CoLa3D: Composable Latent 3D Decomposition

NeurIPS 2026

Xudong Cai1,3 · Yongcai Wang1,2* · Deying Li1 · Chuanxia Zheng3*†
1School of Information, Renmin University of China
2Hebei Key Laboratory of Real-virtual Integrated Autonomous Systems (RIAS)
3PVG, Nanyang Technological University
TL;DR A single feed-forward transformer that decomposes a whole-scene latent into object meshes in one shared coordinate frame — 2× faster than per-object baselines, with <1% of SAM 3D's training data.

Interactive Results

Drag to rotate · scroll to zoom · toggle individual instances below to see the decomposition.

Interactive scene
VIEWER
— objects

Downstream Application

Comparison with State-of-the-Art

BibTeX

@inproceedings{cai2026cola3d,
  title     = {CoLa3D: Composable Latent 3D Decomposition},
  author    = {Xudong Cai and Yongcai Wang and Deying Li and Chuanxia Zheng},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}