T3HG-Editor: Text-driven 3D Human Garment Editing with Body Priors Embedded in SMPL-X

Text-driven 3D human garment editing is a challenging task that requires modifying clothing while preserving the underlying human structure and ensuring consistency across multiple viewpoints. Existing 3D Gaussian Splatting (3DGS) editing methods often suffer from low-fidelity results, including garment distortions, unintended changes to non-target regions, and visual inconsistencies across different camera angles.

The authors propose T3HG-Editor, a system that integrates the SMPL-X (Skinned Multi-Person Linear Model with eXpressive hands and face) model to provide structural and geometric priors. The framework operates in three main stages:

1. Obtainment of Editable Gaussians: The system seeds Gaussians along the surface normals of the SMPL-X model. It then uses a combination of pre-edit and post-edit masks to precisely localize the target garment region, ensuring that both expansion and shrinkage of clothing are handled correctly.
2. Garment Consistent Editing: To maintain consistency across views, the system uses SMPL-X vertices as a unified coordinate reference. It aggregates feature tokens corresponding to the same vertex across multiple key views and propagates them to other views, ensuring that the edited garment texture and shape remain coherent.
3. Gaussian Updating with Overflow Pruning: The system employs a dual-constraint strategy to prevent Gaussians from drifting into non-target regions. It uses a Signed Distance Function (SDF) derived from the SMPL-X model to prune Gaussians that move too far from the body surface, combined with 2D semantic masks to ensure editing remains confined to the target garment.

This work provides a systematic approach to 3D human garment editing that addresses the common failure modes of previous 3DGS editing techniques. By anchoring editing operations to a parametric human body model, T3HG-Editor achieves superior visual fidelity and cross-view consistency, which is critical for applications in virtual try-on, film production, and digital asset creation.