How do you edit 3D objects without starting from scratch?

3D editing typically requires expensive manual work or slow iterative methods. This work introduces Pxform, a dataset of 100K+ paired before/after 3D edits organized by semantic parts (a chair's leg, a cup's handle), and PartFlow, a feedforward network that learns to edit 3D geometry and appearance from these structured examples. The method preserves unmofied regions and maintains consistency across viewpoints without needing 3D masks at test time—outperforming existing approaches on standard editing benchmarks.