ArticuTable: Generating Instance-Level Interactive Rigid-Articulated 3D Tabletop Scenes from a Single Image

Kai Lv^*,1 Yibo Yin^*,2 Lijun Guo¹ Heng Fan³ Kaihao Zhang⁴ Xingping Dong^†,1

¹ Wuhan University² EPFL³ University of North Texas⁴ Australian National University

* Equal contribution · † Corresponding author

arXiv:2610.05249 · 2026

Teaser video. A reconstructed tabletop scene with executable part-level motion.

EXPLORE THE RECONSTRUCTION

One image. An interactive world.

Rotate the tabletop, inspect each object, and control its recovered joints.

Click an object or choose it above, then drag its joint controls. Play follows the Isaac Sim video's joint trajectory. The poster and videos show the original simulator rendering; this WebGL view uses different lighting.

Abstract

Embodied agents benefit from 3D environments that combine visual fidelity to real-world observations with physical interactivity. Existing single-image tabletop reconstruction methods recover plausible scene geometry but typically represent objects as monolithic rigid bodies, limiting interaction to whole-object rigid motion and precluding executable part-level articulation. Meanwhile, recovering a scene layout consistent with the input view remains challenging because a single observation may admit multiple plausible pose-scale configurations.

We present ArticuTable, a single-image 3D tabletop reconstruction framework that recovers both executable part-level articulation and an input-view-consistent scene layout. For object modeling, generation-robust articulation modeling (GRAM) combines joint fitting guided by a multimodal large language model with semantic state reasoning to recover reliable joint parameters and valid motion ranges from imperfect proxy meshes. For scene layout, progressive semantic-geometric scene registration (PSGSR) progressively narrows the pose-scale search space using complementary metric, planar, and input-view constraints. We also contribute ArticuTable-100, a curated collection of 100 simulation-ready tabletop scenes. Extensive evaluation, including a user study, demonstrates strong performance across visual fidelity, scene consistency, articulation quality, physical plausibility, and simulation readiness.

Scenes in Motion

Switch between six additional simulation-ready scenes without opening a wall of video players.

Method

A single-image pipeline for reconstructing articulated assets and registering them in a consistent 3D scene.

ArticuTable pipeline: asset reconstruction, GRAM articulation modeling, PSGSR scene registration, and simulation-ready output — **Figure 2.** Overview of the ArticuTable framework.

01 · ARTICULATION

Generation-Robust Articulation Modeling

GRAM turns imperfect proxy meshes into executable assets by recovering joint structure, parameters, and motion ranges.

02 · REGISTRATION

Progressive Scene Registration

PSGSR aligns object position, orientation, and scale through metric, planar, and input-view constraints.

Reconstruction Results

Comparisons on generated and real-world tabletop images.

One RGB tabletop image reconstructed as an articulated scene under multiple joint configurations — **Figure 1.** Single-image reconstruction of articulated tabletop scenes under different joint configurations.

Qualitative comparisons between input images, baseline 3D reconstruction methods, and ArticuTable — **Figure 3.** Qualitative comparison with existing single-image tabletop reconstruction methods.

Interactive and simulation-ready

In-scene editing and physics validation with object removal, insertion, rearrangement, dragging, and articulated opening — **Figure 6.** In-scene editing and physics validation, including articulated opening.

Quantitative Evaluation

Scene reconstruction, articulated assets, component ablations, and human preference. Values below reproduce Tables 1–5 of the paper.

Scene-level reconstruction

ArticuTable leads on input-view consistency, perceptual assessment, and physical validity.

Table 1 · Scene-level results
Method	LPIPS ↓	DINOv2 ↑	CLIP ↑	VF ↑	IA ↑	PP ↑	Avg. ↑	OR ↓	ColO ↓	ColS ↓
ACDC	0.4447	0.5135	0.8058	3.927	1.920	3.880	3.242	4.298	1.5302%	32.00%
Gen3DSR	0.2716	0.5704	0.8386	2.220	4.680	4.013	3.638	3.349	11.3322%	66.67%
MIDI	0.4302	0.6469	0.8668	2.787	2.933	2.360	2.693	4.142	37.9430%	99.33%
TabletopGen	0.3935	0.7788	0.8908	5.380	4.860	5.893	5.378	2.162	1.4488%	16.00%
ArticuTable	0.2398	0.8991	0.9265	5.867	6.253	6.600	6.240	1.049	0.3819%	10.67%

VF: visual fidelity; IA: input alignment; PP: physical plausibility; OR: overall rank. ColO and ColS report collision rates.

Articulated asset quality

Lightwheel benchmark results, grouped by image- or mesh-conditioned input.

Table 2 · Lightwheel articulation benchmark
Method	Input	Training-free	Prec. ↑	Rec. ↑	Rest mIoU ↑	Art. gIoU ↑	OC ↓	AE ↓	LE ↓
SINGAPO	Image	–	37.60	25.40	0.182	−0.263	0.030	30.30°	0.077
PAct	Image	–	24.60	16.80	0.092	−0.434	0.056	42.90°	0.188
PhysX-Anything	Image	–	20.30	19.20	0.093	−0.459	0.064	38.80°	0.123
URDF-Anything+	Image	–	70.70	23.90	0.260	−0.138	0.033	50.40°	0.128
Articulate AnyMesh	Mesh	✓	89.20	42.50	0.452	0.158	0.010	21.70°	0.043
URDF-Anything+	Mesh	–	77.50	27.40	0.267	−0.110	0.026	49.90°	0.111
Particulate	Mesh	–	89.90	51.50	0.576	0.305	0.009	20.90°	0.040
GRAM	Mesh	✓	91.15	46.54	0.564	0.283	0.014	17.26°	0.034

GRAM ablation

Table 3 · Component ablation
Component	Metric	Enabled	Disabled	Reduction
MLLM-guided joint fitting	AE ↓	17.17°	40.06°	57.1%
MLLM-guided joint fitting	LE ↓	0.0417	0.0527	20.9%
Semantic state reasoning	Art. PC ↓	0.2502	0.2605	3.9%
Semantic state reasoning	OC ↓	0.0152	0.0176	13.6%

PSGSR ablation

Table 4 · Registration stages
Variant	Top	Input	SASPS	LPIPS ↓	DINOv2 ↑	CLIP ↑
Point-cloud only	–	–	–	0.3439	0.8102	0.9077
w/o Top-view	×	✓	–	0.2564	0.8702	0.9210
Top-view only	✓	×	×	0.3235	0.8586	0.9156
w/o SASPS	✓	✓	×	0.2543	0.8762	0.9205
Full model	✓	✓	✓	0.2398	0.8991	0.9265

User study

59 participants; each method received 590 valid rankings.

Table 5 · Human preference rankings
Metric	ACDC	Gen3DSR	MIDI	TabletopGen	ArticuTable
Mean rank ↓	4.08	3.75	3.75	2.24	1.17
Median rank ↓	4	4	4	2	1
First-place rate ↑	0.17%	1.02%	1.02%	12.20%	85.59%
Top-2 rate ↑	6.61%	11.86%	8.81%	74.58%	98.14%

Citation

@misc{lv2026articutablegeneratinginstancelevelinteractive,
  title={ArticuTable: Generating Instance-Level Interactive Rigid-Articulated 3D Tabletop Scenes from a Single Image},
  author={Kai Lv and Yibo Yin and Lijun Guo and Heng Fan and Kaihao Zhang and Xingping Dong},
  year={2026},
  eprint={2610.05249},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.05249}
}