TanGO: Training-Free 3D Editing via Tangent-space Guidance and Optimization

How can we perform precise, localized 3D model editing without retraining generative models or relying on manual spatial masks? Existing methods often struggle with high-level latent representations (like VecSet), where global token interactions cause unintended structural damage—such as surface collapsing or incomplete transformations—when a user attempts to edit only a specific part of an object.

The authors propose TanGO, a framework that treats 3D editing as an instantaneous one-step optimal control problem. Instead of applying a uniform global update to all tokens, TanGO analyzes the velocity fields of the source and target conditions. It computes a token-wise directional discrepancy—modeled using a von Mises-Fisher distribution—to determine exactly which tokens need to be modified. This allows the model to assign a specific 'steering gain' to each token, amplifying the signal for the target region and suppressing it for regions that should remain unchanged. This process is integrated into the flow-matching trajectory, ensuring that edits are both semantically accurate and geometrically coherent.

TanGO significantly outperforms existing baselines like AnchorFlow and FlowEdit in both semantic alignment (CLIP-T) and structural preservation (CLIP-I). Qualitative results demonstrate that TanGO can handle diverse tasks—such as adding wings to a dragon, changing poses, or replacing emblems—without the boundary bleeding or mesh distortions common in previous approaches. The authors also introduce the TanGOEdit dataset, a new benchmark for evaluating rigid and non-rigid 3D editing, to demonstrate the method's robustness across varied object categories.

This work provides a scalable, training-free solution for 3D editing that respects the semantic boundaries of 3D objects. By leveraging the latent structure of modern generative models rather than fighting against it, TanGO enables high-quality, localized control that is essential for practical applications in digital design, gaming, and virtual environment creation.