TanGO: Training-Free 3D Editing via Tangent-Space Guidance and Optimization
TanGO enables training-free 3D editing via tangent-space guidance, significantly reducing structural artifacts.
Key Findings
Methodology
TanGO is a training-free 3D editing framework that performs adaptive per-token steering in the tangent space of generative dynamics. It determines each token's control signal strength using a von Mises-Fisher inspired directional discrepancy.
Key Results
- TanGO achieved a CLIP-T score of 0.2301, significantly outperforming existing baselines, indicating a better balance between semantic modification and localized preservation.
- In DINO-I and CLIP-I evaluations, TanGO scored 0.6510 and 0.7679 respectively, demonstrating superior structural and identity preservation.
- Ablation studies confirmed that TanGO's per-token control mechanism remains stable under strong transformations, avoiding trajectory instability and artifact amplification common in global guidance.
Significance
TanGO is significant in the 3D editing field, addressing common semantic artifact issues in traditional methods. Its training-free nature allows for high-quality 3D editing without additional training, applicable in gaming, entertainment, and industrial design.
Technical Contribution
TanGO contrasts with existing methods by performing per-token control in tangent space, offering new theoretical guarantees and engineering possibilities. Its innovative control signal computation and adaptive gain mechanism provide finer control for 3D editing.
Novelty
TanGO is the first to achieve adaptive per-token control in the tangent space of 3D generative models, offering finer editing capabilities and higher quality compared to traditional global control methods.
Limitations
- TanGO may struggle with very complex geometries, as its control signal calculation relies on directional consistency.
- Under extreme editing demands, higher computational resources may be needed to maintain stability.
Future Work
Future research directions include optimizing TanGO's performance on complex geometries and exploring its potential in real-time 3D editing applications.
AI Executive Summary
3D editing technology has become increasingly important in digital asset creation, yet existing methods often face semantic artifact issues. TanGO introduces a training-free 3D editing framework that performs per-token control in the tangent space of generative dynamics, significantly reducing structural artifacts. Experimental results show that TanGO excels in multiple benchmarks, particularly achieving high scores in CLIP-T, DINO-I, and CLIP-I evaluations, demonstrating its balance between semantic modification and localized preservation.
TanGO's technical contributions lie in its innovative control signal computation and adaptive gain mechanism, making 3D editing more precise and stable. Unlike traditional global control strategies, TanGO achieves higher quality editing through adaptive per-token control.
While TanGO performs well in many aspects, it still has limitations when handling complex geometries. Future research will focus on optimizing its performance in complex structures and exploring its potential in real-time 3D editing applications.
Deep Analysis
Background
Recent advances in 3D generative models have significantly propelled the progress of 3D editing technology. However, existing methods often ignore the natural semantic boundaries of high-level representations, leading to significant semantic artifact issues in editing results.
Core Problem
Traditional 3D editing methods often ignore the natural semantic boundaries of high-level representations, leading to significant semantic artifact issues in editing results. This is a pressing problem that needs to be addressed.
Innovation
TanGO achieves training-free 3D editing by performing adaptive per-token control in the tangent space of generative dynamics. Its innovation lies in using a von Mises-Fisher inspired directional discrepancy to determine each token's control signal strength.
Methodology
- �� TanGO performs adaptive per-token control in the tangent space of generative dynamics.
- �� Uses a von Mises-Fisher inspired directional discrepancy to determine each token's control signal strength.
- �� Achieves selective control through an adaptive gain mechanism.
Experiments
Experiments were conducted using multiple benchmark datasets, including the TanGOEdit dataset, to evaluate TanGO's performance in various editing tasks. Metrics such as CLIP-T, DINO-I, and CLIP-I were used for evaluation.
Results
TanGO achieved a CLIP-T score of 0.2301, significantly outperforming existing baselines. In DINO-I and CLIP-I evaluations, TanGO scored 0.6510 and 0.7679 respectively.
Applications
TanGO is applicable in gaming, entertainment, and industrial design, especially in scenarios requiring high-quality 3D editing.
Limitations & Outlook
TanGO may struggle with very complex geometries. Future research will focus on optimizing its performance in complex structures.
Plain Language Accessible to non-experts
Imagine a kitchen where each ingredient represents a token. TanGO acts like a smart chef who decides how to cook each ingredient based on its characteristics. For example, meat needs to be grilled, while vegetables need to be sautéed. TanGO performs adaptive per-token control in the tangent space of generative dynamics, just like a chef choosing different cooking methods for different ingredients, achieving high-quality 3D editing.
ELI14 Explained like you're 14
Imagine you're playing a 3D modeling game and you want to add a hat to your character. Traditional methods are like using a big brush to paint, which might also paint the character's face. TanGO is like a fine brush that only adds color where you want it, without affecting other areas. It's like a smart assistant that knows what you want and can achieve it accurately.
Glossary
Tangent Space
A space formed by the tangents at a point on a surface or curve.
Used in TanGO to achieve per-token control.
Von Mises-Fisher Distribution
A probability distribution used to describe directional data.
Used to calculate each token's control signal strength.
CLIP-T
A metric for evaluating the semantic consistency of edited results with target text.
Used to evaluate TanGO's semantic modification capability.
Token
A basic unit of representation in generative models.
In TanGO, each token represents a part of a 3D object.
Directional Discrepancy
Measures the directional consistency between source and target velocity fields.
Used to determine each token's control signal strength.
Open Questions Unanswered questions from this research
- 1 How to optimize TanGO's performance on complex geometries? Current methods may lack precision in handling complex structures.
- 2 How to reduce computational resource consumption under extreme editing demands?
Applications
Immediate Applications
Game Character Customization
Game developers can use TanGO for high-quality character customization, meeting players' personalized needs.
Long-term Vision
Industrial Design Automation
TanGO can be used for automating 3D model editing in industrial design, improving design efficiency and quality.
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.