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

2026-07-16

AI Summary

This work addresses a key limitation of existing training-free 3D editing methods, which often suffer from semantic artifacts—such as collapse in preserved regions or incomplete transformations—due to global context sharing. To overcome this, the authors propose a novel training-free 3D editing framework that enables token-wise adaptive guidance within the tangent space of generative dynamics. Specifically, they introduce a one-step optimal control rule that dynamically modulates guidance strength based on directional discrepancies between source and target velocity fields. The approach innovatively combines tangent-space guidance with a von Mises–Fisher distribution-based metric for measuring directional differences, enabling fine-grained and adaptive editing control. Extensive experiments demonstrate that the method substantially reduces structural artifacts and outperforms current baselines across multiple evaluation metrics, achieving state-of-the-art performance.

Abstract

While recent flow-matching 3D generative models (e.g., VecSet) adopt structured representations, their tokens share global context, causing conventional training-free editing to suffer from semantic artifacts such as collapsed preserved regions or incomplete transformations. To address this, we propose TanGO, a training-free framework that enables adaptive per-token steering in the tangent space of generative dynamics. To realize this selective control, we formulate a one-step optimal control rule and determine the strength of each token's control signal using a von Mises-Fisher inspired directional discrepancy derived from the source and target velocity fields. Experiments show that TanGO substantially reduces structural artifacts and achieves state-of-the-art performance, outperforming existing 3D editing baselines. The code is publicly available at https://github.com/siw00-lim/TanGO.

Problem

• 3D editing
• training-free
• semantic artifacts
• structured representations
• generative models

Innovation

• training-free editing
• tangent-space guidance
• flow-matching 3D generation
• per-token control
• directional discrepancy