What makes a generated 3D object editable?
A generated 3D object is editable when a 3D artist can open it and keep working: select one part, change one dimension, move one pivot, while the rest of the object holds its place.
An example: a table lamp
An editable lamp keeps its shade, its stem, its socket and its base as separate parts, each with its own transform. Set the shade radius from 18 cm to 22 cm and the shade grows. The stem, the socket and the base stay exactly where they were.
What it takes
- Separate parts: one object per component, with a name a pipeline can read.
- Real-world scale: dimensions in metres and centimetres, so the object fits a scene, a configurator or a simulation.
- Hierarchy and pivots: parts grouped the way they are assembled, with a pivot at every hinge and joint.
- Clean geometry: surfaces an artist can keep modelling, texturing and rigging.
- Open formats: GLB, FBX and USD carry parts, hierarchy and pivots from one tool to the next.
Part-based generation and part segmentation
Two families of methods lead to a 3D object with parts. Part segmentation starts from a finished mesh and splits it into regions. Part-based generation builds each component as its own object from the start. Both aim at the same result: 3D that stays editable after it is generated.
Sandflow is a frontier lab building foundation models for 3D, held to that standard. Our first model is on its way.
Previous note: what is a foundation model for 3D.