Generating 3D assets with explicit parts is a balancing act: each part must have its own geometric detail, yet all parts must fit together into a coherent whole. A new arXiv paper, Seeing Both the Forest and the Trees: Reusing Holistic 3D Priors for Part-Decomposed Generation, directly addresses this dual requirement.
The authors note that recent image-conditioned 3D generation models adopt a structured 3D representation, but the abstract snippet does not detail the specific architecture or training scheme. The core idea, as reflected in the title, is to reuse holistic 3D priors—models that capture whole-object structure—to inform part-decomposed generation, rather than treating parts in isolation.
Because the source is only an abstract, the exact mechanism remains unspecified. Still, the framing highlights a persistent challenge in 3D asset generation: reconciling local fidelity with global consistency. The paper is available on arXiv under identifier 2610.21850.