Reconstructing interactable articulated assets from a few images is key to robotics and simulation, yet highly underconstrained. Existing optimization-based methods fit an articulated asset under a static 3D prior that scores each state independently, so per-state reconstruction errors leak into the estimated moving part and joint as false motion. Video diffusion models capture how objects move between states, but only implicitly in 2D. To this end, we present TransArt3D, an optimization-based framework that reconstructs an articulated 3D asset from two endpoint images by translating this video transition prior into explicit part structure and joint kinematics. It first lifts frames generated between the endpoints with a transition-coupled 3D consensus that applies the static 3D prior jointly across frames, stabilizing unchanging regions and initializing the parts and joint. A rectified-flow kinematic refinement then lets the video prior supervise a 3D motion field over the whole transition, while part assignment and a shared joint constrain the field throughout optimization, keeping per-state errors out of the articulation. Finally, a motion-aware canonical appearance fusion uses the recovered articulation to map both endpoint images onto one canonical surface, texturing regions that only one state reveals. On PartNet-Mobility, TransArt3D outperforms methods given the same two endpoints in geometry and joint accuracy, and surpasses methods given six observed states in articulated-state geometry, joint localization and appearance.
Figures & tables
Figure 2: Per-state errors become false motion. FreeArt3D [ 3 ] scores each state independently, so reconstruction errors of single states enter the moving part and joint and leave the parts and the unobserved intermediate state It inconsistent. TransArt3D fits one joint to the whole generated transition, so parts and poses stay consistent across states.
Figure 3: Overview of TransArt3D. A frozen video model generates the transition between the two endpoint images. Transition-Coupled 3D Consensus (TC3C) lifts the generated frames to 3D jointly, which yields an initial part assignment and joint. Rectified-Flow Kinematic Refinement (RFKR) then lets the video prior correct a 3D motion field over the whole transition, while one shared joint constrains the field and is refined by it. The recovered articulation finally lets Motion-Aware Canonical Appearance Fusion (MACAF) map both endpoint images onto one canonical surface. The output is a textured articulated asset posed along the recovered joint.
Method
dgIoU↓
dcDist↓
dCD↓
AOR ↓
CLIP ↑
LPIPS ↓
Range ↓
Pivot ↓
RS
AS
RS
AS
RS
AS
Multi- state
FreeArt3D [ 3 ]
0.466
0.376
0.073
0.068
0.052
0.047
0.035
0.886
0.064
0.163
0.134
PARIS [ 12 ]
0.871
0.896
0.201
0.217
0.221
0.184
0.053
0.736
0.102
0.871
0.315
SINGAPO † [ 13 ]
0.525
0.518
0.147
0.196
0.067
0.029
0.005
0.826
0.088
0.853
0.248
Arti-Anything [ 8 ]
0.778
0.758
0.217
0.284
0.103
0.174
0.024
0.829
0.026
0.452
0.447
Dual- state
FreeArt3D [ 3 ]
1.026
0.948
0.228
0.234
0.269
0.296
0.093
0.876
0.072
1.051
0.352
Table 1: Quantitative comparison with six observed states in the multi-state block and two observed endpoints in the dual-state block, with both evaluated at six configurations. Within each input block, bold and underlined values denote the best and second-best results among methods covering all 48 objects. The † row reports SINGAPO on 28 valid objects from its seven supported categories and is excluded from ranking.
Figure 4: Qualitative comparison under the dual-state protocol. Conditioning and adapters for each method are given in the supplementary material, and the shared evaluator renders all outputs with matched cameras.
Ablation setting
AS
AOR ↓
Range ↓
Pivot ↓
dgIoU↓
dcDist↓
dCD↓
Cumulative removal
(a) Full
0.357
0.063
0.041
0.025
0.223
0.077
(b) w/o Seam Term
0.372
0.072
0.046
0.030
0.274
0.084
(c) w/o Joint Refine.
0.436
0.092
0.076
0.021
0.336
0.281
(d) w/o Boundary Refine.
0.504
0.112
0.108
0.036
0.566
0.283
Table 2: Geometry and articulation ablations on 48 objects. Rows (b) to (g) remove components cumulatively. Rows (h) to (j) each replace one component of Full and keep the rest. Row (k) runs CHORD [ 17 ] on the ground-truth mesh and fits the parts and joint afterwards. Complete RS and AS results are provided in the supplementary material.
Figure 5: Geometry and articulation ablations from Tab. 2 . Each column shows the textured mesh and the part assignment with the joint.
Variant
G-CLIP ↑
G-LPIPS ↓
M-PSNR ↑
M-LPIPS ↓
(a) Single-Endpoint
0.885
0.063
8.271
0.520
(b) Average pooling
0.835
0.061
16.317
0.505
(c) w/o Routed Init.
0.894
0.053
11.811
0.487
(d) w/o Artic. Routing
0.901
0.053
14.426
0.524
(e) w/o Meas. Guidance
0.907
0.054
8.787
0.404
(f) Full
0.908
0.052
17.237
0.389
Table 3: Canonical appearance ablations on 48 objects with fixed geometry and articulation. G metrics use the full endpoint images, and M metrics use the visible moving-part regions.
Figure 6: Canonical appearance on a surface revealed only at the articulated endpoint.
Figure 7: Qualitative multi-joint extension showing reconstructed states and recovered movable-part decompositions.
Figure 8: Real-world articulated reconstruction from two casually captured endpoint images per object.
Building interactive digital twins requires recovering both 3D geometry and the kinematic structures that govern how objects articulate. Yet existing methods for articulated object reconstruction require explicitly observable motion from multiple articulation states. We introduce a rest-state formulation that reconstructs articulated objects from a single closed configuration, an inherently ill-posed setting where geometry, semantics, and motion priors compensate for the absence of motion cues. Our framework adopts an explicit mesh as an intermediate representation for cross-model verification and fusion, reconciling noisy outputs from vision-language and segmentation models into spatially consistent part structures. To estimate joint parameters without observed motion, we use a video diffusion model to synthesize articulation hypotheses and validate them through geometric consistency. Our approach achieves accurate part decomposition and physically plausible articulation, performing competitively with motion-observing reconstruction-based, generation-based, and modular pretrained-model baselines.
Reconstructing articulated 3D objects from a single image requires jointly inferring object geometry, part structure, and motion parameters from limited visual evidence. A key difficulty lies in the entanglement between motion cues and object structure, which makes direct articulation regression unstable. Existing methods address this challenge through multi-view supervision, retrieval-based assembly, or auxiliary video generation, often sacrificing scalability or efficiency. We present MonoArt, a unified framework grounded in progressive structural reasoning. Rather than predicting articulation directly from image features, MonoArt progressively transforms visual observations into canonical geometry, structured part representations, and motion-aware embeddings within a single architecture. This structured reasoning process enables stable and interpretable articulation inference without external motion templates or multi-stage pipelines. Extensive experiments on PartNet-Mobility demonstrate that OM achieves state-of-the-art performance in both reconstruction accuracy and inference speed. The framework further generalizes to robotic manipulation and articulated scene reconstruction.
Understanding articulated objects is fundamental for robotic interaction, requiring accurate rigid-part discovery and the recovery of their kinematic relations. Existing approaches often treat articulation as a by-product of reconstructed geometry or recover it through per-instance optimization. We instead build on the hypothesis that articulation is directly observable from persistent motion: points on the same rigid part move coherently, while relative motion between parts reveals their kinematic constraints. We present Track2Art, a motion-centric framework for recovering structured articulated objects from RGB-D interaction videos. Track2Art lifts tracked image points into persistent 3D trajectories and combines pretrained tracking features, visual descriptors, and explicit trajectory geometry. These representations are grouped into a variable number of rigid-part hypotheses and subsequently used to recover directed kinematic relations, joint types, and joint geometry through rotation-equivariant learned--analytic reasoning. On the aligned 20-object PartNet-Mobility suite, Track2Art achieves 0.695 Point IoU and 0.410 end-to-end J@20, while requiring neither ground-truth part counts nor test-time optimization.
Xiaotong Li, Yixiong Jing, Junsheng Ding +4
Department of Engineering, University of Cambridge, Cambridge CB2 1PZ, U.K. · School of Engineering and Design, Technical University of Munich · Munich Center for Machine Learning (MCML), Munich, Germany.