3DRAG: 3D-Native Object Drag Editing with Generative Models

arXiv cs.GR details 3DRAG, a training-free framework for point-based 3D drag editing that operates natively in the explicit latent space of large generative models. The pitch is straightforward: instead of relying on slow 2D diffusion priors or direct geometry warping, 3DRAG tries to make interactive edits feel more like a real 3D authoring tool.

The system extends drag displacement fields into 3D space and uses an Occupancy-Aware Winner-Takes-All strategy to cope with sparse geometry. It also applies a generative editing pipeline that transforms latent representations through that 3D field, which is meant to preserve object identity while producing smoother deformations.

For fidelity, the method adds Logits-Level Token Fusion to carry source geometry and texture into the edited region. The authors also built a new dataset for evaluating point-based editing across diverse assets and multi-point instructions, which matters because this kind of interaction has been hard to benchmark cleanly.

For developers, the practical angle is clear: if the method holds up outside the paper, it could make 3D asset editing more intuitive for artists and technical teams working with generative pipelines. The claimed gains are in drag accuracy, geometric fidelity, and preserving unedited regions, all of which are pain points when trying to turn generative 3D into production-friendly tooling.