Editing 3D scenes from text in a single forward pass

Most 3D scene editing methods edit individual 2D views separately and stitch them back into 3D, which causes blurry textures and geometry mismatches across viewpoints. VGGT-Edit instead works natively in 3D: it injects text instructions aligned to the scene's spatial poses, then uses a residual transformation head to directly predict geometric displacements while leaving the background intact. The authors also release DeltaScene, a large-scale dataset built with an automated 3D-agreement filtering pipeline to provide clean ground truth. VGGT-Edit substantially outperforms 2D-lifting baselines on multi-view consistency and detail sharpness, and targets researchers and practitioners building interactive 3D applications.