MultiCube introduces 3D generation with part-level control

Researchers published a paper on arXiv, MultiCube: Compositional 3D Generation With Part-Level Semantic and Spatial Control, describing a method for generating composite 3D objects with independent control over the semantics and spatial placement of each part. The input consists of a global text prompt, a textual part layout, and a bounding box arrangement. The output is an object made of separate meshes, one for each specified part. The authors use a two-stage diffusion process: first, a monolithic mesh is generated, then it is simultaneously segmented into parts. The authors have not released the code.

What it means

In our reference on this topic, there are no tools with control at the level of individual parts. The closest generators in terms of task are OpenLRM, Hunyuan3D 2.1, and TRELLIS — they operate in the image-to-3d paradigm: they create a single object from one image or text description, without splitting it into semantic parts and without controlling their relative placement. OpenLRM exports the result to obj, TRELLIS to glb, and Hunyuan3D 2.1 generates shape and textures in separate stages. None of them accept a part layout or spatial arrangement as an input condition. In this sense, MultiCube solves a different problem: not “create an object from an image,” but “assemble an object from specified parts in specified locations.” Since the code is not published, the method cannot be tested in practice right now. For users who need component-level control specifically, MultiCube remains a research work rather than a tool for now. The section on 3D generation as a whole can be viewed on the reference page.