GALA-Dataset (3DISH): Multi-Source Geometry-Aware 3D Scene Arrangement Benchmark

SIGGRAPH Asia 2026

Project Page | arXiv | Paper (DOI) | Code (GitHub) | Model (ChuchengXiang/GALA-0.7B)

---

Dataset Overview

GALA-Dataset (3DISH) is a curated multi-source dataset for instruction-driven 3D indoor object placement and geometry-aware point cloud understanding, introduced in our SIGGRAPH Asia 2026 paper "GALA: Geometry-Aware Language Model for Controllable Object Arrangement".

It combines four complementary 3D indoor scene sources (3D-FRONT, Imaginarium, SpatialLM, and Artist-Crafted / Handcrafted scenes) into a unified Z-up coordinate frame, filtered for physical plausibility (out-of-bounds, collision, and floating removal), augmented via 4-way rotation ($0^\circ, 90^\circ, 180^\circ, 270^\circ$ on high-quality subsets), and decomposed into scene-tree autoregressive placement sequences alongside Stage-1 geometric alignment tasks.

Dataset Statistics

Variant	Full File	Total Samples	Train	Val	Test	Description
Separate Pointcloud	3dish_separate_pointcloud.json	331,072	298,811	16,270	15,991	Autoregressive 6D pose prediction with per-object point clouds
Base (Text-Only)	3dish.json	331,072	298,811	16,270	15,991	Autoregressive 6D pose prediction (OBB + instruction only)
Stage-1 Alignment	3dish_alignment.json	500,000	479,702	11,894	8,404	8 point-cloud understanding subtasks for projector alignment

All train/val/test splits are stratified by scene family at an 18:1:1 (90:5:5) ratio so that rotation-augmented variants of the same scene never cross splits.

---

Source Datasets, Processing & Licensing

Important Note: The scene layouts and assets in this repository are processed and curated versions derived from upstream open-source datasets. Specifically, we perform coordinate unification (X-right, Y-forward, Z-up), blacklisted asset removal, tight room boundary computation, out-of-bounds and mesh-collision filtering, 8,192-point surface sampling, and scene-tree autoregressive instruction synthesis.

Please cite and comply with the licenses of the original upstream datasets when using the corresponding subsets:

Source Prefix	Upstream Dataset & Repository	Scenes (Filtered)	Placement Steps	Upstream License & Notes
3dfront_	SSR-3DFRONT (ReSpace) (GitHub) & 3D-FRONT / 3D-FUTURE	8,620	67,685	MIT License (SSR-3DFRONT annotations) / CC BY-NC-SA 4.0 (3D-FRONT & 3D-FUTURE raw meshes, which must be obtained from the official Alibaba Tianchi portal).
imaginarium_	Imaginarium-Dataset (GitHub)	548 (137 × 4 rot)	22,508	CC BY 4.0. We provide our modified and cleaned imaginarium_assets.tar.gz (with pre-converted .obj/.mtl, 8,192-point .ply clouds, and imaginarium_asset_info.csv) and refined scene layouts.
spatiallm_	SpatialLM-Dataset (GitHub)	11,734	201,835	CC BY-NC-SA 4.0. Filtered for rectangular room boundaries and physical collisions; paired with Imaginarium 3D assets.
handcrafted_	GALA Artist-Crafted Scenes (Ours)	1,200 (300 × 4 rot)	39,044	CC BY 4.0. 300 artist-designed scenes with rich within/on/under arrangements using Imaginarium and 3D-FUTURE assets.

---

Repository Structure

ChuchengXiang/GALA-Dataset/
├── README.md
├── assets/
│   └── imaginarium_assets.tar.gz          # Curated Imaginarium 3D assets (.fbx, .obj, .mtl, textures, 8192-pt .ply, imaginarium_asset_info.csv)
├── scenes/
│   ├── Imaginarium_scenes.tar.gz          # Imaginarium scenes (flux_train_data, converted_scenes, filtered_scenes & autoregressive_data)
│   ├── handcrafted_scenes.tar.gz          # 300 artist-crafted scenes (1,200 with rotation augmentation) + autoregressive data
│   ├── SSR-3DFRONT_scenes.tar.gz          # Filtered SSR-3DFRONT scenes, collision logs, autoregressive data & 3dfront_asset_info.csv
│   ├── spatiallm_scenes.tar.gz            # Filtered SpatialLM scenes, collision logs & autoregressive data
│   └── 3dish_autoregressive.tar.gz        # Merged 22,102 autoregressive scene files (331,072 steps)
└── processed/
    ├── DATA_FORMAT.md                     # Detailed schema documentation for the dataset variants
    ├── 3dish_metadata.csv                 # Extended step-by-step metadata CSV
    ├── 3dish_separate_pointcloud.json     # Full Separate Pointcloud dataset (331,072 samples)
    ├── 3dish_separate_pointcloud_331k_{train,val,test}.json   # Separate Pointcloud splits (main GALA benchmark)
    ├── 3dish.json                         # Full Base text-only dataset (331,072 samples)
    ├── 3dish_331k_{train,val,test}.json   # Base (text-only) splits
    ├── 3dish_alignment.json               # Full Stage-1 Projector Alignment dataset (500,000 samples)
    ├── 3dish_alignment_500k_{train,val,test}.json             # Stage-1 Projector Alignment splits
    ├── 3dish_no_imaginarium_handcrafted_separate_pointcloud_{train,val}.json  # Ablation split (w/o artist/high-quality scenes)
    ├── 3dish_imaginarium_handcrafted_separate_pointcloud_test.json            # Stratified test subset: High-quality scenes (2,900 samples)
    └── 3dish_3dfront_spatiallm_separate_pointcloud_test.json                  # Stratified test subset: 3D-FRONT & SpatialLM (13,091 samples)

---

Quick Setup & Local Point Cloud Generation

To keep the repository lightweight, the 2.55M per-step transformed point cloud .ply files (datasets/pointclouds/separate_pointcloud/3dish/sample_XXXXXX/*.ply) are not stored as raw files on the Hub; instead, they are deterministically generated locally from the asset point clouds in a few minutes.

1. Download & Extract

huggingface-cli download ChuchengXiang/GALA-Dataset --repo-type dataset --local-dir datasets_hf

mkdir -p datasets/Imaginarium datasets/processed
cp -r datasets_hf/processed/* datasets/processed/
tar -xzf datasets_hf/assets/imaginarium_assets.tar.gz -C datasets/Imaginarium

for archive in datasets_hf/scenes/*.tar.gz; do
    tar -xzf "$archive" -C datasets/
done

For 3D-FRONT training/evaluation: Place the official 3D-FUTURE-model directory under datasets/3D-FRONT/3D-FUTURE-model and copy datasets/SSR-3DFRONT/3dfront_asset_info.csv into datasets/3D-FRONT/3D-FUTURE-model/3dfront_asset_info.csv, then run Step 1 of [scripts/run_data_pipeline.sh](https://github.com/xccElephant/GALA/blob/main/scripts/run_data_pipeline.sh) to sample 8,192-point clouds for 3D-FUTURE assets.

2. Generate Per-Sample Point Clouds & Utonia Cache

Using the official GALA GitHub repository:

# Export per-sample floor / existing / new_object .ply point clouds (Step 13)
python dataset_pipeline/generation/02_generate_pointcloud_training_data.py \
    --dataset 3dish \
    --output-json-name 3dish_separate_pointcloud.json

# Pre-cache Utonia 0.02m grid-subsampled features (.utonia.npy) for fast training
cd third_party/LLaMA-Factory
python scripts/pointlm_cache_pcd.py \
    --data_files ../../datasets/processed/ \
    --backbone_type utonia \
    --utonia_grid_size 0.02 \
    --workers 64

---

Citation

If you use this dataset in your research, please cite our SIGGRAPH Asia 2026 paper as well as the underlying source datasets:

@inproceedings{xiang2026gala,
  title     = {GALA: Geometry-Aware Language Model for Controllable Object Arrangement},
  author    = {Xiang, Chucheng and Wang, Runze and Deng, Zhi and Bao, Ruchao and Zhang, Yuanwei and Xie, Yuxuan and Wang, Hanliu and Fei, Liangzhen and Wu, Wenzheng and Wan, Cheng and Li, Peifeng and Liu, Zhongyuan and Liu, Ligang},
  booktitle = {SIGGRAPH Asia 2026 Conference Papers (SA Conference Papers '26)},
  year      = {2026},
  doi       = {10.1145/3829340.3842289}
}