Blog ·

CoLa3D online: turn one photo into a 3D scene split into objects

CoLa3D (NeurIPS 2026) builds a 3D scene from a single picture and splits it into separate object meshes. How it works, its MIT license, how to try it in your browser and download the parts, and how to run it locally.

CoLa3D (Composable Latent 3D Decomposition, NeurIPS 2026) turns one picture of a scene into a 3D scene that is already split into separate objects: the sofa, the armchairs and the table each come out as their own mesh, in the same coordinate frame. Most image-to-3D models give you a single fused blob; CoLa3D gives you parts you can move, edit or delete. The code is MIT-licensed, but running it means compiling half a dozen CUDA extensions on Linux. On LiveWeights you just upload the picture.

CoLa3DRenmin University of China & NTU · MIT (TRELLIS.2 and DINOv3 under their own licenses)One picture of a scene in, a 3D scene out, already split into separate object meshes you can move, edit or delete.Try it online →

How it works

  • Built on TRELLIS.2: Microsoft's open image-to-3D model generates the scene's structure and geometry.
  • Composable latents: CoLa3D learns to split the 3D latent into object tokens, so each object is decoded on its own and the parts still fit together.
  • Free, Mask and Point prompts: the model can find objects by itself (Free) or follow masks or points you give it. We run Free mode.

License: can I use it commercially?

CoLa3D's code is MIT. It builds on TRELLIS.2 (MIT) and Meta's DINOv3 (DINOv3 License), which also allow commercial use. The meshes you generate on LiveWeights are yours.

How to try it online

  1. Open the CoLa3D page and sign in, then pick a small credit pack (no subscription).
  2. Upload a picture of a scene with a few clearly separated objects: a corner of a living room, a desk, a table setting. The background is removed automatically.
  3. Press Build 3D scene. It takes about 4 minutes on a 96 GB GPU.
  4. Orbit the scene in the viewer (one colour per object), then download the scene GLB or the ZIP with one GLB per object for Blender, Unity or any 3D tool.

Tips for better results

  • Furniture and everyday objects work best; the model was trained on indoor scenes.
  • Objects hidden behind others have to be guessed, so a view where each object is mostly visible gives cleaner parts.
  • For a single object as a Gaussian splat, try IDOL; for stylized hair, HairDeck.

Run it locally

The code is Master-cai/CoLa3D on GitHub and the checkpoints are on Hugging Face (mastercai/CoLa3D). You need Linux, an NVIDIA GPU, the CUDA 12.8 toolkit and a C++ compiler: ./setup.sh --new-env --all builds FlashAttention, nvdiffrast, CuMesh, FlexGEMM and O-Voxel, then scripts/inference/demo.py --mode free runs the demo.

LiveWeights is not affiliated with the model authors or with the AI Search channel. Facts come from the official pages linked on each model page.