Abstract
Impressive progress in generative models and implicit representations gaverise to methods that can generate 3D shapes of high quality. However, beingable to locally control and edit shapes is another essential property that canunlock several content creation applications. Local control can be achievedwith part-aware models, but existing methods require 3D supervision and cannotproduce textures. In this work, we devise PartNeRF, a novel part-awaregenerative model for editable 3D shape synthesis that does not require anyexplicit 3D supervision. Our model generates objects as a set of locallydefined NeRFs, augmented with an affine transformation. This enables severalediting operations such as applying transformations on parts, mixing parts fromdifferent objects etc. To ensure distinct, manipulable parts we enforce a hardassignment of rays to parts that makes sure that the color of each ray is onlydetermined by a single NeRF. As a result, altering one part does not affect theappearance of the others. Evaluations on various ShapeNet categoriesdemonstrate the ability of our model to generate editable 3D objects ofimproved fidelity, compared to previous part-based generative approaches thatrequire 3D supervision or models relying on NeRFs.