NTH

CoLa3D: Composable Latent 3D Decomposition

AuthorsXudong Cai, Yongcai Wang, Deying Li, Chuanxia Zheng

AffiliationsPVG, Nanyang Technological University[0.3em] https://master-cai.github.io/CoLa3D/

November 13, 2026 2 min read
Watch on YouTube
The one-line take

CoLa3D turns a whole-scene representation into a set of coherent, reusable 3D object meshes.

Key results

27.9%
3D-FRONT scene Chamfer Distance reduction

Reduction over the strongest baseline.

57.1%
MetaScenes scene Chamfer Distance reduction

Reduction for CoLa3D with point queries over the strongest baseline.

53.6%
Replica scene Chamfer Distance reduction

Reduction over the strongest baseline on real-world scenes.

1%
SAM 3D training data used

CoLa3D uses less than this share of SAM 3D’s training data.

What the paper found

CoLa3D turns a single image or an existing scene mesh into individually editable 3D object meshes without generating each object separately and trying to align them afterward. Instead, a query-based transformer decomposes one whole-scene latent in a shared coordinate frame, using 2D masks, 3D points, or learnable queries; the extracted object latents are decoded in parallel. Built on TRELLIS.2 and enhanced with Utonia and DINOv3 features, the method produces coherent scene layouts and reduces object interpenetration. On 3D-FRONT, it cuts scene-level Chamfer Distance by 27.9%; on MetaScenes, the point-query version cuts it by 57.1%. On real-world Replica scenes, it reduces scene-level Chamfer Distance by 53.6% compared with the strongest baseline. The paper also reports that CoLa3D runs about 2× faster than feed-forward baselines and uses less than 1% of SAM 3D’s training data. Its visible-depth estimates are competitive with MoGe2, while producing decomposable meshes suited to simulation and interactive editing.

Original abstract

We introduce CoLa3D, a simple and effective model for composable 3D scene decomposition that, given a single whole-scene latent from either a single image or an existing scene mesh as input, directly outputs a set of object meshes with consistent spatial layout that can be assembled back into a coherent scene. CoLa3D introduces a novel paradigm for 3D scene decomposition within a shared whole-scene latent, unlike alternatives that perform per-object generation in canonical spaces and then glue the objects back together with fragile pose and scale estimation. The network is a lightweight, flexible, and scalable query-based transformer that accepts various types of queries, including 2D masks, 3D points, and learnable queries, and learns to decompose the scene into objects in a compact latent space. On both synthetic and real-world datasets, CoLa3D achieves state-of-the-art performance on composable 3D scene decomposition, substantially improving global scene-level consistency and object-level quality while running 2 times faster and using less than 1% of SAM 3D's training data. Visible depth on par with MoGe2 further confirms that latent decomposition preserves accurate visible scene geometry while producing physically complete, decomposable object meshes.

Read the original paper

More in Computer Vision

Browse all 82 papers →