Xudong Cai, Yongcai Wang, Deying Li, Chuanxia Zheng (Renmin University of China, Hebei Key Laboratory RIAS, Nanyang Technological University) · Scene 3D decomposition

CoLa3D: Composable Latent 3D Decomposition

CoLa3D is a NeurIPS 2026 paper from researchers at Renmin University of China, Hebei Key Laboratory RIAS and Nanyang Technological University. It uses a single feed-forward transformer to decompose a whole-scene latent into object meshes that share one coordinate frame. The authors state it is 2x faster than per-object baselines and uses less than 1% of SAM 3D's training data. No further benchmark figures are given in the text. Code is on GitHub, the paper is on arXiv, and the model, training data and benchmark datasets are on Hugging Face. No licence is stated.