Artic-O

Checkpoint for Artic-O: End-to-End Articulated Object Reconstruction via Latent Geometry Learning, SIGGRAPH Asia 2026 Conference Papers.

Project page · arXiv · Code

Artic-O reconstructs a complete articulated object — geometry, movable-part segmentation, and joint parameters — from a handful of RGB views at two articulation states. It uses a non-pixel-aligned latent as the geometry completion prior, which is what lets it fill in occluded interiors that view-conditioned reconstruction misses.

Files

File	Size	Contents
artic_o_s0_36.pth	2.7 GB	{"model": state_dict, "epoch": 36} — weights only

Optimizer and scheduler state have been stripped; evaluation reads only the model entry.

Usage

Code: https://github.com/Wxyxixixi/Artic-O

hf download wxyxixixi/artic-o artic_o_s0_36.pth --local-dir ./checkpoints/

torchrun --nproc_per_node=4 eval.py \
    --config ./src/configs/artic_o.yaml \
    --ckpt-path ./checkpoints/artic_o_s0_36.pth \
    --run-name release_check

The evaluation data lives in a separate repo: wxyxixixi/artic-o-data.

Results

val-larm-success-all (n=255, deterministic view picks):

metric	Artic-O
CD ↓ (mean over 5 states)	0.01661
F1@0.05 ↑	0.9578
axis_angle@0.25 ↑	0.9490
axis_origin@0.15 ↑	0.9804
Mr@0.3 ↑	0.9216

Evaluation is stochastic (unseeded ODE initialization, subsampled GT); treat CD differences below ~3e-4 as run-to-run noise.

Model

719,478,159 parameters. DINOv2 ViT-L image encoder with a state-embedding table, a frozen flow-matching point decoder, an image-grounded PAT segmentation head, and per-point articulation heads. Trained with a two-stage curriculum on PartNet-Mobility renders.

Citation

@article{wang2026artic,
  title={Artic-O: End-to-End Articulated Object Reconstruction via Latent Geometry Learning},
  author={Wang, Xuyang and Li, Zhenyu and Ding, Jian and Slim, Habib and Wonka, Peter and Li, Hongdong and Elhoseiny, Mohamed},
  journal={arXiv preprint arXiv:2606.21938},
  year={2026}
}

License

Three different licenses are in play here; none extends to the others.

• These weights: non-commercial research use only. The architecture instantiates CroCo-derived (CC BY-NC-SA 4.0) and TripoSG-derived (Tencent Hunyuan Community License) components, so the restrictive upstream terms are the safe reading for the trained model.
• The code at https://github.com/Wxyxixixi/Artic-O is MIT licensed. It does not redistribute either dependency — scripts/fetch_third_party.sh clones them at install time, and each stays under its own terms. See THIRD_PARTY-NOTICES.md there.
• The paper is published by ACM under CC BY 4.0.