SpaceFlow: Locally Controllable 3D Generation

SpaceFlow. Locally Controllable 3D Generation. By Neil De La Fuente and seven others, from ETH Zurich, Stanford University, and one other institution. Published on the eighth of October.

Current methods for generating 3D objects offer control, but only globally. How closely the output adheres to a given geometry is usually set by a single control strength for the whole shape, and appearance cannot be specified part by part at all. If one region should match an input exactly while another is free to vary, or one part should be one material and another part something else, there is no explicit way to say so.

SpaceFlow is a training-free pipeline that adds that local control. Its inputs are text descriptions and a collection of geometric primitives. Each primitive stands in for one part of the object and carries its own control level. A high level tells the pipeline to follow the input shape strictly in that region. A low level leaves room for generative completion. The user is, in effect, painting how much freedom each part of the object gets.

Generation happens in two phases. During structure generation, the spatial constraints set by the primitives are enforced inside the generative flow process itself. For appearance synthesis, the generated structure is first segmented, and the segments are matched back to the primitives. Each part is then conditioned only on the text or image cue assigned to it. That isolation is what limits leakage from one part into another.

Regional geometry metrics show the intended split. In high-control regions, SpaceFlow preserves the specified geometry. In low-control areas, it allows plausible variation in shape. A user study indicates that the resulting balance between geometric fidelity and generative freedom stays competitive in overall quality.

Appearance is evaluated separately, on fixed geometry. There, routing text cues to their parts achieves state-of-the-art prompt faithfulness and state-of-the-art accuracy of colour and material. Qualitative results also show image cues being routed to localised parts.

Local control here comes from two acts of routing rather than from retraining: constraints routed to regions during structure generation, and cues routed to parts during appearance synthesis. Global strength becomes a dial per part.