shards are uncompressed. Depth maps are provided as 16-bit OpenEXR files.

Metadata

Each row in a split JSONL file has the following format:

{
  "id": "00000000",
  "rgb_mask": "train_rgb_mask-*.tar:00000000",
  "geometry": "train_geometry-*.tar:00000000",
  "edited_frame_index": 24,
  "edit_prompt": "...",
  "num_frames": 49
}

The shard fields identify the tar family and sample prefix. Use the matching sample ID to read RGB/mask and geometry data together.

Loading

The shards can be read with WebDataset or any standard tar reader. Select a split from metadata/, locate the corresponding sample ID in both tar families, and load the before/after frames, masks, depth maps, and camera parameters as needed.

License and attribution

SceneEdit3D-15K is derived from Imaginarium scene assets and is released under CC BY-NC-SA 4.0. Please retain attribution to the Imaginarium authors and distribute derivative works under the same license. This is a non-commercial research release.

Citation

@article{zhu2026jointedit3d,
  title   = {JointEdit3D: Feed-Forward 3D Scene Editing in a Unified Latent Space},
  author  = {Zhu, Xinnan and Xu, Ruijie and Ying, Jiayu and Dong, Daoguo and Xu, Jiachen and Xie, Yuan and Tan, Xin},
  journal = {arXiv preprint arXiv:2606.13345},
  year    = {2026},
  url     = {https://arxiv.org/abs/2606.13345}
}