GALA-0.7B: Geometry-Aware Language Model for Controllable Object Arrangement

SIGGRAPH Asia 2026

Project Page | arXiv | Paper (DOI) | Code (GitHub) | Dataset (ChuchengXiang/GALA-Dataset)

---

Model Overview

GALA-0.7B is a geometry-aware multimodal language model for controllable 3D indoor object arrangement. Unlike methods that rely solely on oriented bounding boxes (OBBs), GALA perceives explicit per-object 3D point clouds (floor, existing objects, and target object) alongside natural language placement instructions to autoregressively predict the 6-DoF pose [x, y, z, roll, pitch, yaw] of the target object. This explicit geometric grounding enables accurate reasoning over fine-grained spatial constraints such as containment (within open shelves or cabinets) and physical support (on, under).

Architecture & Training Details

Component	Specification
Point Cloud Encoder	Utonia (PTv3 Mode3, 9-channel xyz+rgb+normal, 0.02m grid size, 576-dim output, float32, frozen)
Multimodal Projector	2-layer MLP with GELU (576 -> 1024, bfloat16)
LLM Backbone	Qwen/Qwen3-0.6B (28 layers, hidden_size=1024, bfloat16)
Prompt Placeholders	<floor_pointcloud>, <object_pointcloud>, <new_object_pointcloud>
Tokenizer Special Tokens	<|point_start|> (151669), <|point_end|> (151670), <|point_pad|> (151671)
Stage 1 (Alignment)	3 epochs on 500k point cloud understanding tasks (3dish_alignment_500k_train), training projector only (lr=1e-3, global batch size 256)
Stage 2 (Placement)	Fine-tuned on 331k autoregressive placement steps (3dish_separate_pointcloud_331k_train, llm_lr=2e-5, projector_lr=5e-4, global batch size 128)
Checkpoint Step	Step 6240 (checkpoint-6240, Epoch 2.67) — the exact checkpoint reported in the SIGGRAPH Asia 2026 paper

---

Benchmark Performance (Held-Out Test Split, 15,991 Samples)

Evaluated on 3dish_separate_pointcloud_331k_test.json (15,991 total placement steps, including 3,165 within containment samples):

Method	Input	Within Valid ($\uparrow$)	Instruction Valid ($\uparrow$)	Floating Rate ($\downarrow$)	Support Ratio ($\uparrow$)
ATISS	OBB + Category	--	--	0.93%	99.20%
ReSpace	OBB + Category	--	--	5.79%	94.20%
LayoutVLM	OBB + Instruction	10.96%	37.71%	17.05%	82.91%
GPT-OSS-120B	OBB + Instruction	22.53%	42.29%	7.64%	93.54%
GALA-0.7B (Step 6240)	Point Cloud + Instruction	65.62%	78.59%	4.63%	96.03%

---

Citation

@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}
}

License

This model checkpoint is released under the Apache License 2.0.