VGGT-Edit: Feed-forward Native 3D Scene Editing with Residual Field Prediction

Researchers introduced VGGT-Edit, a feed-forward native 3D scene editing framework that uses text conditioning and direct geometry displacement prediction.

The method injects semantic guidance in sync with scene depth, improving spatial alignment and enabling more consistent edits across views.

They also built the DeltaScene Dataset with automated quality filtering to support training and evaluation.

VGGT-Edit outperformed 2D-lifting baselines on detail quality, cross-view consistency, and inference speed, making interactive 3D editing faster and more reliable.