CNS-Edit++: Category-Agnostic 3D Editing with Coupled Neural Shape Representation

CNS-Edit++ extends the earlier CNS-Edit pipeline into category-agnostic 3D editing, aiming to make neural shape manipulation work across broader asset types instead of a narrow object class. The core idea is a coupled neural shape representation: a global latent code carries high-level semantics, while a 3D neural feature volume preserves local spatial context for edits.

The system turns common editing actions into optimization objectives, including copy, resize, delete, mix, point-wise drag, and region-wise drag. That matters because it moves 3D editing closer to the kind of direct manipulation artists and tools teams expect, but in latent space rather than through manual mesh surgery or category-specific pipelines.

To keep untouched areas stable, the framework adds two region-control mechanisms: KV-cache replacement and latent feature regularization. Those safeguards are important for production use, where a useful edit is often judged as much by what stays intact as by what changes.

The work is designed to run on both category-specific inversion models and category-agnostic 3D foundation models, and it was evaluated across multiple generative backends. For game teams experimenting with AI-assisted asset workflows, the main takeaway is that editable 3D generation is becoming more flexible, more model-agnostic, and more compatible with iterative art direction.